python_code
stringlengths
0
780k
repo_name
stringlengths
7
38
file_path
stringlengths
5
103
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Importance-weighted actor-learner architecture (IMPALA) agent.""" from acme.agents.jax.impala.builder import IMPALABuilder from acme.agents.jax.impala.config import IMPALAConfig from acme.agents.jax.impala.learning import IMPALALearner from acme.agents.jax.impala.networks import IMPALANetworks from acme.agents.jax.impala.networks import make_atari_networks
acme-master
acme/agents/jax/impala/__init__.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Some types/assumptions used in the IMPALA agent.""" from typing import Callable, Tuple from acme.agents.jax.actor_core import RecurrentState from acme.jax import networks from acme.jax import types as jax_types import jax.numpy as jnp # Only simple observations & discrete action spaces for now. Observation = jnp.ndarray Action = int Outputs = Tuple[Tuple[networks.Logits, networks.Value], RecurrentState] PolicyValueInitFn = Callable[[networks.PRNGKey, RecurrentState], networks.Params] PolicyValueFn = Callable[[networks.Params, Observation, RecurrentState], Outputs] RecurrentStateFn = Callable[[jax_types.PRNGKey], RecurrentState]
acme-master
acme/agents/jax/impala/types.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """IMPALA Builder.""" from typing import Any, Callable, Generic, Iterator, List, Optional import acme from acme import adders from acme import core from acme import specs from acme.adders import reverb as reverb_adders from acme.agents.jax import actor_core as actor_core_lib from acme.agents.jax import actors as actors_lib from acme.agents.jax import builders from acme.agents.jax.impala import acting from acme.agents.jax.impala import config as impala_config from acme.agents.jax.impala import learning from acme.agents.jax.impala import networks as impala_networks from acme.datasets import reverb as datasets from acme.jax import networks as networks_lib from acme.jax import utils from acme.jax import variable_utils from acme.utils import counting from acme.utils import loggers import jax import optax import reverb class IMPALABuilder(Generic[actor_core_lib.RecurrentState], builders.ActorLearnerBuilder[impala_networks.IMPALANetworks, acting.ImpalaPolicy, reverb.ReplaySample]): """IMPALA Builder.""" def __init__( self, config: impala_config.IMPALAConfig, table_extension: Optional[Callable[[], Any]] = None, ): """Creates an IMPALA learner.""" self._config = config self._sequence_length = self._config.sequence_length self._table_extension = table_extension def make_replay_tables( self, environment_spec: specs.EnvironmentSpec, policy: acting.ImpalaPolicy, ) -> List[reverb.Table]: """The queue; use XData or INFO log.""" dummy_actor_state = policy.init(jax.random.PRNGKey(0)) signature = reverb_adders.SequenceAdder.signature( environment_spec, policy.get_extras(dummy_actor_state), sequence_length=self._config.sequence_length) # Maybe create rate limiter. # Setting the samples_per_insert ratio less than the default of 1.0, allows # the agent to drop data for the benefit of using data from most up-to-date # policies to compute its learner updates. samples_per_insert = self._config.samples_per_insert if samples_per_insert: if samples_per_insert > 1.0 or samples_per_insert <= 0.0: raise ValueError( 'Impala requires a samples_per_insert ratio in the range (0, 1],' f' but received {samples_per_insert}.') limiter = reverb.rate_limiters.SampleToInsertRatio( samples_per_insert=samples_per_insert, min_size_to_sample=1, error_buffer=self._config.batch_size) else: limiter = reverb.rate_limiters.MinSize(1) table_extensions = [] if self._table_extension is not None: table_extensions = [self._table_extension()] queue = reverb.Table( name=self._config.replay_table_name, sampler=reverb.selectors.Uniform(), remover=reverb.selectors.Fifo(), max_size=self._config.max_queue_size, max_times_sampled=1, rate_limiter=limiter, extensions=table_extensions, signature=signature) return [queue] def make_dataset_iterator( self, replay_client: reverb.Client) -> Iterator[reverb.ReplaySample]: """Creates a dataset.""" batch_size_per_learner = self._config.batch_size // jax.process_count() batch_size_per_device, ragged = divmod(self._config.batch_size, jax.device_count()) if ragged: raise ValueError( 'Learner batch size must be divisible by total number of devices!') dataset = datasets.make_reverb_dataset( table=self._config.replay_table_name, server_address=replay_client.server_address, batch_size=batch_size_per_device, num_parallel_calls=None, max_in_flight_samples_per_worker=2 * batch_size_per_learner) return utils.multi_device_put(dataset.as_numpy_iterator(), jax.local_devices()) def make_adder( self, replay_client: reverb.Client, environment_spec: Optional[specs.EnvironmentSpec], policy: Optional[acting.ImpalaPolicy], ) -> Optional[adders.Adder]: """Creates an adder which handles observations.""" del environment_spec, policy # Note that the last transition in the sequence is used for bootstrapping # only and is ignored otherwise. So we need to make sure that sequences # overlap on one transition, thus "-1" in the period length computation. return reverb_adders.SequenceAdder( client=replay_client, priority_fns={self._config.replay_table_name: None}, period=self._config.sequence_period or (self._sequence_length - 1), sequence_length=self._sequence_length, ) def make_learner( self, random_key: networks_lib.PRNGKey, networks: impala_networks.IMPALANetworks, dataset: Iterator[reverb.ReplaySample], logger_fn: loggers.LoggerFactory, environment_spec: specs.EnvironmentSpec, replay_client: Optional[reverb.Client] = None, counter: Optional[counting.Counter] = None, ) -> core.Learner: del environment_spec, replay_client optimizer = optax.chain( optax.clip_by_global_norm(self._config.max_gradient_norm), optax.adam( self._config.learning_rate, b1=self._config.adam_momentum_decay, b2=self._config.adam_variance_decay, eps=self._config.adam_eps, eps_root=self._config.adam_eps_root)) return learning.IMPALALearner( networks=networks, iterator=dataset, optimizer=optimizer, random_key=random_key, discount=self._config.discount, entropy_cost=self._config.entropy_cost, baseline_cost=self._config.baseline_cost, max_abs_reward=self._config.max_abs_reward, counter=counter, logger=logger_fn('learner'), ) def make_actor( self, random_key: networks_lib.PRNGKey, policy: acting.ImpalaPolicy, environment_spec: specs.EnvironmentSpec, variable_source: Optional[core.VariableSource] = None, adder: Optional[adders.Adder] = None, ) -> acme.Actor: del environment_spec variable_client = variable_utils.VariableClient( client=variable_source, key='network', update_period=self._config.variable_update_period) return actors_lib.GenericActor(policy, random_key, variable_client, adder) def make_policy(self, networks: impala_networks.IMPALANetworks, environment_spec: specs.EnvironmentSpec, evaluation: bool = False) -> acting.ImpalaPolicy: return acting.get_actor_core(networks, environment_spec, evaluation)
acme-master
acme/agents/jax/impala/builder.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """IMPALA networks definition.""" from acme import specs from acme.jax import networks as networks_lib IMPALANetworks = networks_lib.UnrollableNetwork def make_atari_networks(env_spec: specs.EnvironmentSpec) -> IMPALANetworks: """Builds default IMPALA networks for Atari games.""" def make_core_module() -> networks_lib.DeepIMPALAAtariNetwork: return networks_lib.DeepIMPALAAtariNetwork(env_spec.actions.num_values) return networks_lib.make_unrollable_network(env_spec, make_core_module)
acme-master
acme/agents/jax/impala/networks.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Learner for the IMPALA actor-critic agent.""" import time from typing import Dict, Iterator, List, NamedTuple, Optional, Sequence, Tuple from absl import logging import acme from acme.agents.jax.impala import networks as impala_networks from acme.jax import losses from acme.jax import networks as networks_lib from acme.jax import utils from acme.utils import counting from acme.utils import loggers import jax import jax.numpy as jnp import numpy as np import optax import reverb _PMAP_AXIS_NAME = 'data' class TrainingState(NamedTuple): """Training state consists of network parameters and optimiser state.""" params: networks_lib.Params opt_state: optax.OptState class IMPALALearner(acme.Learner): """Learner for an importanced-weighted advantage actor-critic.""" def __init__( self, networks: impala_networks.IMPALANetworks, iterator: Iterator[reverb.ReplaySample], optimizer: optax.GradientTransformation, random_key: networks_lib.PRNGKey, discount: float = 0.99, entropy_cost: float = 0.0, baseline_cost: float = 1.0, max_abs_reward: float = np.inf, counter: Optional[counting.Counter] = None, logger: Optional[loggers.Logger] = None, devices: Optional[Sequence[jax.Device]] = None, prefetch_size: int = 2, ): local_devices = jax.local_devices() process_id = jax.process_index() logging.info('Learner process id: %s. Devices passed: %s', process_id, devices) logging.info('Learner process id: %s. Local devices from JAX API: %s', process_id, local_devices) self._devices = devices or local_devices self._local_devices = [d for d in self._devices if d in local_devices] self._iterator = iterator def unroll_without_rng( params: networks_lib.Params, observations: networks_lib.Observation, initial_state: networks_lib.RecurrentState ) -> Tuple[networks_lib.NetworkOutput, networks_lib.RecurrentState]: unused_rng = jax.random.PRNGKey(0) return networks.unroll(params, unused_rng, observations, initial_state) loss_fn = losses.impala_loss( # TODO(b/244319884): Consider supporting the use of RNG in impala_loss. unroll_fn=unroll_without_rng, discount=discount, max_abs_reward=max_abs_reward, baseline_cost=baseline_cost, entropy_cost=entropy_cost) @jax.jit def sgd_step( state: TrainingState, sample: reverb.ReplaySample ) -> Tuple[TrainingState, Dict[str, jnp.ndarray]]: """Computes an SGD step, returning new state and metrics for logging.""" # Compute gradients. grad_fn = jax.value_and_grad(loss_fn, has_aux=True) (loss_value, metrics), gradients = grad_fn(state.params, sample) # Average gradients over pmap replicas before optimizer update. gradients = jax.lax.pmean(gradients, _PMAP_AXIS_NAME) # Apply updates. updates, new_opt_state = optimizer.update(gradients, state.opt_state) new_params = optax.apply_updates(state.params, updates) metrics.update({ 'loss': loss_value, 'param_norm': optax.global_norm(new_params), 'param_updates_norm': optax.global_norm(updates), }) new_state = TrainingState(params=new_params, opt_state=new_opt_state) return new_state, metrics def make_initial_state(key: jnp.ndarray) -> TrainingState: """Initialises the training state (parameters and optimiser state).""" initial_params = networks.init(key) return TrainingState( params=initial_params, opt_state=optimizer.init(initial_params)) # Initialise training state (parameters and optimiser state). state = make_initial_state(random_key) self._state = utils.replicate_in_all_devices(state, self._local_devices) self._sgd_step = jax.pmap( sgd_step, axis_name=_PMAP_AXIS_NAME, devices=self._devices) # Set up logging/counting. self._counter = counter or counting.Counter() self._logger = logger or loggers.make_default_logger( 'learner', steps_key=self._counter.get_steps_key()) def step(self): """Does a step of SGD and logs the results.""" samples = next(self._iterator) # Do a batch of SGD. start = time.time() self._state, results = self._sgd_step(self._state, samples) # Take results from first replica. # NOTE: This measure will be a noisy estimate for the purposes of the logs # as it does not pmean over all devices. results = utils.get_from_first_device(results) # Update our counts and record them. counts = self._counter.increment(steps=1, time_elapsed=time.time() - start) # Maybe write logs. self._logger.write({**results, **counts}) def get_variables(self, names: Sequence[str]) -> List[networks_lib.Params]: # Return first replica of parameters. return utils.get_from_first_device([self._state.params], as_numpy=False) def save(self) -> TrainingState: # Serialize only the first replica of parameters and optimizer state. return utils.get_from_first_device(self._state) def restore(self, state: TrainingState): self._state = utils.replicate_in_all_devices(state, self._local_devices)
acme-master
acme/agents/jax/impala/learning.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """IMPALA actor implementation.""" from typing import Generic, Mapping, Tuple from acme import specs from acme.agents.jax import actor_core as actor_core_lib from acme.agents.jax.impala import networks as impala_networks from acme.jax import networks as networks_lib from acme.jax import types as jax_types import chex import jax import jax.numpy as jnp ImpalaExtras = Mapping[str, jnp.ndarray] @chex.dataclass(frozen=True, mappable_dataclass=False) class ImpalaActorState(Generic[actor_core_lib.RecurrentState]): rng: jax_types.PRNGKey logits: networks_lib.Logits recurrent_state: actor_core_lib.RecurrentState prev_recurrent_state: actor_core_lib.RecurrentState ImpalaPolicy = actor_core_lib.ActorCore[ ImpalaActorState[actor_core_lib.RecurrentState], ImpalaExtras] def get_actor_core( networks: impala_networks.IMPALANetworks, environment_spec: specs.EnvironmentSpec, evaluation: bool = False, ) -> ImpalaPolicy: """Creates an Impala ActorCore.""" dummy_logits = jnp.zeros(environment_spec.actions.num_values) def init( rng: jax_types.PRNGKey ) -> ImpalaActorState[actor_core_lib.RecurrentState]: rng, init_state_rng = jax.random.split(rng) initial_state = networks.init_recurrent_state(init_state_rng, None) return ImpalaActorState( rng=rng, logits=dummy_logits, recurrent_state=initial_state, prev_recurrent_state=initial_state) def select_action( params: networks_lib.Params, observation: networks_lib.Observation, state: ImpalaActorState[actor_core_lib.RecurrentState], ) -> Tuple[networks_lib.Action, ImpalaActorState[actor_core_lib.RecurrentState]]: rng, apply_rng, policy_rng = jax.random.split(state.rng, 3) (logits, _), new_recurrent_state = networks.apply( params, apply_rng, observation, state.recurrent_state, ) if evaluation: action = jnp.argmax(logits, axis=-1) else: action = jax.random.categorical(policy_rng, logits) return action, ImpalaActorState( rng=rng, logits=logits, recurrent_state=new_recurrent_state, prev_recurrent_state=state.recurrent_state) def get_extras( state: ImpalaActorState[actor_core_lib.RecurrentState]) -> ImpalaExtras: return {'logits': state.logits, 'core_state': state.prev_recurrent_state} return actor_core_lib.ActorCore( init=init, select_action=select_action, get_extras=get_extras)
acme-master
acme/agents/jax/impala/acting.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """RND config.""" import dataclasses @dataclasses.dataclass class RNDConfig: """Configuration options for RND.""" # Learning rate for the predictor. predictor_learning_rate: float = 1e-4 # If True, the direct rl algorithm is using the SequenceAdder data format. is_sequence_based: bool = False # How many gradient updates to perform per step. num_sgd_steps_per_step: int = 1
acme-master
acme/agents/jax/rnd/config.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """RND agent.""" from acme.agents.jax.rnd.builder import RNDBuilder from acme.agents.jax.rnd.config import RNDConfig from acme.agents.jax.rnd.learning import rnd_loss from acme.agents.jax.rnd.learning import rnd_update_step from acme.agents.jax.rnd.learning import RNDLearner from acme.agents.jax.rnd.learning import RNDTrainingState from acme.agents.jax.rnd.networks import compute_rnd_reward from acme.agents.jax.rnd.networks import make_networks from acme.agents.jax.rnd.networks import rnd_reward_fn from acme.agents.jax.rnd.networks import RNDNetworks
acme-master
acme/agents/jax/rnd/__init__.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """RND Builder.""" from typing import Callable, Generic, Iterator, List, Optional from acme import adders from acme import core from acme import specs from acme.agents.jax import actor_core as actor_core_lib from acme.agents.jax import builders from acme.agents.jax.rnd import config as rnd_config from acme.agents.jax.rnd import learning as rnd_learning from acme.agents.jax.rnd import networks as rnd_networks from acme.jax import networks as networks_lib from acme.jax.types import Policy from acme.utils import counting from acme.utils import loggers import jax import optax import reverb class RNDBuilder(Generic[rnd_networks.DirectRLNetworks, Policy], builders.ActorLearnerBuilder[rnd_networks.RNDNetworks, Policy, reverb.ReplaySample]): """RND Builder.""" def __init__( self, rl_agent: builders.ActorLearnerBuilder[rnd_networks.DirectRLNetworks, Policy, reverb.ReplaySample], config: rnd_config.RNDConfig, logger_fn: Callable[[], loggers.Logger] = lambda: None, ): """Implements a builder for RND using rl_agent as forward RL algorithm. Args: rl_agent: The standard RL agent used by RND to optimize the generator. config: A config with RND HPs. logger_fn: a logger factory for the rl_agent's learner. """ self._rl_agent = rl_agent self._config = config self._logger_fn = logger_fn def make_learner( self, random_key: networks_lib.PRNGKey, networks: rnd_networks.RNDNetworks[rnd_networks.DirectRLNetworks], dataset: Iterator[reverb.ReplaySample], logger_fn: loggers.LoggerFactory, environment_spec: specs.EnvironmentSpec, replay_client: Optional[reverb.Client] = None, counter: Optional[counting.Counter] = None, ) -> core.Learner: direct_rl_learner_key, rnd_learner_key = jax.random.split(random_key) counter = counter or counting.Counter() direct_rl_counter = counting.Counter(counter, 'direct_rl') def direct_rl_learner_factory( networks: rnd_networks.DirectRLNetworks, dataset: Iterator[reverb.ReplaySample]) -> core.Learner: return self._rl_agent.make_learner( direct_rl_learner_key, networks, dataset, logger_fn=lambda name: self._logger_fn(), environment_spec=environment_spec, replay_client=replay_client, counter=direct_rl_counter) optimizer = optax.adam(learning_rate=self._config.predictor_learning_rate) return rnd_learning.RNDLearner( direct_rl_learner_factory=direct_rl_learner_factory, iterator=dataset, optimizer=optimizer, rnd_network=networks, rng_key=rnd_learner_key, is_sequence_based=self._config.is_sequence_based, grad_updates_per_batch=self._config.num_sgd_steps_per_step, counter=counter, logger=logger_fn('learner')) def make_replay_tables( self, environment_spec: specs.EnvironmentSpec, policy: Policy, ) -> List[reverb.Table]: return self._rl_agent.make_replay_tables(environment_spec, policy) def make_dataset_iterator( # pytype: disable=signature-mismatch # overriding-return-type-checks self, replay_client: reverb.Client) -> Optional[Iterator[reverb.ReplaySample]]: return self._rl_agent.make_dataset_iterator(replay_client) def make_adder(self, replay_client: reverb.Client, environment_spec: Optional[specs.EnvironmentSpec], policy: Optional[Policy]) -> Optional[adders.Adder]: return self._rl_agent.make_adder(replay_client, environment_spec, policy) def make_actor( self, random_key: networks_lib.PRNGKey, policy: Policy, environment_spec: specs.EnvironmentSpec, variable_source: Optional[core.VariableSource] = None, adder: Optional[adders.Adder] = None, ) -> core.Actor: return self._rl_agent.make_actor(random_key, policy, environment_spec, variable_source, adder) def make_policy(self, networks: rnd_networks.RNDNetworks, environment_spec: specs.EnvironmentSpec, evaluation: bool = False) -> actor_core_lib.FeedForwardPolicy: """Construct the policy.""" return self._rl_agent.make_policy(networks.direct_rl_networks, environment_spec, evaluation)
acme-master
acme/agents/jax/rnd/builder.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Networks definitions for the BC agent.""" import dataclasses import functools from typing import Callable, Generic, Tuple, TypeVar from acme import specs from acme import types from acme.jax import networks as networks_lib from acme.jax import utils import haiku as hk import jax.numpy as jnp DirectRLNetworks = TypeVar('DirectRLNetworks') @dataclasses.dataclass class RNDNetworks(Generic[DirectRLNetworks]): """Container of RND networks factories.""" target: networks_lib.FeedForwardNetwork predictor: networks_lib.FeedForwardNetwork # Function from predictor output, target output, and original reward to reward get_reward: Callable[ [networks_lib.NetworkOutput, networks_lib.NetworkOutput, jnp.ndarray], jnp.ndarray] direct_rl_networks: DirectRLNetworks = None # See Appendix A.2 of https://arxiv.org/pdf/1810.12894.pdf def rnd_reward_fn( predictor_output: networks_lib.NetworkOutput, target_output: networks_lib.NetworkOutput, original_reward: jnp.ndarray, intrinsic_reward_coefficient: float = 1.0, extrinsic_reward_coefficient: float = 0.0, ) -> jnp.ndarray: intrinsic_reward = jnp.mean( jnp.square(predictor_output - target_output), axis=-1) return (intrinsic_reward_coefficient * intrinsic_reward + extrinsic_reward_coefficient * original_reward) def make_networks( spec: specs.EnvironmentSpec, direct_rl_networks: DirectRLNetworks, layer_sizes: Tuple[int, ...] = (256, 256), intrinsic_reward_coefficient: float = 1.0, extrinsic_reward_coefficient: float = 0.0, ) -> RNDNetworks[DirectRLNetworks]: """Creates networks used by the agent and returns RNDNetworks. Args: spec: Environment spec. direct_rl_networks: Networks used by a direct rl algorithm. layer_sizes: Layer sizes. intrinsic_reward_coefficient: Multiplier on intrinsic reward. extrinsic_reward_coefficient: Multiplier on extrinsic reward. Returns: The RND networks. """ def _rnd_fn(obs, act): # RND does not use the action but other variants like RED do. del act network = networks_lib.LayerNormMLP(list(layer_sizes)) return network(obs) target = hk.without_apply_rng(hk.transform(_rnd_fn)) predictor = hk.without_apply_rng(hk.transform(_rnd_fn)) # Create dummy observations and actions to create network parameters. dummy_obs = utils.zeros_like(spec.observations) dummy_obs = utils.add_batch_dim(dummy_obs) return RNDNetworks( target=networks_lib.FeedForwardNetwork( lambda key: target.init(key, dummy_obs, ()), target.apply), predictor=networks_lib.FeedForwardNetwork( lambda key: predictor.init(key, dummy_obs, ()), predictor.apply), direct_rl_networks=direct_rl_networks, get_reward=functools.partial( rnd_reward_fn, intrinsic_reward_coefficient=intrinsic_reward_coefficient, extrinsic_reward_coefficient=extrinsic_reward_coefficient)) def compute_rnd_reward(predictor_params: networks_lib.Params, target_params: networks_lib.Params, transitions: types.Transition, networks: RNDNetworks) -> jnp.ndarray: """Computes the intrinsic RND reward for a given transition. Args: predictor_params: Parameters of the predictor network. target_params: Parameters of the target network. transitions: The sample to compute rewards for. networks: RND networks Returns: The rewards as an ndarray. """ target_output = networks.target.apply(target_params, transitions.observation, transitions.action) predictor_output = networks.predictor.apply(predictor_params, transitions.observation, transitions.action) return networks.get_reward(predictor_output, target_output, transitions.reward)
acme-master
acme/agents/jax/rnd/networks.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """RND learner implementation.""" import functools import time from typing import Any, Callable, Dict, Iterator, List, NamedTuple, Optional, Tuple import acme from acme import types from acme.agents.jax.rnd import networks as rnd_networks from acme.jax import networks as networks_lib from acme.jax import utils from acme.utils import counting from acme.utils import loggers from acme.utils import reverb_utils import jax import jax.numpy as jnp import optax import reverb class RNDTrainingState(NamedTuple): """Contains training state for the learner.""" optimizer_state: optax.OptState params: networks_lib.Params target_params: networks_lib.Params steps: int class GlobalTrainingState(NamedTuple): """Contains training state of the RND learner.""" rewarder_state: RNDTrainingState learner_state: Any RNDLoss = Callable[[networks_lib.Params, networks_lib.Params, types.Transition], float] def rnd_update_step( state: RNDTrainingState, transitions: types.Transition, loss_fn: RNDLoss, optimizer: optax.GradientTransformation ) -> Tuple[RNDTrainingState, Dict[str, jnp.ndarray]]: """Run an update steps on the given transitions. Args: state: The learner state. transitions: Transitions to update on. loss_fn: The loss function. optimizer: The optimizer of the predictor network. Returns: A new state and metrics. """ loss, grads = jax.value_and_grad(loss_fn)( state.params, state.target_params, transitions=transitions) update, optimizer_state = optimizer.update(grads, state.optimizer_state) params = optax.apply_updates(state.params, update) new_state = RNDTrainingState( optimizer_state=optimizer_state, params=params, target_params=state.target_params, steps=state.steps + 1, ) return new_state, {'rnd_loss': loss} def rnd_loss( predictor_params: networks_lib.Params, target_params: networks_lib.Params, transitions: types.Transition, networks: rnd_networks.RNDNetworks, ) -> float: """The Random Network Distillation loss. See https://arxiv.org/pdf/1810.12894.pdf A.2 Args: predictor_params: Parameters of the predictor target_params: Parameters of the target transitions: Transitions to compute the loss on. networks: RND networks Returns: The MSE loss as a float. """ target_output = networks.target.apply(target_params, transitions.observation, transitions.action) predictor_output = networks.predictor.apply(predictor_params, transitions.observation, transitions.action) return jnp.mean(jnp.square(target_output - predictor_output)) class RNDLearner(acme.Learner): """RND learner.""" def __init__( self, direct_rl_learner_factory: Callable[[Any, Iterator[reverb.ReplaySample]], acme.Learner], iterator: Iterator[reverb.ReplaySample], optimizer: optax.GradientTransformation, rnd_network: rnd_networks.RNDNetworks, rng_key: jnp.ndarray, grad_updates_per_batch: int, is_sequence_based: bool, counter: Optional[counting.Counter] = None, logger: Optional[loggers.Logger] = None): self._is_sequence_based = is_sequence_based target_key, predictor_key = jax.random.split(rng_key) target_params = rnd_network.target.init(target_key) predictor_params = rnd_network.predictor.init(predictor_key) optimizer_state = optimizer.init(predictor_params) self._state = RNDTrainingState( optimizer_state=optimizer_state, params=predictor_params, target_params=target_params, steps=0) # General learner book-keeping and loggers. self._counter = counter or counting.Counter() self._logger = logger or loggers.make_default_logger( 'learner', asynchronous=True, serialize_fn=utils.fetch_devicearray, steps_key=self._counter.get_steps_key()) loss = functools.partial(rnd_loss, networks=rnd_network) self._update = functools.partial(rnd_update_step, loss_fn=loss, optimizer=optimizer) self._update = utils.process_multiple_batches(self._update, grad_updates_per_batch) self._update = jax.jit(self._update) self._get_reward = jax.jit( functools.partial( rnd_networks.compute_rnd_reward, networks=rnd_network)) # Generator expression that works the same as an iterator. # https://pymbook.readthedocs.io/en/latest/igd.html#generator-expressions updated_iterator = (self._process_sample(sample) for sample in iterator) self._direct_rl_learner = direct_rl_learner_factory( rnd_network.direct_rl_networks, updated_iterator) # Do not record timestamps until after the first learning step is done. # This is to avoid including the time it takes for actors to come online and # fill the replay buffer. self._timestamp = None def _process_sample(self, sample: reverb.ReplaySample) -> reverb.ReplaySample: """Uses the replay sample to train and update its reward. Args: sample: Replay sample to train on. Returns: The sample replay sample with an updated reward. """ transitions = reverb_utils.replay_sample_to_sars_transition( sample, is_sequence=self._is_sequence_based) self._state, metrics = self._update(self._state, transitions) rewards = self._get_reward(self._state.params, self._state.target_params, transitions) # Compute elapsed time. timestamp = time.time() elapsed_time = timestamp - self._timestamp if self._timestamp else 0 self._timestamp = timestamp # Increment counts and record the current time counts = self._counter.increment(steps=1, walltime=elapsed_time) # Attempts to write the logs. self._logger.write({**metrics, **counts}) return sample._replace(data=sample.data._replace(reward=rewards)) def step(self): self._direct_rl_learner.step() def get_variables(self, names: List[str]) -> List[Any]: rnd_variables = { 'target_params': self._state.target_params, 'predictor_params': self._state.params } learner_names = [name for name in names if name not in rnd_variables] learner_dict = {} if learner_names: learner_dict = dict( zip(learner_names, self._direct_rl_learner.get_variables(learner_names))) variables = [ rnd_variables.get(name, learner_dict.get(name, None)) for name in names ] return variables def save(self) -> GlobalTrainingState: return GlobalTrainingState( rewarder_state=self._state, learner_state=self._direct_rl_learner.save()) def restore(self, state: GlobalTrainingState): self._state = state.rewarder_state self._direct_rl_learner.restore(state.learner_state)
acme-master
acme/agents/jax/rnd/learning.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """DQN config.""" import dataclasses from typing import Callable, Optional, Sequence, Union from acme.adders import reverb as adders_reverb import jax.numpy as jnp import numpy as np @dataclasses.dataclass class DQNConfig: """Configuration options for DQN agent. Attributes: epsilon: for use by epsilon-greedy policies. If multiple, the epsilons are alternated randomly per-episode. eval_epsilon: for use by evaluation epsilon-greedy policies. seed: Random seed. learning_rate: Learning rate for Adam optimizer. Could be a number or a function defining a schedule. adam_eps: Epsilon for Adam optimizer. discount: Discount rate applied to value per timestep. n_step: N-step TD learning. target_update_period: Update target network every period. max_gradient_norm: For gradient clipping. batch_size: Number of transitions per batch. min_replay_size: Minimum replay size. max_replay_size: Maximum replay size. replay_table_name: Reverb table, defaults to DEFAULT_PRIORITY_TABLE. importance_sampling_exponent: Importance sampling for replay. priority_exponent: Priority exponent for replay. prefetch_size: Prefetch size for reverb replay performance. samples_per_insert: Ratio of learning samples to insert. samples_per_insert_tolerance_rate: Rate to be used for the SampleToInsertRatio rate limitter tolerance. See a formula in make_replay_tables for more details. num_sgd_steps_per_step: How many gradient updates to perform per learner step. """ epsilon: Union[float, Sequence[float]] = 0.05 eval_epsilon: Optional[float] = None # TODO(b/191706065): update all clients and remove this field. seed: int = 1 # Learning rule learning_rate: Union[float, Callable[[int], float]] = 1e-3 adam_eps: float = 1e-8 # Eps for Adam optimizer. discount: float = 0.99 # Discount rate applied to value per timestep. n_step: int = 5 # N-step TD learning. target_update_period: int = 100 # Update target network every period. max_gradient_norm: float = np.inf # For gradient clipping. # Replay options batch_size: int = 256 min_replay_size: int = 1_000 max_replay_size: int = 1_000_000 replay_table_name: str = adders_reverb.DEFAULT_PRIORITY_TABLE importance_sampling_exponent: float = 0.2 priority_exponent: float = 0.6 prefetch_size: int = 4 samples_per_insert: float = 0.5 samples_per_insert_tolerance_rate: float = 0.1 num_sgd_steps_per_step: int = 1 def logspace_epsilons( num_epsilons: int, epsilon: float = 0.017 ) -> Union[Sequence[float], jnp.ndarray]: """`num_epsilons` of logspace-distributed values, with median `epsilon`.""" if num_epsilons <= 1: return (epsilon,) return jnp.logspace(1, 8, num_epsilons, base=epsilon ** (2./9.))
acme-master
acme/agents/jax/dqn/config.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """SgdLearner takes steps of SGD on a LossFn.""" import functools import time from typing import Dict, Iterator, List, NamedTuple, Optional, Tuple import acme from acme.adders import reverb as adders from acme.jax import networks as networks_lib from acme.jax import utils from acme.utils import async_utils from acme.utils import counting from acme.utils import loggers import jax import jax.numpy as jnp import optax import reverb import tree import typing_extensions # The pmap axis name. Data means data parallelization. PMAP_AXIS_NAME = 'data' class ReverbUpdate(NamedTuple): """Tuple for updating reverb priority information.""" keys: jnp.ndarray priorities: jnp.ndarray class LossExtra(NamedTuple): """Extra information that is returned along with loss value.""" metrics: Dict[str, jax.Array] # New optional updated priorities for the samples. reverb_priorities: Optional[jax.Array] = None class LossFn(typing_extensions.Protocol): """A LossFn calculates a loss on a single batch of data.""" def __call__( self, network: networks_lib.TypedFeedForwardNetwork, params: networks_lib.Params, target_params: networks_lib.Params, batch: reverb.ReplaySample, key: networks_lib.PRNGKey, ) -> Tuple[jax.Array, LossExtra]: """Calculates a loss on a single batch of data.""" class TrainingState(NamedTuple): """Holds the agent's training state.""" params: networks_lib.Params target_params: networks_lib.Params opt_state: optax.OptState steps: int rng_key: networks_lib.PRNGKey class SGDLearner(acme.Learner): """An Acme learner based around SGD on batches. This learner currently supports optional prioritized replay and assumes a TrainingState as described above. """ def __init__(self, network: networks_lib.TypedFeedForwardNetwork, loss_fn: LossFn, optimizer: optax.GradientTransformation, data_iterator: Iterator[utils.PrefetchingSplit], target_update_period: int, random_key: networks_lib.PRNGKey, replay_client: Optional[reverb.Client] = None, replay_table_name: str = adders.DEFAULT_PRIORITY_TABLE, counter: Optional[counting.Counter] = None, logger: Optional[loggers.Logger] = None, num_sgd_steps_per_step: int = 1): """Initialize the SGD learner.""" self.network = network # Internalize the loss_fn with network. self._loss = jax.jit(functools.partial(loss_fn, self.network)) # SGD performs the loss, optimizer update and periodic target net update. def sgd_step(state: TrainingState, batch: reverb.ReplaySample) -> Tuple[TrainingState, LossExtra]: next_rng_key, rng_key = jax.random.split(state.rng_key) # Implements one SGD step of the loss and updates training state (loss, extra), grads = jax.value_and_grad( self._loss, has_aux=True)(state.params, state.target_params, batch, rng_key) loss = jax.lax.pmean(loss, axis_name=PMAP_AXIS_NAME) # Average gradients over pmap replicas before optimizer update. grads = jax.lax.pmean(grads, axis_name=PMAP_AXIS_NAME) # Apply the optimizer updates updates, new_opt_state = optimizer.update(grads, state.opt_state) new_params = optax.apply_updates(state.params, updates) extra.metrics.update({'total_loss': loss}) # Periodically update target networks. steps = state.steps + 1 target_params = optax.periodic_update(new_params, state.target_params, # pytype: disable=wrong-arg-types # numpy-scalars steps, target_update_period) new_training_state = TrainingState( new_params, target_params, new_opt_state, steps, next_rng_key) return new_training_state, extra def postprocess_aux(extra: LossExtra) -> LossExtra: reverb_priorities = jax.tree_util.tree_map( lambda a: jnp.reshape(a, (-1, *a.shape[2:])), extra.reverb_priorities) return extra._replace( metrics=jax.tree_util.tree_map(jnp.mean, extra.metrics), reverb_priorities=reverb_priorities) self._num_sgd_steps_per_step = num_sgd_steps_per_step sgd_step = utils.process_multiple_batches(sgd_step, num_sgd_steps_per_step, postprocess_aux) self._sgd_step = jax.pmap( sgd_step, axis_name=PMAP_AXIS_NAME, devices=jax.devices()) # Internalise agent components self._data_iterator = data_iterator self._target_update_period = target_update_period self._counter = counter or counting.Counter() self._logger = logger or loggers.TerminalLogger('learner', time_delta=1.) # Do not record timestamps until after the first learning step is done. # This is to avoid including the time it takes for actors to come online and # fill the replay buffer. self._timestamp = None # Initialize the network parameters key_params, key_target, key_state = jax.random.split(random_key, 3) initial_params = self.network.init(key_params) initial_target_params = self.network.init(key_target) state = TrainingState( params=initial_params, target_params=initial_target_params, opt_state=optimizer.init(initial_params), steps=0, rng_key=key_state, ) self._state = utils.replicate_in_all_devices(state, jax.local_devices()) # Update replay priorities def update_priorities(reverb_update: ReverbUpdate) -> None: if replay_client is None: return keys, priorities = tree.map_structure( # Fetch array and combine device and batch dimensions. lambda x: utils.fetch_devicearray(x).reshape((-1,) + x.shape[2:]), (reverb_update.keys, reverb_update.priorities)) replay_client.mutate_priorities( table=replay_table_name, updates=dict(zip(keys, priorities))) self._replay_client = replay_client self._async_priority_updater = async_utils.AsyncExecutor(update_priorities) self._current_step = 0 def step(self): """Takes one SGD step on the learner.""" with jax.profiler.StepTraceAnnotation('step', step_num=self._current_step): prefetching_split = next(self._data_iterator) # In this case the host property of the prefetching split contains only # replay keys and the device property is the prefetched full original # sample. Key is on host since it's uint64 type. reverb_keys = prefetching_split.host batch: reverb.ReplaySample = prefetching_split.device self._state, extra = self._sgd_step(self._state, batch) # Compute elapsed time. timestamp = time.time() elapsed = timestamp - self._timestamp if self._timestamp else 0 self._timestamp = timestamp if self._replay_client and extra.reverb_priorities is not None: reverb_update = ReverbUpdate(reverb_keys, extra.reverb_priorities) self._async_priority_updater.put(reverb_update) steps_per_sec = (self._num_sgd_steps_per_step / elapsed) if elapsed else 0 self._current_step, metrics = utils.get_from_first_device( (self._state.steps, extra.metrics)) metrics['steps_per_second'] = steps_per_sec # Update our counts and record it. result = self._counter.increment( steps=self._num_sgd_steps_per_step, walltime=elapsed) result.update(metrics) self._logger.write(result) def get_variables(self, names: List[str]) -> List[networks_lib.Params]: # Return first replica of parameters. return utils.get_from_first_device([self._state.params]) def save(self) -> TrainingState: # Serialize only the first replica of parameters and optimizer state. return utils.get_from_first_device(self._state) def restore(self, state: TrainingState): self._state = utils.replicate_in_all_devices(state, jax.local_devices())
acme-master
acme/agents/jax/dqn/learning_lib.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Implementation of a deep Q-networks (DQN) agent.""" from acme.agents.jax.dqn.actor import behavior_policy from acme.agents.jax.dqn.actor import default_behavior_policy from acme.agents.jax.dqn.actor import DQNPolicy from acme.agents.jax.dqn.actor import Epsilon from acme.agents.jax.dqn.actor import EpsilonPolicy from acme.agents.jax.dqn.builder import DistributionalDQNBuilder from acme.agents.jax.dqn.builder import DQNBuilder from acme.agents.jax.dqn.config import DQNConfig from acme.agents.jax.dqn.learning import DQNLearner from acme.agents.jax.dqn.learning_lib import LossExtra from acme.agents.jax.dqn.learning_lib import LossFn from acme.agents.jax.dqn.learning_lib import ReverbUpdate from acme.agents.jax.dqn.learning_lib import SGDLearner from acme.agents.jax.dqn.losses import PrioritizedCategoricalDoubleQLearning from acme.agents.jax.dqn.losses import PrioritizedDoubleQLearning from acme.agents.jax.dqn.losses import QLearning from acme.agents.jax.dqn.losses import QrDqn from acme.agents.jax.dqn.networks import DQNNetworks
acme-master
acme/agents/jax/dqn/__init__.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """DQN Builder.""" from typing import Iterator, List, Optional, Sequence from acme import adders from acme import core from acme import specs from acme.adders import reverb as adders_reverb from acme.agents.jax import actors from acme.agents.jax import builders from acme.agents.jax.dqn import actor as dqn_actor from acme.agents.jax.dqn import config as dqn_config from acme.agents.jax.dqn import learning_lib from acme.agents.jax.dqn import networks as dqn_networks from acme.datasets import reverb as datasets from acme.jax import networks as networks_lib from acme.jax import utils from acme.jax import variable_utils from acme.utils import counting from acme.utils import loggers import jax import optax import reverb from reverb import rate_limiters class DQNBuilder(builders.ActorLearnerBuilder[dqn_networks.DQNNetworks, dqn_actor.DQNPolicy, utils.PrefetchingSplit]): """DQN Builder.""" def __init__(self, config: dqn_config.DQNConfig, loss_fn: learning_lib.LossFn, actor_backend: Optional[str] = 'cpu'): """Creates DQN learner and the behavior policies. Args: config: DQN config. loss_fn: A loss function. actor_backend: Which backend to use when jitting the policy. """ self._config = config self._loss_fn = loss_fn self._actor_backend = actor_backend def make_learner( self, random_key: networks_lib.PRNGKey, networks: dqn_networks.DQNNetworks, dataset: Iterator[utils.PrefetchingSplit], logger_fn: loggers.LoggerFactory, environment_spec: Optional[specs.EnvironmentSpec], replay_client: Optional[reverb.Client] = None, counter: Optional[counting.Counter] = None, ) -> core.Learner: del environment_spec return learning_lib.SGDLearner( network=networks.policy_network, random_key=random_key, optimizer=optax.adam( self._config.learning_rate, eps=self._config.adam_eps), target_update_period=self._config.target_update_period, data_iterator=dataset, loss_fn=self._loss_fn, replay_client=replay_client, replay_table_name=self._config.replay_table_name, counter=counter, num_sgd_steps_per_step=self._config.num_sgd_steps_per_step, logger=logger_fn('learner')) def make_actor( self, random_key: networks_lib.PRNGKey, policy: dqn_actor.DQNPolicy, environment_spec: Optional[specs.EnvironmentSpec], variable_source: Optional[core.VariableSource] = None, adder: Optional[adders.Adder] = None, ) -> core.Actor: del environment_spec assert variable_source is not None # Inference happens on CPU, so it's better to move variables there too. variable_client = variable_utils.VariableClient( variable_source, '', device='cpu') return actors.GenericActor( actor=policy, random_key=random_key, variable_client=variable_client, adder=adder, backend=self._actor_backend) def make_replay_tables( self, environment_spec: specs.EnvironmentSpec, policy: dqn_actor.DQNPolicy, ) -> List[reverb.Table]: """Creates reverb tables for the algorithm.""" del policy samples_per_insert_tolerance = ( self._config.samples_per_insert_tolerance_rate * self._config.samples_per_insert) error_buffer = self._config.min_replay_size * samples_per_insert_tolerance limiter = rate_limiters.SampleToInsertRatio( min_size_to_sample=self._config.min_replay_size, samples_per_insert=self._config.samples_per_insert, error_buffer=error_buffer) return [ reverb.Table( name=self._config.replay_table_name, sampler=reverb.selectors.Prioritized( self._config.priority_exponent), remover=reverb.selectors.Fifo(), max_size=self._config.max_replay_size, rate_limiter=limiter, signature=adders_reverb.NStepTransitionAdder.signature( environment_spec)) ] @property def batch_size_per_device(self) -> int: """Splits the batch size across local devices.""" # Account for the number of SGD steps per step. batch_size = self._config.batch_size * self._config.num_sgd_steps_per_step num_devices = jax.local_device_count() # TODO(bshahr): Using jax.device_count will not be valid when colocating # learning and inference. if batch_size % num_devices != 0: raise ValueError( 'The DQN learner received a batch size that is not divisible by the ' f'number of available learner devices. Got: batch_size={batch_size}, ' f'num_devices={num_devices}.') return batch_size // num_devices def make_dataset_iterator( self, replay_client: reverb.Client, ) -> Iterator[utils.PrefetchingSplit]: """Creates a dataset iterator to use for learning.""" dataset = datasets.make_reverb_dataset( table=self._config.replay_table_name, server_address=replay_client.server_address, batch_size=self.batch_size_per_device, prefetch_size=self._config.prefetch_size) return utils.multi_device_put( dataset.as_numpy_iterator(), jax.local_devices(), split_fn=utils.keep_key_on_host) def make_adder( self, replay_client: reverb.Client, environment_spec: Optional[specs.EnvironmentSpec], policy: Optional[dqn_actor.DQNPolicy], ) -> Optional[adders.Adder]: """Creates an adder which handles observations.""" del environment_spec, policy return adders_reverb.NStepTransitionAdder( priority_fns={self._config.replay_table_name: None}, client=replay_client, n_step=self._config.n_step, discount=self._config.discount) def _policy_epsilons(self, evaluation: bool) -> Sequence[float]: if evaluation and self._config.eval_epsilon: epsilon = self._config.eval_epsilon else: epsilon = self._config.epsilon epsilons = epsilon if isinstance(epsilon, Sequence) else (epsilon,) return epsilons def make_policy(self, networks: dqn_networks.DQNNetworks, environment_spec: specs.EnvironmentSpec, evaluation: bool = False) -> dqn_actor.DQNPolicy: """Creates the policy.""" del environment_spec return dqn_actor.alternating_epsilons_actor_core( dqn_actor.behavior_policy(networks), epsilons=self._policy_epsilons(evaluation)) class DistributionalDQNBuilder(DQNBuilder): """Distributional DQN Builder.""" def make_policy(self, networks: dqn_networks.DQNNetworks, environment_spec: specs.EnvironmentSpec, evaluation: bool = False) -> dqn_actor.DQNPolicy: """Creates the policy. Expects network head which returns a tuple with the first entry representing q-values. Creates the agent policy given the collection of network components and environment spec. An optional boolean can be given to indicate if the policy will be used for evaluation. Args: networks: struct describing the networks needed to generate the policy. environment_spec: struct describing the specs of the environment. evaluation: when true, a version of the policy to use for evaluation should be returned. This is algorithm-specific so if an algorithm makes no distinction between behavior and evaluation policies this boolean may be ignored. Returns: Behavior policy or evaluation policy for the agent. """ del environment_spec def get_action_values(params: networks_lib.Params, observation: networks_lib.Observation, *args, **kwargs) -> networks_lib.NetworkOutput: return networks.policy_network.apply(params, observation, *args, **kwargs)[0] typed_network = networks_lib.TypedFeedForwardNetwork( init=networks.policy_network.init, apply=get_action_values) behavior_policy = dqn_actor.behavior_policy( dqn_networks.DQNNetworks(policy_network=typed_network)) return dqn_actor.alternating_epsilons_actor_core( behavior_policy, epsilons=self._policy_epsilons(evaluation))
acme-master
acme/agents/jax/dqn/builder.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Network definitions for DQN.""" import dataclasses from typing import Callable, Optional from acme.jax import networks as networks_lib from acme.jax import types import rlax Epsilon = float EpsilonPolicy = Callable[[ networks_lib.Params, networks_lib.PRNGKey, networks_lib.Observation, Epsilon ], networks_lib.Action] EpsilonSampleFn = Callable[[networks_lib.NetworkOutput, types.PRNGKey, Epsilon], networks_lib.Action] EpsilonLogProbFn = Callable[ [networks_lib.NetworkOutput, networks_lib.Action, Epsilon], networks_lib.LogProb] def default_sample_fn(action_values: networks_lib.NetworkOutput, key: types.PRNGKey, epsilon: Epsilon) -> networks_lib.Action: return rlax.epsilon_greedy(epsilon).sample(key, action_values) @dataclasses.dataclass class DQNNetworks: """The network and pure functions for the DQN agent. Attributes: policy_network: The policy network. sample_fn: A pure function. Samples an action based on the network output. log_prob: A pure function. Computes log-probability for an action. """ policy_network: networks_lib.TypedFeedForwardNetwork sample_fn: EpsilonSampleFn = default_sample_fn log_prob: Optional[EpsilonLogProbFn] = None
acme-master
acme/agents/jax/dqn/networks.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """DQN losses.""" import dataclasses from typing import Tuple from acme import types from acme.agents.jax.dqn import learning_lib from acme.jax import networks as networks_lib import chex import jax import jax.numpy as jnp import reverb import rlax @dataclasses.dataclass class PrioritizedDoubleQLearning(learning_lib.LossFn): """Clipped double q learning with prioritization on TD error.""" discount: float = 0.99 importance_sampling_exponent: float = 0.2 max_abs_reward: float = 1. huber_loss_parameter: float = 1. def __call__( self, network: networks_lib.TypedFeedForwardNetwork, params: networks_lib.Params, target_params: networks_lib.Params, batch: reverb.ReplaySample, key: networks_lib.PRNGKey, ) -> Tuple[jax.Array, learning_lib.LossExtra]: """Calculate a loss on a single batch of data.""" transitions: types.Transition = batch.data probs = batch.info.probability # Forward pass. key1, key2, key3 = jax.random.split(key, 3) q_tm1 = network.apply( params, transitions.observation, is_training=True, key=key1) q_t_value = network.apply( target_params, transitions.next_observation, is_training=True, key=key2) q_t_selector = network.apply( params, transitions.next_observation, is_training=True, key=key3) # Cast and clip rewards. d_t = (transitions.discount * self.discount).astype(jnp.float32) r_t = jnp.clip(transitions.reward, -self.max_abs_reward, self.max_abs_reward).astype(jnp.float32) # Compute double Q-learning n-step TD-error. batch_error = jax.vmap(rlax.double_q_learning) td_error = batch_error(q_tm1, transitions.action, r_t, d_t, q_t_value, q_t_selector) batch_loss = rlax.huber_loss(td_error, self.huber_loss_parameter) # Importance weighting. importance_weights = (1. / probs).astype(jnp.float32) importance_weights **= self.importance_sampling_exponent importance_weights /= jnp.max(importance_weights) # Reweight. loss = jnp.mean(importance_weights * batch_loss) # [] extra = learning_lib.LossExtra( metrics={}, reverb_priorities=jnp.abs(td_error).astype(jnp.float64)) return loss, extra @dataclasses.dataclass class QrDqn(learning_lib.LossFn): """Quantile Regression DQN. https://arxiv.org/abs/1710.10044 """ num_atoms: int = 51 huber_param: float = 1.0 def __call__( self, network: networks_lib.TypedFeedForwardNetwork, params: networks_lib.Params, target_params: networks_lib.Params, batch: reverb.ReplaySample, key: networks_lib.PRNGKey, ) -> Tuple[jax.Array, learning_lib.LossExtra]: """Calculate a loss on a single batch of data.""" transitions: types.Transition = batch.data key1, key2 = jax.random.split(key) _, dist_q_tm1 = network.apply( params, transitions.observation, is_training=True, key=key1) _, dist_q_target_t = network.apply( target_params, transitions.next_observation, is_training=True, key=key2) batch_size = len(transitions.observation) chex.assert_shape( dist_q_tm1, ( batch_size, None, self.num_atoms, ), custom_message=f'Expected (batch_size, num_actions, num_atoms), got: {dist_q_tm1.shape}', include_default_message=True) chex.assert_shape( dist_q_target_t, ( batch_size, None, self.num_atoms, ), custom_message=f'Expected (batch_size, num_actions, num_atoms), got: {dist_q_target_t.shape}', include_default_message=True) # Swap distribution and action dimension, since # rlax.quantile_q_learning expects it that way. dist_q_tm1 = jnp.swapaxes(dist_q_tm1, 1, 2) dist_q_target_t = jnp.swapaxes(dist_q_target_t, 1, 2) quantiles = ( (jnp.arange(self.num_atoms, dtype=jnp.float32) + 0.5) / self.num_atoms) batch_quantile_q_learning = jax.vmap( rlax.quantile_q_learning, in_axes=(0, None, 0, 0, 0, 0, 0, None)) losses = batch_quantile_q_learning( dist_q_tm1, quantiles, transitions.action, transitions.reward, transitions.discount, dist_q_target_t, # No double Q-learning here. dist_q_target_t, self.huber_param, ) loss = jnp.mean(losses) chex.assert_shape(losses, (batch_size,)) extra = learning_lib.LossExtra(metrics={'mean_loss': loss}) return loss, extra @dataclasses.dataclass class PrioritizedCategoricalDoubleQLearning(learning_lib.LossFn): """Categorical double q learning with prioritization on TD error.""" discount: float = 0.99 importance_sampling_exponent: float = 0.2 max_abs_reward: float = 1. def __call__( self, network: networks_lib.TypedFeedForwardNetwork, params: networks_lib.Params, target_params: networks_lib.Params, batch: reverb.ReplaySample, key: networks_lib.PRNGKey, ) -> Tuple[jax.Array, learning_lib.LossExtra]: """Calculate a loss on a single batch of data.""" transitions: types.Transition = batch.data probs = batch.info.probability # Forward pass. key1, key2, key3 = jax.random.split(key, 3) _, logits_tm1, atoms_tm1 = network.apply( params, transitions.observation, is_training=True, key=key1) _, logits_t, atoms_t = network.apply( target_params, transitions.next_observation, is_training=True, key=key2) q_t_selector, _, _ = network.apply( params, transitions.next_observation, is_training=True, key=key3) # Cast and clip rewards. d_t = (transitions.discount * self.discount).astype(jnp.float32) r_t = jnp.clip(transitions.reward, -self.max_abs_reward, self.max_abs_reward).astype(jnp.float32) # Compute categorical double Q-learning loss. batch_loss_fn = jax.vmap( rlax.categorical_double_q_learning, in_axes=(None, 0, 0, 0, 0, None, 0, 0)) batch_loss = batch_loss_fn(atoms_tm1, logits_tm1, transitions.action, r_t, d_t, atoms_t, logits_t, q_t_selector) # Importance weighting. importance_weights = (1. / probs).astype(jnp.float32) importance_weights **= self.importance_sampling_exponent importance_weights /= jnp.max(importance_weights) # Reweight. loss = jnp.mean(importance_weights * batch_loss) # [] extra = learning_lib.LossExtra( metrics={}, reverb_priorities=jnp.abs(batch_loss).astype(jnp.float64)) return loss, extra @dataclasses.dataclass class QLearning(learning_lib.LossFn): """Deep q learning. This matches the original DQN loss: https://arxiv.org/abs/1312.5602. It differs by two aspects that improve it on the optimization side - it uses Adam instead of RMSProp as an optimizer - it uses a square loss instead of the Huber one. """ discount: float = 0.99 max_abs_reward: float = 1. def __call__( self, network: networks_lib.TypedFeedForwardNetwork, params: networks_lib.Params, target_params: networks_lib.Params, batch: reverb.ReplaySample, key: networks_lib.PRNGKey, ) -> Tuple[jax.Array, learning_lib.LossExtra]: """Calculate a loss on a single batch of data.""" transitions: types.Transition = batch.data # Forward pass. key1, key2 = jax.random.split(key) q_tm1 = network.apply( params, transitions.observation, is_training=True, key=key1) q_t = network.apply( target_params, transitions.next_observation, is_training=True, key=key2) # Cast and clip rewards. d_t = (transitions.discount * self.discount).astype(jnp.float32) r_t = jnp.clip(transitions.reward, -self.max_abs_reward, self.max_abs_reward).astype(jnp.float32) # Compute Q-learning TD-error. batch_error = jax.vmap(rlax.q_learning) td_error = batch_error(q_tm1, transitions.action, r_t, d_t, q_t) batch_loss = jnp.square(td_error) loss = jnp.mean(batch_loss) extra = learning_lib.LossExtra(metrics={}) return loss, extra @dataclasses.dataclass class RegularizedQLearning(learning_lib.LossFn): """Regularized Q-learning. Implements DQNReg loss function: https://arxiv.org/abs/2101.03958. This is almost identical to QLearning except: 1) Adds a regularization term; 2) Uses vanilla TD error without huber loss. 3) No reward clipping. """ discount: float = 0.99 regularizer_coeff = 0.1 def __call__( self, network: networks_lib.TypedFeedForwardNetwork, params: networks_lib.Params, target_params: networks_lib.Params, batch: reverb.ReplaySample, key: networks_lib.PRNGKey, ) -> Tuple[jax.Array, learning_lib.LossExtra]: """Calculate a loss on a single batch of data.""" transitions: types.Transition = batch.data # Forward pass. key1, key2 = jax.random.split(key) q_tm1 = network.apply( params, transitions.observation, is_training=True, key=key1) q_t = network.apply( target_params, transitions.next_observation, is_training=True, key=key2) d_t = (transitions.discount * self.discount).astype(jnp.float32) # Compute Q-learning TD-error. batch_error = jax.vmap(rlax.q_learning) td_error = batch_error( q_tm1, transitions.action, transitions.reward, d_t, q_t) td_error = 0.5 * jnp.square(td_error) def select(qtm1, action): return qtm1[action] q_regularizer = jax.vmap(select)(q_tm1, transitions.action) loss = self.regularizer_coeff * jnp.mean(q_regularizer) + jnp.mean(td_error) extra = learning_lib.LossExtra(metrics={}) return loss, extra @dataclasses.dataclass class MunchausenQLearning(learning_lib.LossFn): """Munchausen q learning. Implements M-DQN: https://arxiv.org/abs/2007.14430. """ entropy_temperature: float = 0.03 # tau parameter munchausen_coefficient: float = 0.9 # alpha parameter clip_value_min: float = -1e3 discount: float = 0.99 max_abs_reward: float = 1. huber_loss_parameter: float = 1. def __call__( self, network: networks_lib.TypedFeedForwardNetwork, params: networks_lib.Params, target_params: networks_lib.Params, batch: reverb.ReplaySample, key: networks_lib.PRNGKey, ) -> Tuple[jax.Array, learning_lib.LossExtra]: """Calculate a loss on a single batch of data.""" transitions: types.Transition = batch.data # Forward pass. key1, key2, key3 = jax.random.split(key, 3) q_online_s = network.apply( params, transitions.observation, is_training=True, key=key1) action_one_hot = jax.nn.one_hot(transitions.action, q_online_s.shape[-1]) q_online_sa = jnp.sum(action_one_hot * q_online_s, axis=-1) q_target_s = network.apply( target_params, transitions.observation, is_training=True, key=key2) q_target_next = network.apply( target_params, transitions.next_observation, is_training=True, key=key3) # Cast and clip rewards. d_t = (transitions.discount * self.discount).astype(jnp.float32) r_t = jnp.clip(transitions.reward, -self.max_abs_reward, self.max_abs_reward).astype(jnp.float32) # Munchausen term : tau * log_pi(a|s) munchausen_term = self.entropy_temperature * jax.nn.log_softmax( q_target_s / self.entropy_temperature, axis=-1) munchausen_term_a = jnp.sum(action_one_hot * munchausen_term, axis=-1) munchausen_term_a = jnp.clip(munchausen_term_a, a_min=self.clip_value_min, a_max=0.) # Soft Bellman operator applied to q next_v = self.entropy_temperature * jax.nn.logsumexp( q_target_next / self.entropy_temperature, axis=-1) target_q = jax.lax.stop_gradient(r_t + self.munchausen_coefficient * munchausen_term_a + d_t * next_v) batch_loss = rlax.huber_loss(target_q - q_online_sa, self.huber_loss_parameter) loss = jnp.mean(batch_loss) extra = learning_lib.LossExtra(metrics={}) return loss, extra
acme-master
acme/agents/jax/dqn/losses.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """DQN learner implementation.""" from typing import Iterator, Optional from acme.adders import reverb as adders from acme.agents.jax.dqn import learning_lib from acme.agents.jax.dqn import losses from acme.jax import networks as networks_lib from acme.jax import utils from acme.utils import counting from acme.utils import loggers import optax import reverb class DQNLearner(learning_lib.SGDLearner): """DQN learner. We are in the process of migrating towards a more general SGDLearner to allow for easy configuration of the loss. This is maintained now for compatibility. """ def __init__(self, network: networks_lib.TypedFeedForwardNetwork, discount: float, importance_sampling_exponent: float, target_update_period: int, iterator: Iterator[utils.PrefetchingSplit], optimizer: optax.GradientTransformation, random_key: networks_lib.PRNGKey, max_abs_reward: float = 1., huber_loss_parameter: float = 1., replay_client: Optional[reverb.Client] = None, replay_table_name: str = adders.DEFAULT_PRIORITY_TABLE, counter: Optional[counting.Counter] = None, logger: Optional[loggers.Logger] = None, num_sgd_steps_per_step: int = 1): """Initializes the learner.""" loss_fn = losses.PrioritizedDoubleQLearning( discount=discount, importance_sampling_exponent=importance_sampling_exponent, max_abs_reward=max_abs_reward, huber_loss_parameter=huber_loss_parameter, ) super().__init__( network=network, loss_fn=loss_fn, optimizer=optimizer, data_iterator=iterator, target_update_period=target_update_period, random_key=random_key, replay_client=replay_client, replay_table_name=replay_table_name, counter=counter, logger=logger, num_sgd_steps_per_step=num_sgd_steps_per_step, )
acme-master
acme/agents/jax/dqn/learning.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """DQN actor helpers.""" from typing import Callable, Sequence from acme.agents.jax import actor_core as actor_core_lib from acme.agents.jax.dqn import networks as dqn_networks from acme.jax import networks as networks_lib from acme.jax import utils import chex import jax import jax.numpy as jnp Epsilon = float EpsilonPolicy = Callable[[ networks_lib.Params, networks_lib.PRNGKey, networks_lib .Observation, Epsilon ], networks_lib.Action] @chex.dataclass(frozen=True, mappable_dataclass=False) class EpsilonActorState: rng: networks_lib.PRNGKey epsilon: jnp.ndarray DQNPolicy = actor_core_lib.ActorCore[EpsilonActorState, None] def alternating_epsilons_actor_core(policy_network: EpsilonPolicy, epsilons: Sequence[float]) -> DQNPolicy: """Returns actor components for alternating epsilon exploration. Args: policy_network: A feedforward action selecting function. epsilons: epsilons to alternate per-episode for epsilon-greedy exploration. Returns: A feedforward policy. """ epsilons = jnp.array(epsilons) def apply_and_sample(params: networks_lib.Params, observation: networks_lib.Observation, state: EpsilonActorState): random_key, key = jax.random.split(state.rng) actions = policy_network(params, key, observation, state.epsilon) # pytype: disable=wrong-arg-types # jax-ndarray return (actions.astype(jnp.int32), EpsilonActorState(rng=random_key, epsilon=state.epsilon)) def policy_init(random_key: networks_lib.PRNGKey): random_key, key = jax.random.split(random_key) epsilon = jax.random.choice(key, epsilons) return EpsilonActorState(rng=random_key, epsilon=epsilon) return actor_core_lib.ActorCore( init=policy_init, select_action=apply_and_sample, get_extras=lambda _: None) def behavior_policy(networks: dqn_networks.DQNNetworks) -> EpsilonPolicy: """A policy with parameterized epsilon-greedy exploration.""" def apply_and_sample(params: networks_lib.Params, key: networks_lib.PRNGKey, observation: networks_lib.Observation, epsilon: Epsilon ) -> networks_lib.Action: # TODO(b/161332815): Make JAX Actor work with batched or unbatched inputs. observation = utils.add_batch_dim(observation) action_values = networks.policy_network.apply( params, observation, is_training=False) action_values = utils.squeeze_batch_dim(action_values) return networks.sample_fn(action_values, key, epsilon) return apply_and_sample def default_behavior_policy(networks: dqn_networks.DQNNetworks, epsilon: Epsilon) -> EpsilonPolicy: """A policy with a fixed-epsilon epsilon-greedy exploration. DEPRECATED: use behavior_policy instead. Args: networks: DQN networks epsilon: sampling parameter that overrides the one in EpsilonPolicy Returns: epsilon-greedy behavior policy with fixed epsilon """ # TODO(lukstafi): remove this function and migrate its users. def apply_and_sample(params: networks_lib.Params, key: networks_lib.PRNGKey, observation: networks_lib.Observation, _: Epsilon ) -> networks_lib.Action: # TODO(b/161332815): Make JAX Actor work with batched or unbatched inputs. observation = utils.add_batch_dim(observation) action_values = networks.policy_network.apply( params, observation, is_training=False) action_values = utils.squeeze_batch_dim(action_values) return networks.sample_fn(action_values, key, epsilon) return apply_and_sample
acme-master
acme/agents/jax/dqn/actor.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Defines Rainbow DQN, using JAX.""" import dataclasses from typing import Callable from acme import specs from acme.agents.jax.dqn import actor as dqn_actor from acme.agents.jax.dqn import builder from acme.agents.jax.dqn import config as dqn_config from acme.agents.jax.dqn import losses from acme.jax import networks as networks_lib from acme.jax import utils import rlax NetworkFactory = Callable[[specs.EnvironmentSpec], networks_lib.FeedForwardNetwork] @dataclasses.dataclass class RainbowConfig(dqn_config.DQNConfig): """(Additional) configuration options for RainbowDQN.""" max_abs_reward: float = 1.0 # For clipping reward def apply_policy_and_sample( network: networks_lib.FeedForwardNetwork,) -> dqn_actor.EpsilonPolicy: """Returns a function that computes actions. Note that this differs from default_behavior_policy with that it expects c51-style network head which returns a tuple with the first entry representing q-values. Args: network: A c51-style feedforward network. Returns: A feedforward policy. """ def apply_and_sample(params, key, obs, epsilon): # TODO(b/161332815): Make JAX Actor work with batched or unbatched inputs. obs = utils.add_batch_dim(obs) action_values = network.apply(params, obs)[0] action_values = utils.squeeze_batch_dim(action_values) return rlax.epsilon_greedy(epsilon).sample(key, action_values) return apply_and_sample def eval_policy(network: networks_lib.FeedForwardNetwork, eval_epsilon: float) -> dqn_actor.EpsilonPolicy: """Returns a function that computes actions. Note that this differs from default_behavior_policy with that it expects c51-style network head which returns a tuple with the first entry representing q-values. Args: network: A c51-style feedforward network. eval_epsilon: for epsilon-greedy exploration. Returns: A feedforward policy. """ policy = apply_policy_and_sample(network) def apply_and_sample(params, key, obs, _): return policy(params, key, obs, eval_epsilon) return apply_and_sample def make_builder(config: RainbowConfig): """Returns a DQNBuilder with a pre-built loss function.""" loss_fn = losses.PrioritizedCategoricalDoubleQLearning( discount=config.discount, importance_sampling_exponent=config.importance_sampling_exponent, max_abs_reward=config.max_abs_reward, ) return builder.DQNBuilder(config, loss_fn=loss_fn)
acme-master
acme/agents/jax/dqn/rainbow.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """MBOP config.""" import dataclasses from acme.agents.jax.mbop import mppi @dataclasses.dataclass(frozen=True) class MBOPConfig: """Configuration options for the MBOP agent. Attributes: mppi_config: Planner hyperparameters. learning_rate: Learning rate. num_networks: Number of networks in the ensembles. num_sgd_steps_per_step: How many gradient updates to perform per learner step. """ mppi_config: mppi.MPPIConfig = dataclasses.field( default_factory=mppi.MPPIConfig ) learning_rate: float = 3e-4 num_networks: int = 5 num_sgd_steps_per_step: int = 1
acme-master
acme/agents/jax/mbop/config.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """MBOP models.""" import functools from typing import Callable, Generic, Optional, Tuple from acme import specs from acme.agents.jax import actor_core from acme.agents.jax.mbop import ensemble from acme.agents.jax.mbop import networks as mbop_networks from acme.jax import networks from acme.jax import utils import chex import jax # World, policy prior and n-step return models. These are backed by the # corresponding networks. WorldModel = Callable[[networks.Params, networks.Observation, networks.Action], Tuple[networks.Observation, networks.Value]] MakeWorldModel = Callable[[mbop_networks.WorldModelNetwork], WorldModel] PolicyPrior = actor_core.ActorCore MakePolicyPrior = Callable[ [mbop_networks.PolicyPriorNetwork, specs.EnvironmentSpec], PolicyPrior] NStepReturn = Callable[[networks.Params, networks.Observation, networks.Action], networks.Value] MakeNStepReturn = Callable[[mbop_networks.NStepReturnNetwork], NStepReturn] @chex.dataclass(frozen=True, mappable_dataclass=False) class PolicyPriorState(Generic[actor_core.RecurrentState]): """State of a policy prior. Attributes: rng: Random key. action_tm1: Previous action. recurrent_state: Recurrent state. It will be none for non-recurrent, e.g. feed forward, policies. """ rng: networks.PRNGKey action_tm1: networks.Action recurrent_state: Optional[actor_core.RecurrentState] = None FeedForwardPolicyState = PolicyPriorState[actor_core.NoneType] def feed_forward_policy_prior_to_actor_core( policy: actor_core.RecurrentPolicy, initial_action_tm1: networks.Action ) -> actor_core.ActorCore[PolicyPriorState, actor_core.NoneType]: """A convenience adaptor from a feed forward policy prior to ActorCore. Args: policy: A feed forward policy prior. In the planner and other components, the previous action is explicitly passed as an argument to the policy prior together with the observation to infer the next action. Therefore, we model feed forward policy priors as recurrent ActorCore policies with previous action being the recurrent state. initial_action_tm1: Initial previous action. This will usually be a zero tensor. Returns: an ActorCore representing the feed forward policy prior. """ def select_action(params: networks.Params, observation: networks.Observation, state: FeedForwardPolicyState): rng, policy_rng = jax.random.split(state.rng) action = policy(params, policy_rng, observation, state.action_tm1) return action, PolicyPriorState(rng, action) def init(rng: networks.PRNGKey) -> FeedForwardPolicyState: return PolicyPriorState(rng, initial_action_tm1) def get_extras(unused_state: FeedForwardPolicyState) -> actor_core.NoneType: return None return actor_core.ActorCore( init=init, select_action=select_action, get_extras=get_extras) def make_ensemble_world_model( world_model_network: mbop_networks.WorldModelNetwork) -> WorldModel: """Creates an ensemble world model from its network.""" return functools.partial(ensemble.apply_round_robin, world_model_network.apply) def make_ensemble_policy_prior( policy_prior_network: mbop_networks.PolicyPriorNetwork, spec: specs.EnvironmentSpec, use_round_robin: bool = True) -> PolicyPrior: """Creates an ensemble policy prior from its network. Args: policy_prior_network: The policy prior network. spec: Environment specification. use_round_robin: Whether to use round robin or mean to calculate the policy prior over the ensemble members. Returns: A policy prior. """ def _policy_prior(params: networks.Params, key: networks.PRNGKey, observation_t: networks.Observation, action_tm1: networks.Action) -> networks.Action: # Regressor policies are deterministic. del key apply_fn = ( ensemble.apply_round_robin if use_round_robin else ensemble.apply_mean) return apply_fn( policy_prior_network.apply, params, observation_t=observation_t, action_tm1=action_tm1) dummy_action = utils.zeros_like(spec.actions) dummy_action = utils.add_batch_dim(dummy_action) return feed_forward_policy_prior_to_actor_core(_policy_prior, dummy_action) def make_ensemble_n_step_return( n_step_return_network: mbop_networks.NStepReturnNetwork) -> NStepReturn: """Creates an ensemble n-step return model from its network.""" return functools.partial(ensemble.apply_mean, n_step_return_network.apply)
acme-master
acme/agents/jax/mbop/models.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Tests for ensemble.""" import functools from typing import Any from acme.agents.jax.mbop import ensemble from acme.jax import networks from flax import linen as nn import jax import jax.numpy as jnp import numpy as np from absl.testing import absltest class RandomFFN(nn.Module): @nn.compact def __call__(self, x): return nn.Dense(15)(x) def params_adding_ffn(x: jnp.ndarray) -> networks.FeedForwardNetwork: """Apply adds the parameters to the inputs.""" return networks.FeedForwardNetwork( init=lambda key, x=x: jax.random.uniform(key, x.shape), apply=lambda params, x: params + x) def funny_args_ffn(x: jnp.ndarray) -> networks.FeedForwardNetwork: """Apply takes additional parameters, returns `params + x + foo - bar`.""" return networks.FeedForwardNetwork( init=lambda key, x=x: jax.random.uniform(key, x.shape), apply=lambda params, x, foo, bar: params + x + foo - bar) def struct_params_adding_ffn(sx: Any) -> networks.FeedForwardNetwork: """Like params_adding_ffn, but with pytree inputs, preserves structure.""" def init_fn(key, sx=sx): return jax.tree_map(lambda x: jax.random.uniform(key, x.shape), sx) def apply_fn(params, x): return jax.tree_map(lambda p, v: p + v, params, x) return networks.FeedForwardNetwork(init=init_fn, apply=apply_fn) class EnsembleTest(absltest.TestCase): def test_ensemble_init(self): x = jnp.ones(10) # Base input wrapped_ffn = params_adding_ffn(x) rr_ensemble = ensemble.make_ensemble( wrapped_ffn, ensemble.apply_round_robin, num_networks=3) key = jax.random.PRNGKey(0) params = rr_ensemble.init(key) self.assertTupleEqual(params.shape, (3,) + x.shape) # The ensemble dimension is the lead dimension. self.assertFalse((params[0, ...] == params[1, ...]).all()) def test_apply_all(self): x = jnp.ones(10) # Base input bx = jnp.ones((7, 10)) # Batched input wrapped_ffn = params_adding_ffn(x) rr_ensemble = ensemble.make_ensemble( wrapped_ffn, ensemble.apply_all, num_networks=3) key = jax.random.PRNGKey(0) params = rr_ensemble.init(key) self.assertTupleEqual(params.shape, (3,) + x.shape) y = rr_ensemble.apply(params, x) self.assertTupleEqual(y.shape, (3,) + x.shape) np.testing.assert_allclose(params, y - jnp.broadcast_to(x, (3,) + x.shape)) by = rr_ensemble.apply(params, bx) # Note: the batch dimension is no longer the leading dimension. self.assertTupleEqual(by.shape, (3,) + bx.shape) def test_apply_round_robin(self): x = jnp.ones(10) # Base input bx = jnp.ones((7, 10)) # Batched input wrapped_ffn = params_adding_ffn(x) rr_ensemble = ensemble.make_ensemble( wrapped_ffn, ensemble.apply_round_robin, num_networks=3) key = jax.random.PRNGKey(0) params = rr_ensemble.init(key) self.assertTupleEqual(params.shape, (3,) + x.shape) y = rr_ensemble.apply(params, jnp.broadcast_to(x, (3,) + x.shape)) self.assertTupleEqual(y.shape, (3,) + x.shape) np.testing.assert_allclose(params, y - x) # Note: the ensemble dimension must lead, the batch dimension is no longer # the leading dimension. by = rr_ensemble.apply( params, jnp.broadcast_to(jnp.expand_dims(bx, axis=0), (3,) + bx.shape)) self.assertTupleEqual(by.shape, (3,) + bx.shape) # If num_networks=3, then `round_robin(params, input)[4]` should be equal # to `apply(params[1], input[4])`, etc. yy = rr_ensemble.apply(params, jnp.broadcast_to(x, (6,) + x.shape)) self.assertTupleEqual(yy.shape, (6,) + x.shape) np.testing.assert_allclose( jnp.concatenate([params, params], axis=0), yy - jnp.expand_dims(x, axis=0)) def test_apply_mean(self): x = jnp.ones(10) # Base input bx = jnp.ones((7, 10)) # Batched input wrapped_ffn = params_adding_ffn(x) rr_ensemble = ensemble.make_ensemble( wrapped_ffn, ensemble.apply_mean, num_networks=3) key = jax.random.PRNGKey(0) params = rr_ensemble.init(key) self.assertTupleEqual(params.shape, (3,) + x.shape) self.assertFalse((params[0, ...] == params[1, ...]).all()) y = rr_ensemble.apply(params, x) self.assertTupleEqual(y.shape, x.shape) np.testing.assert_allclose( jnp.mean(params, axis=0), y - x, atol=1E-5, rtol=1E-5) by = rr_ensemble.apply(params, bx) self.assertTupleEqual(by.shape, bx.shape) def test_apply_all_multiargs(self): x = jnp.ones(10) # Base input wrapped_ffn = funny_args_ffn(x) rr_ensemble = ensemble.make_ensemble( wrapped_ffn, ensemble.apply_all, num_networks=3) key = jax.random.PRNGKey(0) params = rr_ensemble.init(key) self.assertTupleEqual(params.shape, (3,) + x.shape) y = rr_ensemble.apply(params, x, 2 * x, x) self.assertTupleEqual(y.shape, (3,) + x.shape) np.testing.assert_allclose( params, y - jnp.broadcast_to(2 * x, (3,) + x.shape), atol=1E-5, rtol=1E-5) y = rr_ensemble.apply(params, x, bar=x, foo=2 * x) self.assertTupleEqual(y.shape, (3,) + x.shape) np.testing.assert_allclose( params, y - jnp.broadcast_to(2 * x, (3,) + x.shape), atol=1E-5, rtol=1E-5) def test_apply_all_structured(self): x = jnp.ones(10) sx = [(3 * x, 2 * x), 5 * x] # Base input wrapped_ffn = struct_params_adding_ffn(sx) rr_ensemble = ensemble.make_ensemble( wrapped_ffn, ensemble.apply_all, num_networks=3) key = jax.random.PRNGKey(0) params = rr_ensemble.init(key) y = rr_ensemble.apply(params, sx) ex = jnp.broadcast_to(x, (3,) + x.shape) np.testing.assert_allclose(y[0][0], params[0][0] + 3 * ex) def test_apply_round_robin_multiargs(self): x = jnp.ones(10) # Base input wrapped_ffn = funny_args_ffn(x) rr_ensemble = ensemble.make_ensemble( wrapped_ffn, ensemble.apply_round_robin, num_networks=3) key = jax.random.PRNGKey(0) params = rr_ensemble.init(key) self.assertTupleEqual(params.shape, (3,) + x.shape) ex = jnp.broadcast_to(x, (3,) + x.shape) y = rr_ensemble.apply(params, ex, 2 * ex, ex) self.assertTupleEqual(y.shape, (3,) + x.shape) np.testing.assert_allclose( params, y - jnp.broadcast_to(2 * x, (3,) + x.shape), atol=1E-5, rtol=1E-5) y = rr_ensemble.apply(params, ex, bar=ex, foo=2 * ex) self.assertTupleEqual(y.shape, (3,) + x.shape) np.testing.assert_allclose( params, y - jnp.broadcast_to(2 * x, (3,) + x.shape), atol=1E-5, rtol=1E-5) def test_apply_round_robin_structured(self): x = jnp.ones(10) sx = [(3 * x, 2 * x), 5 * x] # Base input wrapped_ffn = struct_params_adding_ffn(sx) rr_ensemble = ensemble.make_ensemble( wrapped_ffn, ensemble.apply_round_robin, num_networks=3) key = jax.random.PRNGKey(0) params = rr_ensemble.init(key) ex = jnp.broadcast_to(x, (3,) + x.shape) esx = [(3 * ex, 2 * ex), 5 * ex] y = rr_ensemble.apply(params, esx) np.testing.assert_allclose(y[0][0], params[0][0] + 3 * ex) def test_apply_mean_multiargs(self): x = jnp.ones(10) # Base input wrapped_ffn = funny_args_ffn(x) rr_ensemble = ensemble.make_ensemble( wrapped_ffn, ensemble.apply_mean, num_networks=3) key = jax.random.PRNGKey(0) params = rr_ensemble.init(key) self.assertTupleEqual(params.shape, (3,) + x.shape) y = rr_ensemble.apply(params, x, 2 * x, x) self.assertTupleEqual(y.shape, x.shape) np.testing.assert_allclose( jnp.mean(params, axis=0), y - 2 * x, atol=1E-5, rtol=1E-5) y = rr_ensemble.apply(params, x, bar=x, foo=2 * x) self.assertTupleEqual(y.shape, x.shape) np.testing.assert_allclose( jnp.mean(params, axis=0), y - 2 * x, atol=1E-5, rtol=1E-5) def test_apply_mean_structured(self): x = jnp.ones(10) sx = [(3 * x, 2 * x), 5 * x] # Base input wrapped_ffn = struct_params_adding_ffn(sx) rr_ensemble = ensemble.make_ensemble( wrapped_ffn, ensemble.apply_mean, num_networks=3) key = jax.random.PRNGKey(0) params = rr_ensemble.init(key) y = rr_ensemble.apply(params, sx) np.testing.assert_allclose( y[0][0], jnp.mean(params[0][0], axis=0) + 3 * x, atol=1E-5, rtol=1E-5) def test_round_robin_random(self): x = jnp.ones(10) # Base input bx = jnp.ones((9, 10)) # Batched input ffn = RandomFFN() wrapped_ffn = networks.FeedForwardNetwork( init=functools.partial(ffn.init, x=x), apply=ffn.apply) rr_ensemble = ensemble.make_ensemble( wrapped_ffn, ensemble.apply_round_robin, num_networks=3) key = jax.random.PRNGKey(0) params = rr_ensemble.init(key) out = rr_ensemble.apply(params, bx) # The output should be the same every 3 rows: blocks = jnp.split(out, 3, axis=0) np.testing.assert_array_equal(blocks[0], blocks[1]) np.testing.assert_array_equal(blocks[0], blocks[2]) self.assertTrue((out[0] != out[1]).any()) for i in range(9): np.testing.assert_allclose( out[i], ffn.apply(jax.tree_map(lambda p, i=i: p[i % 3], params), bx[i]), atol=1E-5, rtol=1E-5) def test_mean_random(self): x = jnp.ones(10) bx = jnp.ones((9, 10)) ffn = RandomFFN() wrapped_ffn = networks.FeedForwardNetwork( init=functools.partial(ffn.init, x=x), apply=ffn.apply) mean_ensemble = ensemble.make_ensemble( wrapped_ffn, ensemble.apply_mean, num_networks=3) key = jax.random.PRNGKey(0) params = mean_ensemble.init(key) single_output = mean_ensemble.apply(params, x) self.assertEqual(single_output.shape, (15,)) batch_output = mean_ensemble.apply(params, bx) # Make sure all rows are equal: np.testing.assert_allclose( jnp.broadcast_to(batch_output[0], batch_output.shape), batch_output, atol=1E-5, rtol=1E-5) # Check results explicitly: all_members = jnp.concatenate([ jnp.expand_dims( ffn.apply(jax.tree_map(lambda p, i=i: p[i], params), bx), axis=0) for i in range(3) ]) batch_means = jnp.mean(all_members, axis=0) np.testing.assert_allclose(batch_output, batch_means, atol=1E-5, rtol=1E-5) if __name__ == '__main__': absltest.main()
acme-master
acme/agents/jax/mbop/ensemble_test.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Provides the extended MPPI planner used in MBOP [https://arxiv.org/abs/2008.05556]. In this context, MPPI refers to Model-Predictive Path Integral control, originally introduced in "Model Predictive Path Integral Control: From Theory to Parallel Computation" Grady Williams, Andrew Aldrich and Evangelos A. Theodorou. This is a modified implementation of MPPI that adds a policy prior and n-step return extension as described in Algorithm 2 of "Model-Based Offline Planning" [https://arxiv.org/abs/2008.05556]. Notation is taken from the paper. This planner can be 'degraded' to provide both 'basic' MPPI or PDDM-style [https://arxiv.org/abs/1909.11652] planning by removing the n-step return, providing a Gaussian policy prior, or single-head ensembles. """ import dataclasses import functools from typing import Callable, Optional from acme import specs from acme.agents.jax.mbop import models from acme.jax import networks import jax from jax import random import jax.numpy as jnp # Function that takes (n_trajectories, horizon, action_dim) tensor of action # trajectories and (n_trajectories) vector of corresponding cumulative rewards, # i.e. returns, for each trajectory as input and returns a single action # trajectory. ActionAggregationFn = Callable[[jnp.ndarray, jnp.ndarray], jnp.ndarray] def return_weighted_average(action_trajectories: jnp.ndarray, cum_reward: jnp.ndarray, kappa: float) -> jnp.ndarray: r"""Calculates return-weighted average over all trajectories. This will calculate the return-weighted average over a set of trajectories as defined on l.17 of Alg. 2 in the MBOP paper: [https://arxiv.org/abs/2008.05556]. Note: Clipping will be performed for `cum_reward` values > 80 to avoid NaNs. Args: action_trajectories: (n_trajectories, horizon, action_dim) tensor of action trajectories, corresponds to `A` in Alg. 2. cum_reward: (n_trajectories) vector of corresponding cumulative rewards (returns) for each trajectory. Corresponds to `\mathcal{R}` in Alg. 2. kappa: `\kappa` constant, changes the 'peakiness' of the exponential averaging. Returns: Single action trajectory corresponding to the return-weighted average of the trajectories. """ # Substract maximum reward to avoid NaNs: cum_reward = cum_reward - cum_reward.max() # Remove the batch dimension of cum_reward allows for an implicit broadcast in # jnp.average: exp_cum_reward = jnp.exp(kappa * jnp.squeeze(cum_reward)) return jnp.average(action_trajectories, weights=exp_cum_reward, axis=0) def return_top_k_average(action_trajectories: jnp.ndarray, cum_reward: jnp.ndarray, k: int = 10) -> jnp.ndarray: r"""Calculates the top-k average over all trajectories. This will calculate the top-k average over a set of trajectories as defined in the POIR Paper: Note: top-k average is more numerically stable than the weighted average. Args: action_trajectories: (n_trajectories, horizon, action_dim) tensor of action trajectories. cum_reward: (n_trajectories) vector of corresponding cumulative rewards (returns) for each trajectory. k: the number of trajectories to average. Returns: Single action trajectory corresponding to the average of the k best trajectories. """ top_k_trajectories = action_trajectories[jnp.argsort( jnp.squeeze(cum_reward))[-int(k):]] return jnp.mean(top_k_trajectories, axis=0) @dataclasses.dataclass class MPPIConfig: """Config dataclass for MPPI-style planning, used in mppi.py. These variables correspond to different parameters of `MBOP-Trajopt` as defined in MBOP [https://arxiv.org/abs/2008.05556] (Alg. 2). Attributes: sigma: Variance of action-additive noise. beta: Mixture parameter between old trajectory and new action. horizon: Planning horizon, corresponds to H in Alg. 2 line 8. n_trajectories: Number of trajectories used in `mppi_planner` to sample the best action. Corresponds to `N` in Alg. 2 line. 5. previous_trajectory_clip: Value to clip the previous_trajectory's actions to. Disabled if None. action_aggregation_fn: Function that aggregates action trajectories and returns a single action trajectory. """ sigma: float = 0.8 beta: float = 0.2 horizon: int = 15 n_trajectories: int = 1000 previous_trajectory_clip: Optional[float] = None action_aggregation_fn: ActionAggregationFn = ( functools.partial(return_weighted_average, kappa=0.5)) def get_initial_trajectory(config: MPPIConfig, env_spec: specs.EnvironmentSpec): """Returns the initial empty trajectory `T_0`.""" return jnp.zeros((max(1, config.horizon),) + env_spec.actions.shape) def _repeat_n(new_batch: int, data: jnp.ndarray) -> jnp.ndarray: """Create new batch dimension of size `new_batch` by repeating `data`.""" return jnp.broadcast_to(data, (new_batch,) + data.shape) def mppi_planner( config: MPPIConfig, world_model: models.WorldModel, policy_prior: models.PolicyPrior, n_step_return: models.NStepReturn, world_model_params: networks.Params, policy_prior_params: networks.Params, n_step_return_params: networks.Params, random_key: networks.PRNGKey, observation: networks.Observation, previous_trajectory: jnp.ndarray, ) -> jnp.ndarray: """MPPI-extended trajectory optimizer. This implements the trajectory optimizer described in MBOP [https://arxiv.org/abs/2008.05556] (Alg. 2) which is an extended version of MPPI that adds support for arbitrary sampling distributions and extends the return horizon using an approximate model of returns. There are a couple notation changes for readability: A -> action_trajectories T -> action_trajectory If the horizon is set to 0, the planner will simply call the policy_prior and average the action over the ensemble heads. Args: config: Base configuration parameters of MPPI. world_model: Corresponds to `f_m(s_t, a_t)_s` in Alg. 2. policy_prior: Corresponds to `f_b(s_t, a_tm1)` in Alg. 2. n_step_return: Corresponds to `f_R(s_t, a_t)` in Alg. 2. world_model_params: Parameters for world model. policy_prior_params: Parameters for policy prior. n_step_return_params: Parameters for n_step return. random_key: JAX random key seed. observation: Normalized current observation from the environment, `s` in Alg. 2. previous_trajectory: Normalized previous action trajectory. `T` in Alg 2. Shape is [n_trajectories, horizon, action_dims]. Returns: jnp.ndarray: Average action trajectory of shape [horizon, action_dims]. """ action_trajectory_tm1 = previous_trajectory policy_prior_state = policy_prior.init(random_key) # Broadcast so that we have n_trajectories copies of each: observation_t = jax.tree_map( functools.partial(_repeat_n, config.n_trajectories), observation) action_tm1 = jnp.broadcast_to(action_trajectory_tm1[0], (config.n_trajectories,) + action_trajectory_tm1[0].shape) if config.previous_trajectory_clip is not None: action_tm1 = jnp.clip( action_tm1, a_min=-config.previous_trajectory_clip, a_max=config.previous_trajectory_clip) # First check if planning is unnecessary: if config.horizon == 0: if hasattr(policy_prior_state, 'action_tm1'): policy_prior_state = policy_prior_state.replace(action_tm1=action_tm1) action_set, _ = policy_prior.select_action(policy_prior_params, observation_t, policy_prior_state) # Need to re-create an action trajectory from a single action. return jnp.broadcast_to( jnp.mean(action_set, axis=0), (1, action_set.shape[-1])) # Accumulators for returns and trajectories: cum_reward = jnp.zeros((config.n_trajectories, 1)) # Generate noise once: random_key, noise_key = random.split(random_key) action_noise = config.sigma * random.normal(noise_key, ( (config.horizon,) + action_tm1.shape)) # Initialize empty set of action trajectories for concatenation in loop: action_trajectories = jnp.zeros((config.n_trajectories, 0) + action_trajectory_tm1[0].shape) for t in range(config.horizon): # Query policy prior for proposed action: if hasattr(policy_prior_state, 'action_tm1'): policy_prior_state = policy_prior_state.replace(action_tm1=action_tm1) action_t, policy_prior_state = policy_prior.select_action( policy_prior_params, observation_t, policy_prior_state) # Add action noise: action_t = action_t + action_noise[t] # Mix action with previous trajectory's corresponding action: action_t = (1 - config.beta) * action_t + config.beta * action_trajectory_tm1[t] # Query world model to get next observation and reward: observation_tp1, reward_t = world_model(world_model_params, observation_t, action_t) cum_reward += reward_t # Insert actions into trajectory matrix: action_trajectories = jnp.concatenate( [action_trajectories, jnp.expand_dims(action_t, axis=1)], axis=1) # Bump variable timesteps for next loop: observation_t = observation_tp1 action_tm1 = action_t # De-normalize and append the final n_step return prediction: n_step_return_t = n_step_return(n_step_return_params, observation_t, action_t) cum_reward += n_step_return_t # Average the set of `n_trajectories` trajectories into a single trajectory. return config.action_aggregation_fn(action_trajectories, cum_reward)
acme-master
acme/agents/jax/mbop/mppi.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Implementation of the Model-Based Offline Planning (MBOP) agent.""" from acme.agents.jax.mbop.acting import ActorCore from acme.agents.jax.mbop.acting import make_actor from acme.agents.jax.mbop.acting import make_actor_core from acme.agents.jax.mbop.acting import make_ensemble_actor_core from acme.agents.jax.mbop.builder import MBOPBuilder from acme.agents.jax.mbop.config import MBOPConfig from acme.agents.jax.mbop.dataset import EPISODE_RETURN from acme.agents.jax.mbop.dataset import episodes_to_timestep_batched_transitions from acme.agents.jax.mbop.dataset import get_normalization_stats from acme.agents.jax.mbop.dataset import N_STEP_RETURN from acme.agents.jax.mbop.learning import LoggerFn from acme.agents.jax.mbop.learning import make_ensemble_regressor_learner from acme.agents.jax.mbop.learning import MakeNStepReturnLearner from acme.agents.jax.mbop.learning import MakePolicyPriorLearner from acme.agents.jax.mbop.learning import MakeWorldModelLearner from acme.agents.jax.mbop.learning import MBOPLearner from acme.agents.jax.mbop.learning import TrainingState from acme.agents.jax.mbop.losses import MBOPLosses from acme.agents.jax.mbop.losses import policy_prior_loss from acme.agents.jax.mbop.losses import TransitionLoss from acme.agents.jax.mbop.losses import world_model_loss from acme.agents.jax.mbop.models import make_ensemble_n_step_return from acme.agents.jax.mbop.models import make_ensemble_policy_prior from acme.agents.jax.mbop.models import make_ensemble_world_model from acme.agents.jax.mbop.models import MakeNStepReturn from acme.agents.jax.mbop.models import MakePolicyPrior from acme.agents.jax.mbop.models import MakeWorldModel from acme.agents.jax.mbop.mppi import mppi_planner from acme.agents.jax.mbop.mppi import MPPIConfig from acme.agents.jax.mbop.mppi import return_top_k_average from acme.agents.jax.mbop.mppi import return_weighted_average from acme.agents.jax.mbop.networks import make_networks from acme.agents.jax.mbop.networks import make_policy_prior_network from acme.agents.jax.mbop.networks import make_world_model_network from acme.agents.jax.mbop.networks import MBOPNetworks
acme-master
acme/agents/jax/mbop/__init__.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Tests for dataset.""" from acme.agents.jax.mbop import dataset as dataset_lib import rlds from rlds.transformations import transformations_testlib import tensorflow as tf from absl.testing import absltest def sample_episode() -> rlds.Episode: """Returns a sample episode.""" steps = { rlds.OBSERVATION: [ [1, 1], [2, 2], [3, 3], [4, 4], [5, 5], ], rlds.ACTION: [[1], [2], [3], [4], [5]], rlds.REWARD: [1.0, 2.0, 3.0, 4.0, 5.0], rlds.DISCOUNT: [1, 1, 1, 1, 1], rlds.IS_FIRST: [True, False, False, False, False], rlds.IS_LAST: [False, False, False, False, True], rlds.IS_TERMINAL: [False, False, False, False, True], } return {rlds.STEPS: tf.data.Dataset.from_tensor_slices(steps)} class DatasetTest(transformations_testlib.TransformationsTest): def test_episode_to_timestep_batch(self): batched = dataset_lib.episode_to_timestep_batch( sample_episode(), return_horizon=2) # Scalars should be expanded and the n-step return should be present. Each # element of a step should be a triplet containing the previous, current and # next values of the corresponding fields. Since the return horizon is 2 and # the number of steps in the episode is 5, there can be only 2 triplets for # time steps 1 and 2. expected_steps = { rlds.OBSERVATION: [ [[1, 1], [2, 2], [3, 3]], [[2, 2], [3, 3], [4, 4]], ], rlds.ACTION: [ [[1], [2], [3]], [[2], [3], [4]], ], rlds.REWARD: [ [[1.0], [2.0], [3.0]], [[2.0], [3.0], [4.0]], ], rlds.DISCOUNT: [ [[1], [1], [1]], [[1], [1], [1]], ], rlds.IS_FIRST: [ [[True], [False], [False]], [[False], [False], [False]], ], rlds.IS_LAST: [ [[False], [False], [False]], [[False], [False], [False]], ], rlds.IS_TERMINAL: [ [[False], [False], [False]], [[False], [False], [False]], ], dataset_lib.N_STEP_RETURN: [ [[3.0], [5.0], [7.0]], [[5.0], [7.0], [9.0]], ], } self.expect_equal_datasets( batched, tf.data.Dataset.from_tensor_slices(expected_steps)) def test_episode_to_timestep_batch_episode_return(self): batched = dataset_lib.episode_to_timestep_batch( sample_episode(), return_horizon=3, calculate_episode_return=True) expected_steps = { rlds.OBSERVATION: [[[1, 1], [2, 2], [3, 3]]], rlds.ACTION: [[[1], [2], [3]]], rlds.REWARD: [[[1.0], [2.0], [3.0]]], rlds.DISCOUNT: [[[1], [1], [1]]], rlds.IS_FIRST: [[[True], [False], [False]]], rlds.IS_LAST: [[[False], [False], [False]]], rlds.IS_TERMINAL: [[[False], [False], [False]]], dataset_lib.N_STEP_RETURN: [[[6.0], [9.0], [12.0]]], # This should match to the sum of the rewards in the input. dataset_lib.EPISODE_RETURN: [[[15.0], [15.0], [15.0]]], } self.expect_equal_datasets( batched, tf.data.Dataset.from_tensor_slices(expected_steps)) def test_episode_to_timestep_batch_no_return_horizon(self): batched = dataset_lib.episode_to_timestep_batch( sample_episode(), return_horizon=1) expected_steps = { rlds.OBSERVATION: [ [[1, 1], [2, 2], [3, 3]], [[2, 2], [3, 3], [4, 4]], [[3, 3], [4, 4], [5, 5]], ], rlds.ACTION: [ [[1], [2], [3]], [[2], [3], [4]], [[3], [4], [5]], ], rlds.REWARD: [ [[1.0], [2.0], [3.0]], [[2.0], [3.0], [4.0]], [[3.0], [4.0], [5.0]], ], rlds.DISCOUNT: [ [[1], [1], [1]], [[1], [1], [1]], [[1], [1], [1]], ], rlds.IS_FIRST: [ [[True], [False], [False]], [[False], [False], [False]], [[False], [False], [False]], ], rlds.IS_LAST: [ [[False], [False], [False]], [[False], [False], [False]], [[False], [False], [True]], ], rlds.IS_TERMINAL: [ [[False], [False], [False]], [[False], [False], [False]], [[False], [False], [True]], ], # n-step return should be equal to the rewards. dataset_lib.N_STEP_RETURN: [ [[1.0], [2.0], [3.0]], [[2.0], [3.0], [4.0]], [[3.0], [4.0], [5.0]], ], } self.expect_equal_datasets( batched, tf.data.Dataset.from_tensor_slices(expected_steps)) def test_episode_to_timestep_batch_drop_return_horizon(self): steps = { rlds.OBSERVATION: [[1], [2], [3], [4], [5], [6]], rlds.REWARD: [1.0, 2.0, 3.0, 4.0, 5.0, 6.0], } episode = {rlds.STEPS: tf.data.Dataset.from_tensor_slices(steps)} batched = dataset_lib.episode_to_timestep_batch( episode, return_horizon=2, calculate_episode_return=True, drop_return_horizon=True) # The two steps of the episode should be dropped. There will be 4 steps left # and since the return horizon is 2, only a single 3-batched step should be # emitted. The episode return should be the sum of the rewards of the first # 4 steps. expected_steps = { rlds.OBSERVATION: [[[1], [2], [3]]], rlds.REWARD: [[[1.0], [2.0], [3.0]]], dataset_lib.N_STEP_RETURN: [[[3.0], [5.0], [7.0]]], dataset_lib.EPISODE_RETURN: [[[10.0], [10.0], [10.0]]], } self.expect_equal_datasets( batched, tf.data.Dataset.from_tensor_slices(expected_steps)) if __name__ == '__main__': absltest.main()
acme-master
acme/agents/jax/mbop/dataset_test.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """MBOP Builder.""" import functools from typing import Iterator, Optional from acme import core from acme import specs from acme import types from acme.agents.jax import builders from acme.agents.jax.mbop import acting from acme.agents.jax.mbop import config as mbop_config from acme.agents.jax.mbop import learning from acme.agents.jax.mbop import losses as mbop_losses from acme.agents.jax.mbop import networks as mbop_networks from acme.jax import networks as networks_lib from acme.jax import running_statistics from acme.utils import counting from acme.utils import loggers import optax class MBOPBuilder(builders.OfflineBuilder[mbop_networks.MBOPNetworks, acting.ActorCore, types.Transition]): """MBOP Builder. This builder uses ensemble regressor learners for the world model, policy prior and the n-step return models with fixed learning rates. The ensembles and the learning rate are configured in the config. """ def __init__( self, config: mbop_config.MBOPConfig, losses: mbop_losses.MBOPLosses, mean_std: Optional[running_statistics.NestedMeanStd] = None, ): """Initializes an MBOP builder. Args: config: a config with MBOP hyperparameters. losses: MBOP losses. mean_std: NestedMeanStd used to normalize the samples. """ self._config = config self._losses = losses self._mean_std = mean_std def make_learner( self, random_key: networks_lib.PRNGKey, networks: mbop_networks.MBOPNetworks, dataset: Iterator[types.Transition], logger_fn: loggers.LoggerFactory, environment_spec: specs.EnvironmentSpec, counter: Optional[counting.Counter] = None, ) -> core.Learner: """See base class.""" def make_ensemble_regressor_learner( name: str, logger_fn: loggers.LoggerFactory, counter: counting.Counter, rng_key: networks_lib.PRNGKey, iterator: Iterator[types.Transition], network: networks_lib.FeedForwardNetwork, loss: mbop_losses.TransitionLoss, ) -> core.Learner: """Creates an ensemble regressor learner.""" return learning.make_ensemble_regressor_learner( name, self._config.num_networks, logger_fn, counter, rng_key, iterator, network, loss, optax.adam(self._config.learning_rate), self._config.num_sgd_steps_per_step, ) make_world_model_learner = functools.partial( make_ensemble_regressor_learner, 'world_model') make_policy_prior_learner = functools.partial( make_ensemble_regressor_learner, 'policy_prior') make_n_step_return_learner = functools.partial( make_ensemble_regressor_learner, 'n_step_return') counter = counter or counting.Counter(time_delta=0.) return learning.MBOPLearner( networks, self._losses, dataset, random_key, logger_fn, make_world_model_learner, make_policy_prior_learner, make_n_step_return_learner, counter, ) def make_actor( self, random_key: networks_lib.PRNGKey, policy: acting.ActorCore, environment_spec: specs.EnvironmentSpec, variable_source: Optional[core.VariableSource] = None, ) -> core.Actor: """See base class.""" del environment_spec return acting.make_actor(policy, random_key, variable_source) def make_policy( self, networks: mbop_networks.MBOPNetworks, environment_spec: specs.EnvironmentSpec, evaluation: bool, ) -> acting.ActorCore: """See base class.""" return acting.make_ensemble_actor_core( networks, self._config.mppi_config, environment_spec, self._mean_std, use_round_robin=not evaluation)
acme-master
acme/agents/jax/mbop/builder.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Dataset related definitions and methods.""" import functools import itertools from typing import Iterator, Optional from acme import types from acme.jax import running_statistics import jax import jax.numpy as jnp import rlds import tensorflow as tf import tree # Keys in extras dictionary of the transitions. # Total return over n-steps. N_STEP_RETURN: str = 'n_step_return' # Total return of the episode that the transition belongs to. EPISODE_RETURN: str = 'episode_return' # Indices of the time-batched transitions. PREVIOUS: int = 0 CURRENT: int = 1 NEXT: int = 2 def _append_n_step_return(output, n_step_return): """Append n-step return to an output step.""" output[N_STEP_RETURN] = n_step_return return output def _append_episode_return(output, episode_return): """Append episode return to an output step.""" output[EPISODE_RETURN] = episode_return return output def _expand_scalars(output): """If rewards are scalar, expand them.""" return tree.map_structure(tf.experimental.numpy.atleast_1d, output) def episode_to_timestep_batch( episode: rlds.BatchedStep, return_horizon: int = 0, drop_return_horizon: bool = False, calculate_episode_return: bool = False) -> tf.data.Dataset: """Converts an episode into multi-timestep batches. Args: episode: Batched steps as provided directly by RLDS. return_horizon: int describing the horizon to which we should accumulate the return. drop_return_horizon: bool whether we should drop the last `return_horizon` steps to avoid mis-calculated returns near the end of the episode. calculate_episode_return: Whether to calculate episode return. Can be an expensive operation on datasets with many episodes. Returns: rl_dataset.DatasetType of 3-batched transitions, with scalar rewards expanded to 1D rewards This means that for every step, the corresponding elements will be a batch of size 3, with the first batched element corresponding to *_t-1, the second to *_t and the third to *_t+1, e.g. you can access the previous observation as: ``` o_tm1 = el[types.OBSERVATION][0] ``` Two additional keys can be added: 'R_t' which corresponds to the undiscounted return for horizon `return_horizon` from time t (always present), and 'R_total' which corresponds to the total return of the associated episode (if `calculate_episode_return` is True). Rewards are converted to be (at least) one-dimensional, prior to batching (to avoid ()-shaped elements). In this example, 0-valued observations correspond to o_{t-1}, 1-valued observations correspond to o_t, and 2-valued observations correspond to s_{t+1}. This same structure is true for all keys, except 'R_t' and 'R_total' which are both scalars. ``` ipdb> el[types.OBSERVATION] <tf.Tensor: shape=(3, 11), dtype=float32, numpy= array([[0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.], [1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.], [2., 2., 2., 2., 2., 2., 2., 2., 2., 2., 2.]], dtype=float32)> ``` """ steps = episode[rlds.STEPS] if drop_return_horizon: episode_length = steps.cardinality() steps = steps.take(episode_length - return_horizon) # Calculate n-step return: rewards = steps.map(lambda step: step[rlds.REWARD]) batched_rewards = rlds.transformations.batch( rewards, size=return_horizon, shift=1, stride=1, drop_remainder=True) returns = batched_rewards.map(tf.math.reduce_sum) output = tf.data.Dataset.zip((steps, returns)).map(_append_n_step_return) # Calculate total episode return for potential filtering, use total # of steps # to calculate return. if calculate_episode_return: dtype = jnp.float64 if jax.config.jax_enable_x64 else jnp.float32 # Need to redefine this here to avoid a tf.data crash. rewards = steps.map(lambda step: step[rlds.REWARD]) episode_return = rewards.reduce(dtype(0), lambda x, y: x + y) output = output.map( functools.partial( _append_episode_return, episode_return=episode_return)) output = output.map(_expand_scalars) output = rlds.transformations.batch( output, size=3, shift=1, drop_remainder=True) return output def _step_to_transition(rlds_step: rlds.BatchedStep) -> types.Transition: """Converts batched RLDS steps to batched transitions.""" return types.Transition( observation=rlds_step[rlds.OBSERVATION], action=rlds_step[rlds.ACTION], reward=rlds_step[rlds.REWARD], discount=rlds_step[rlds.DISCOUNT], # We provide next_observation if an algorithm needs it, however note that # it will only contain s_t and s_t+1, so will be one element short of all # other attributes (which contain s_t-1, s_t, s_t+1). next_observation=tree.map_structure(lambda x: x[1:], rlds_step[rlds.OBSERVATION]), extras={ N_STEP_RETURN: rlds_step[N_STEP_RETURN], }) def episodes_to_timestep_batched_transitions( episode_dataset: tf.data.Dataset, return_horizon: int = 10, drop_return_horizon: bool = False, min_return_filter: Optional[float] = None) -> tf.data.Dataset: """Process an existing dataset converting it to episode to 3-transitions. A 3-transition is an Transition with each attribute having an extra dimension of size 3, representing 3 consecutive timesteps. Each 3-step object will be in random order relative to each other. See `episode_to_timestep_batch` for more information. Args: episode_dataset: An RLDS dataset to process. return_horizon: The horizon we want calculate Monte-Carlo returns to. drop_return_horizon: Whether we should drop the last `return_horizon` steps. min_return_filter: Minimum episode return below which we drop an episode. Returns: A tf.data.Dataset of 3-transitions. """ dataset = episode_dataset.interleave( functools.partial( episode_to_timestep_batch, return_horizon=return_horizon, drop_return_horizon=drop_return_horizon, calculate_episode_return=min_return_filter is not None), num_parallel_calls=tf.data.experimental.AUTOTUNE, deterministic=False) if min_return_filter is not None: def filter_on_return(step): return step[EPISODE_RETURN][0][0] > min_return_filter dataset = dataset.filter(filter_on_return) dataset = dataset.map( _step_to_transition, num_parallel_calls=tf.data.experimental.AUTOTUNE) return dataset def get_normalization_stats( iterator: Iterator[types.Transition], num_normalization_batches: int = 50 ) -> running_statistics.RunningStatisticsState: """Precomputes normalization statistics over a fixed number of batches. The iterator should contain batches of 3-transitions, i.e. with two leading dimensions, the first one denoting the batch dimension and the second one the previous, current and next timesteps. The statistics are calculated using the data of the previous timestep. Args: iterator: Iterator of batchs of 3-transitions. num_normalization_batches: Number of batches to calculate the statistics. Returns: RunningStatisticsState containing the normalization statistics. """ # Set up normalization: example = next(iterator) unbatched_single_example = jax.tree_map(lambda x: x[0, PREVIOUS, :], example) mean_std = running_statistics.init_state(unbatched_single_example) for batch in itertools.islice(iterator, num_normalization_batches - 1): example = jax.tree_map(lambda x: x[:, PREVIOUS, :], batch) mean_std = running_statistics.update(mean_std, example) return mean_std
acme-master
acme/agents/jax/mbop/dataset.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """MBOP networks definitions.""" import dataclasses from typing import Any, Tuple from acme import specs from acme.jax import networks from acme.jax import utils import haiku as hk import jax.numpy as jnp import numpy as np # The term network is used in a general sense, e.g. for the CRR policy prior, it # will be a dataclass that encapsulates the networks used by the CRR (learner). WorldModelNetwork = Any PolicyPriorNetwork = Any NStepReturnNetwork = Any @dataclasses.dataclass class MBOPNetworks: """Container class to hold MBOP networks.""" world_model_network: WorldModelNetwork policy_prior_network: PolicyPriorNetwork n_step_return_network: NStepReturnNetwork def make_network_from_module( module: hk.Transformed, spec: specs.EnvironmentSpec) -> networks.FeedForwardNetwork: """Creates a network with dummy init arguments using the specified module. Args: module: Module that expects one batch axis and one features axis for its inputs. spec: EnvironmentSpec shapes to derive dummy inputs. Returns: FeedForwardNetwork whose `init` method only takes a random key, and `apply` takes an observation and action and produces an output. """ dummy_obs = utils.add_batch_dim(utils.zeros_like(spec.observations)) dummy_action = utils.add_batch_dim(utils.zeros_like(spec.actions)) return networks.FeedForwardNetwork( lambda key: module.init(key, dummy_obs, dummy_action), module.apply) def make_world_model_network( spec: specs.EnvironmentSpec, hidden_layer_sizes: Tuple[int, ...] = (64, 64) ) -> networks.FeedForwardNetwork: """Creates a world model network used by the agent.""" observation_size = np.prod(spec.observations.shape, dtype=int) def _world_model_fn(observation_t, action_t, is_training=False, key=None): # is_training and key allows to defined train/test dependant modules # like dropout. del is_training del key network = hk.nets.MLP(hidden_layer_sizes + (observation_size + 1,)) # World model returns both an observation and a reward. observation_tp1, reward_t = jnp.split( network(jnp.concatenate([observation_t, action_t], axis=-1)), [observation_size], axis=-1) return observation_tp1, reward_t world_model = hk.without_apply_rng(hk.transform(_world_model_fn)) return make_network_from_module(world_model, spec) def make_policy_prior_network( spec: specs.EnvironmentSpec, hidden_layer_sizes: Tuple[int, ...] = (64, 64) ) -> networks.FeedForwardNetwork: """Creates a policy prior network used by the agent.""" action_size = np.prod(spec.actions.shape, dtype=int) def _policy_prior_fn(observation_t, action_tm1, is_training=False, key=None): # is_training and key allows to defined train/test dependant modules # like dropout. del is_training del key network = hk.nets.MLP(hidden_layer_sizes + (action_size,)) # Policy prior returns an action. return network(jnp.concatenate([observation_t, action_tm1], axis=-1)) policy_prior = hk.without_apply_rng(hk.transform(_policy_prior_fn)) return make_network_from_module(policy_prior, spec) def make_n_step_return_network( spec: specs.EnvironmentSpec, hidden_layer_sizes: Tuple[int, ...] = (64, 64) ) -> networks.FeedForwardNetwork: """Creates an N-step return network used by the agent.""" def _n_step_return_fn(observation_t, action_t, is_training=False, key=None): # is_training and key allows to defined train/test dependant modules # like dropout. del is_training del key network = hk.nets.MLP(hidden_layer_sizes + (1,)) return network(jnp.concatenate([observation_t, action_t], axis=-1)) n_step_return = hk.without_apply_rng(hk.transform(_n_step_return_fn)) return make_network_from_module(n_step_return, spec) def make_networks( spec: specs.EnvironmentSpec, hidden_layer_sizes: Tuple[int, ...] = (64, 64), ) -> MBOPNetworks: """Creates networks used by the agent.""" world_model_network = make_world_model_network( spec, hidden_layer_sizes=hidden_layer_sizes) policy_prior_network = make_policy_prior_network( spec, hidden_layer_sizes=hidden_layer_sizes) n_step_return_network = make_n_step_return_network( spec, hidden_layer_sizes=hidden_layer_sizes) return MBOPNetworks( world_model_network=world_model_network, policy_prior_network=policy_prior_network, n_step_return_network=n_step_return_network)
acme-master
acme/agents/jax/mbop/networks.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Tests for the MBOP agent.""" import functools from acme import specs from acme import types from acme.agents.jax.mbop import learning from acme.agents.jax.mbop import losses as mbop_losses from acme.agents.jax.mbop import networks as mbop_networks from acme.testing import fakes from acme.utils import loggers import chex import jax import optax import rlds from absl.testing import absltest class MBOPTest(absltest.TestCase): def test_learner(self): with chex.fake_pmap_and_jit(): num_sgd_steps_per_step = 1 num_steps = 5 num_networks = 7 # Create a fake environment to test with. environment = fakes.ContinuousEnvironment( episode_length=10, bounded=True, observation_dim=3, action_dim=2) spec = specs.make_environment_spec(environment) dataset = fakes.transition_dataset(environment) # Add dummy n-step return to the transitions. def _add_dummy_n_step_return(sample): return types.Transition(*sample.data)._replace( extras={'n_step_return': 1.0}) dataset = dataset.map(_add_dummy_n_step_return) # Convert into time-batched format with previous, current and next # transitions. dataset = rlds.transformations.batch(dataset, 3) dataset = dataset.batch(8).as_numpy_iterator() # Use the default networks and losses. networks = mbop_networks.make_networks(spec) losses = mbop_losses.MBOPLosses() def logger_fn(label: str, steps_key: str): return loggers.make_default_logger(label, steps_key=steps_key) def make_learner_fn(name, logger_fn, counter, rng_key, dataset, network, loss): return learning.make_ensemble_regressor_learner(name, num_networks, logger_fn, counter, rng_key, dataset, network, loss, optax.adam(0.01), num_sgd_steps_per_step) learner = learning.MBOPLearner( networks, losses, dataset, jax.random.PRNGKey(0), logger_fn, functools.partial(make_learner_fn, 'world_model'), functools.partial(make_learner_fn, 'policy_prior'), functools.partial(make_learner_fn, 'n_step_return')) # Train the agent for _ in range(num_steps): learner.step() # Save and restore. learner_state = learner.save() learner.restore(learner_state) if __name__ == '__main__': absltest.main()
acme-master
acme/agents/jax/mbop/agent_test.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Tests for mppi.""" import functools from typing import Any from acme import specs from acme.agents.jax.mbop import ensemble from acme.agents.jax.mbop import models from acme.agents.jax.mbop import mppi from acme.jax import networks as networks_lib import jax import jax.numpy as jnp import numpy as np from absl.testing import absltest from absl.testing import parameterized def get_fake_world_model() -> networks_lib.FeedForwardNetwork: def apply(params: Any, observation_t: jnp.ndarray, action_t: jnp.ndarray): del params return observation_t, jnp.ones(( action_t.shape[0], 1, )) return networks_lib.FeedForwardNetwork(init=lambda: None, apply=apply) def get_fake_policy_prior() -> networks_lib.FeedForwardNetwork: return networks_lib.FeedForwardNetwork( init=lambda: None, apply=lambda params, observation_t, action_tm1: action_tm1) def get_fake_n_step_return() -> networks_lib.FeedForwardNetwork: def apply(params, observation_t, action_t): del params, action_t return jnp.ones((observation_t.shape[0], 1)) return networks_lib.FeedForwardNetwork(init=lambda: None, apply=apply) class WeightedAverageTests(parameterized.TestCase): @parameterized.parameters((np.array([1, 1, 1]), 1), (np.array([0, 1, 0]), 10), (np.array([-1, 1, -1]), 4), (np.array([-10, 30, 0]), -0.5)) def test_weighted_averages(self, cum_reward, kappa): """Compares method with a local version of the exp-weighted averaging.""" action_trajectories = jnp.reshape( jnp.arange(3 * 10 * 4), (3, 10, 4), order='F') averaged_trajectory = mppi.return_weighted_average( action_trajectories=action_trajectories, cum_reward=cum_reward, kappa=kappa) exp_weights = jnp.exp(kappa * cum_reward) # Verify single-value averaging lines up with the global averaging call: for i in range(10): for j in range(4): np.testing.assert_allclose( averaged_trajectory[i, j], jnp.sum(exp_weights * action_trajectories[:, i, j]) / jnp.sum(exp_weights), atol=1E-5, rtol=1E-5) class MPPITest(parameterized.TestCase): """This tests the MPPI planner to make sure it is correctly rolling out. It does not check the actual performance of the planner, as this would be a bit more complicated to set up. """ # TODO(dulacarnold): Look at how we can check this is actually finding an # optimal path through the model. def setUp(self): super().setUp() self.state_dims = 8 self.action_dims = 4 self.params = { 'world': jnp.ones((3,)), 'policy': jnp.ones((3,)), 'value': jnp.ones((3,)) } self.env_spec = specs.EnvironmentSpec( observations=specs.Array(shape=(self.state_dims,), dtype=float), actions=specs.Array(shape=(self.action_dims,), dtype=float), rewards=specs.Array(shape=(1,), dtype=float, name='reward'), discounts=specs.BoundedArray( shape=(), dtype=float, minimum=0., maximum=1., name='discount')) @parameterized.named_parameters(('NO-PLAN', 0), ('NORMAL', 10)) def test_planner_init(self, horizon: int): world_model = get_fake_world_model() rr_world_model = functools.partial(ensemble.apply_round_robin, world_model.apply) policy_prior = get_fake_policy_prior() def _rr_policy_prior(params, key, observation_t, action_tm1): del key return ensemble.apply_round_robin( policy_prior.apply, params, observation_t=observation_t, action_tm1=action_tm1) rr_policy_prior = models.feed_forward_policy_prior_to_actor_core( _rr_policy_prior, jnp.zeros((1, self.action_dims))) n_step_return = get_fake_n_step_return() n_step_return = functools.partial(ensemble.apply_mean, n_step_return.apply) config = mppi.MPPIConfig( sigma=1, beta=0.2, horizon=horizon, n_trajectories=9, action_aggregation_fn=functools.partial( mppi.return_weighted_average, kappa=1)) previous_trajectory = mppi.get_initial_trajectory(config, self.env_spec) key = jax.random.PRNGKey(0) for _ in range(5): previous_trajectory = mppi.mppi_planner( config, world_model=rr_world_model, policy_prior=rr_policy_prior, n_step_return=n_step_return, world_model_params=self.params, policy_prior_params=self.params, n_step_return_params=self.params, random_key=key, observation=jnp.ones(self.state_dims), previous_trajectory=previous_trajectory) if __name__ == '__main__': absltest.main()
acme-master
acme/agents/jax/mbop/mppi_test.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Module to provide ensembling support on top of a base network.""" import functools from typing import (Any, Callable) from acme.jax import networks import jax import jax.numpy as jnp def _split_batch_dimension(new_batch: int, data: jnp.ndarray) -> jnp.ndarray: """Splits the batch dimension and introduces new one with size `new_batch`. The result has two batch dimensions, first one of size `new_batch`, second one of size `data.shape[0]/new_batch`. It expects that `data.shape[0]` is divisible by `new_batch`. Args: new_batch: Dimension of outer batch dimension. data: jnp.ndarray to be reshaped. Returns: jnp.ndarray with extra batch dimension at start and updated second dimension. """ # The first dimension will be used for allocating to a specific ensemble # member, and the second dimension is the parallelized batch dimension, and # the remaining dimensions are passed as-is to the wrapped network. # We use Fortan (F) order so that each input batch i is allocated to # ensemble member k = i % new_batch. return jnp.reshape(data, (new_batch, -1) + data.shape[1:], order='F') def _repeat_n(new_batch: int, data: jnp.ndarray) -> jnp.ndarray: """Create new batch dimension of size `new_batch` by repeating `data`.""" return jnp.broadcast_to(data, (new_batch,) + data.shape) def ensemble_init(base_init: Callable[[networks.PRNGKey], networks.Params], num_networks: int, rnd: jnp.ndarray): """Initializes the ensemble parameters. Args: base_init: An init function that takes only a PRNGKey, if a network's init function requires other parameters such as example inputs they need to have been previously wrapped i.e. with functool.partial using kwargs. num_networks: Number of networks to generate parameters for. rnd: PRNGKey to split from when generating parameters. Returns: `params` for the set of ensemble networks. """ rnds = jax.random.split(rnd, num_networks) return jax.vmap(base_init)(rnds) def apply_round_robin(base_apply: Callable[[networks.Params, Any], Any], params: networks.Params, *args, **kwargs) -> Any: """Passes the input in a round-robin manner. The round-robin application means that each element of the input batch will be passed through a single ensemble member in a deterministic round-robin manner, i.e. element_i -> member_k where k = i % num_networks. It expects that: * `base_apply(member_params, *member_args, **member_kwargs)` is a valid call, where: * `member_params.shape = params.shape[1:]` * `member_args` and `member_kwargs` have the same structure as `args` and `kwargs`. * `params[k]` contains the params of the k-th member of the ensemble. * All jax arrays in `args` and `kwargs` have a batch dimension at axis 0 of the same size, which is divisible by `params.shape[0]`. Args: base_apply: Base network `apply` function that will be used for round-robin. NOTE -- This will not work with mutable/stateful apply functions. -- params: Model parameters. Number of networks is deduced from this. *args: Allows for arbitrary call signatures for `base_apply`. **kwargs: Allows for arbitrary call signatures for `base_apply`. Returns: pytree of the round-robin application. Output shape will be [initial_batch_size, <remaining dimensions>]. """ # `num_networks` is the size of the batch dimension in `params`. num_networks = jax.tree_util.tree_leaves(params)[0].shape[0] # Reshape args and kwargs for the round-robin: args = jax.tree_map( functools.partial(_split_batch_dimension, num_networks), args) kwargs = jax.tree_map( functools.partial(_split_batch_dimension, num_networks), kwargs) # `out.shape` is `(num_networks, initial_batch_size/num_networks, ...) out = jax.vmap(base_apply)(params, *args, **kwargs) # Reshape to [initial_batch_size, <remaining dimensions>]. Using the 'F' order # forces the original values to the last dimension. return jax.tree_map(lambda x: x.reshape((-1,) + x.shape[2:], order='F'), out) def apply_all(base_apply: Callable[[networks.Params, Any], Any], params: networks.Params, *args, **kwargs) -> Any: """Pass the input to all ensemble members. Inputs can either have a batch dimension which will get implicitly vmapped over, or can be a single vector which will get sent to all ensemble members. e.g. [<inputs_dims>] or [batch_size, <input_dims>]. Args: base_apply: Base network `apply` function that will be used for averaging. NOTE -- This will not work with mutable/stateful apply functions. -- params: Model parameters. Number of networks is deduced from this. *args: Allows for arbitrary call signatures for `base_apply`. **kwargs: Allows for arbitrary call signatures for `base_apply`. Returns: pytree of the resulting output of passing input to all ensemble members. Output shape will be [num_members, batch_size, <network output dims>]. """ # `num_networks` is the size of the batch dimension in `params`. num_networks = jax.tree_util.tree_leaves(params)[0].shape[0] args = jax.tree_map(functools.partial(_repeat_n, num_networks), args) kwargs = jax.tree_map(functools.partial(_repeat_n, num_networks), kwargs) # `out` is of shape `(num_networks, batch_size, <remaining dimensions>)`. return jax.vmap(base_apply)(params, *args, **kwargs) def apply_mean(base_apply: Callable[[networks.Params, Any], Any], params: networks.Params, *args, **kwargs) -> Any: """Calculates the mean over all ensemble members for each batch element. Args: base_apply: Base network `apply` function that will be used for averaging. NOTE -- This will not work with mutable/stateful apply functions. -- params: Model parameters. Number of networks is deduced from this. *args: Allows for arbitrary call signatures for `base_apply`. **kwargs: Allows for arbitrary call signatures for `base_apply`. Returns: pytree of the average over all ensembles for each element. Output shape will be [batch_size, <network output_dims>] """ out = apply_all(base_apply, params, *args, **kwargs) return jax.tree_map(functools.partial(jnp.mean, axis=0), out) def make_ensemble(base_network: networks.FeedForwardNetwork, ensemble_apply: Callable[..., Any], num_networks: int) -> networks.FeedForwardNetwork: return networks.FeedForwardNetwork( init=functools.partial(ensemble_init, base_network.init, num_networks), apply=functools.partial(ensemble_apply, base_network.apply))
acme-master
acme/agents/jax/mbop/ensemble.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Loss function wrappers, assuming a leading batch axis.""" import dataclasses from typing import Any, Callable, Optional, Tuple, Union from acme import types from acme.agents.jax.mbop import dataset from acme.jax import networks import jax import jax.numpy as jnp # The apply function takes an observation (and an action) as arguments, and is # usually a network with bound parameters. TransitionApplyFn = Callable[[networks.Observation, networks.Action], Any] ObservationOnlyTransitionApplyFn = Callable[[networks.Observation], Any] TransitionLoss = Callable[[ Union[TransitionApplyFn, ObservationOnlyTransitionApplyFn], types.Transition ], jnp.ndarray] def mse(a: jnp.ndarray, b: jnp.ndarray) -> jnp.ndarray: """MSE distance.""" return jnp.mean(jnp.square(a - b)) def world_model_loss(apply_fn: Callable[[networks.Observation, networks.Action], Tuple[networks.Observation, jnp.ndarray]], steps: types.Transition) -> jnp.ndarray: """Returns the loss for the world model. Args: apply_fn: applies a transition model (o_t, a_t) -> (o_t+1, r), expects the leading axis to index the batch and the second axis to index the transition triplet (t-1, t, t+1). steps: RLDS dictionary of transition triplets as prepared by `rlds_loader.episode_to_timestep_batch`. Returns: A scalar loss value as jnp.ndarray. """ observation_t = jax.tree_map(lambda obs: obs[:, dataset.CURRENT, ...], steps.observation) action_t = steps.action[:, dataset.CURRENT, ...] observation_tp1 = jax.tree_map(lambda obs: obs[:, dataset.NEXT, ...], steps.observation) reward_t = steps.reward[:, dataset.CURRENT, ...] (predicted_observation_tp1, predicted_reward_t) = apply_fn(observation_t, action_t) # predicted_* variables may have an extra outer dimension due to ensembling, # the mse loss still works due to broadcasting however. if len(observation_tp1.shape) != len(reward_t.shape): # The rewards from the transitions may not have the last singular dimension. reward_t = jnp.expand_dims(reward_t, axis=-1) return mse( jnp.concatenate([predicted_observation_tp1, predicted_reward_t], axis=-1), jnp.concatenate([observation_tp1, reward_t], axis=-1)) def policy_prior_loss( apply_fn: Callable[[networks.Observation, networks.Action], networks.Action], steps: types.Transition): """Returns the loss for the policy prior. Args: apply_fn: applies a policy prior (o_t, a_t) -> a_t+1, expects the leading axis to index the batch and the second axis to index the transition triplet (t-1, t, t+1). steps: RLDS dictionary of transition triplets as prepared by `rlds_loader.episode_to_timestep_batch`. Returns: A scalar loss value as jnp.ndarray. """ observation_t = jax.tree_map(lambda obs: obs[:, dataset.CURRENT, ...], steps.observation) action_tm1 = steps.action[:, dataset.PREVIOUS, ...] action_t = steps.action[:, dataset.CURRENT, ...] predicted_action_t = apply_fn(observation_t, action_tm1) return mse(predicted_action_t, action_t) def return_loss(apply_fn: Callable[[networks.Observation, networks.Action], jnp.ndarray], steps: types.Transition): """Returns the loss for the n-step return model. Args: apply_fn: applies an n-step return model (o_t, a_t) -> r, expects the leading axis to index the batch and the second axis to index the transition triplet (t-1, t, t+1). steps: RLDS dictionary of transition triplets as prepared by `rlds_loader.episode_to_timestep_batch`. Returns: A scalar loss value as jnp.ndarray. """ observation_t = jax.tree_map(lambda obs: obs[:, dataset.CURRENT, ...], steps.observation) action_t = steps.action[:, dataset.CURRENT, ...] n_step_return_t = steps.extras[dataset.N_STEP_RETURN][:, dataset.CURRENT, ...] predicted_n_step_return_t = apply_fn(observation_t, action_t) return mse(predicted_n_step_return_t, n_step_return_t) @dataclasses.dataclass class MBOPLosses: """Losses for the world model, policy prior and the n-step return.""" world_model_loss: Optional[TransitionLoss] = world_model_loss policy_prior_loss: Optional[TransitionLoss] = policy_prior_loss n_step_return_loss: Optional[TransitionLoss] = return_loss
acme-master
acme/agents/jax/mbop/losses.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Combined offline learning of world model, policy and N-step return.""" import dataclasses import functools import itertools import time from typing import Any, Callable, Iterator, List, Optional from acme import core from acme import types from acme.agents.jax import bc from acme.agents.jax.mbop import ensemble from acme.agents.jax.mbop import losses as mbop_losses from acme.agents.jax.mbop import networks as mbop_networks from acme.jax import networks as networks_lib from acme.jax import types as jax_types from acme.jax import utils from acme.utils import counting from acme.utils import loggers import jax import jax.numpy as jnp import optax @dataclasses.dataclass class TrainingState: """States of the world model, policy prior and n-step return learners.""" world_model: Any policy_prior: Any n_step_return: Any LoggerFn = Callable[[str, str], loggers.Logger] # Creates a world model learner. MakeWorldModelLearner = Callable[[ LoggerFn, counting.Counter, jax_types.PRNGKey, Iterator[types.Transition], mbop_networks.WorldModelNetwork, mbop_losses.TransitionLoss, ], core.Learner] # Creates a policy prior learner. MakePolicyPriorLearner = Callable[[ LoggerFn, counting.Counter, jax_types.PRNGKey, Iterator[types.Transition], mbop_networks.PolicyPriorNetwork, mbop_losses.TransitionLoss, ], core.Learner] # Creates an n-step return model learner. MakeNStepReturnLearner = Callable[[ LoggerFn, counting.Counter, jax_types.PRNGKey, Iterator[types.Transition], mbop_networks.NStepReturnNetwork, mbop_losses.TransitionLoss, ], core.Learner] def make_ensemble_regressor_learner( name: str, num_networks: int, logger_fn: loggers.LoggerFactory, counter: counting.Counter, rng_key: jnp.ndarray, iterator: Iterator[types.Transition], base_network: networks_lib.FeedForwardNetwork, loss: mbop_losses.TransitionLoss, optimizer: optax.GradientTransformation, num_sgd_steps_per_step: int, ): """Creates an ensemble regressor learner from the base network. Args: name: Name of the learner used for logging and counters. num_networks: Number of networks in the ensemble. logger_fn: Constructs a logger for a label. counter: Parent counter object. rng_key: Random key. iterator: An iterator of time-batched transitions used to train the networks. base_network: Base network for the ensemble. loss: Training loss to use. optimizer: Optax optimizer. num_sgd_steps_per_step: Number of gradient updates per step. Returns: An ensemble regressor learner. """ mbop_ensemble = ensemble.make_ensemble(base_network, ensemble.apply_all, num_networks) local_counter = counting.Counter(parent=counter, prefix=name) local_logger = logger_fn(name, local_counter.get_steps_key()) if logger_fn else None def loss_fn(networks: bc.BCNetworks, params: networks_lib.Params, key: jax_types.PRNGKey, transitions: types.Transition) -> jnp.ndarray: del key return loss( functools.partial(networks.policy_network.apply, params), transitions) bc_policy_network = bc.convert_to_bc_network(mbop_ensemble) bc_networks = bc.BCNetworks(bc_policy_network) # This is effectively a regressor learner. return bc.BCLearner( bc_networks, rng_key, loss_fn, optimizer, iterator, num_sgd_steps_per_step, logger=local_logger, counter=local_counter) class MBOPLearner(core.Learner): """Model-Based Offline Planning (MBOP) learner. See https://arxiv.org/abs/2008.05556 for more information. """ def __init__(self, networks: mbop_networks.MBOPNetworks, losses: mbop_losses.MBOPLosses, iterator: Iterator[types.Transition], rng_key: jax_types.PRNGKey, logger_fn: LoggerFn, make_world_model_learner: MakeWorldModelLearner, make_policy_prior_learner: MakePolicyPriorLearner, make_n_step_return_learner: MakeNStepReturnLearner, counter: Optional[counting.Counter] = None): """Creates an MBOP learner. Args: networks: One network per model. losses: One loss per model. iterator: An iterator of time-batched transitions used to train the networks. rng_key: Random key. logger_fn: Constructs a logger for a label. make_world_model_learner: Function to create the world model learner. make_policy_prior_learner: Function to create the policy prior learner. make_n_step_return_learner: Function to create the n-step return learner. counter: Parent counter object. """ # General learner book-keeping and loggers. self._counter = counter or counting.Counter() self._logger = logger_fn('mbop', 'steps') # Prepare iterators for the learners, to not split the data (preserve sample # efficiency). sharded_prefetching_dataset = utils.sharded_prefetch(iterator) world_model_iterator, policy_prior_iterator, n_step_return_iterator = ( itertools.tee(sharded_prefetching_dataset, 3)) world_model_key, policy_prior_key, n_step_return_key = jax.random.split( rng_key, 3) self._world_model = make_world_model_learner(logger_fn, self._counter, world_model_key, world_model_iterator, networks.world_model_network, losses.world_model_loss) self._policy_prior = make_policy_prior_learner( logger_fn, self._counter, policy_prior_key, policy_prior_iterator, networks.policy_prior_network, losses.policy_prior_loss) self._n_step_return = make_n_step_return_learner( logger_fn, self._counter, n_step_return_key, n_step_return_iterator, networks.n_step_return_network, losses.n_step_return_loss) # Start recording timestamps after the first learning step to not report # "warmup" time. self._timestamp = None self._learners = { 'world_model': self._world_model, 'policy_prior': self._policy_prior, 'n_step_return': self._n_step_return } def step(self): # Step the world model, policy learner and n-step return learners. self._world_model.step() self._policy_prior.step() self._n_step_return.step() # Compute the elapsed time. timestamp = time.time() elapsed_time = timestamp - self._timestamp if self._timestamp else 0 self._timestamp = timestamp # Increment counts and record the current time. counts = self._counter.increment(steps=1, walltime=elapsed_time) # Attempt to write the logs. self._logger.write({**counts}) def get_variables(self, names: List[str]) -> List[types.NestedArray]: variables = [] for name in names: # Variables will be prefixed by the learner names. If separator is not # found, learner_name=name, which is OK. learner_name, _, variable_name = name.partition('-') learner = self._learners[learner_name] variables.extend(learner.get_variables([variable_name])) return variables def save(self) -> TrainingState: return TrainingState( world_model=self._world_model.save(), policy_prior=self._policy_prior.save(), n_step_return=self._n_step_return.save()) def restore(self, state: TrainingState): self._world_model.restore(state.world_model) self._policy_prior.restore(state.policy_prior) self._n_step_return.restore(state.n_step_return)
acme-master
acme/agents/jax/mbop/learning.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """The MPPI-inspired JAX actor.""" from typing import List, Mapping, Optional, Tuple from acme import adders from acme import core from acme import specs from acme.agents.jax import actor_core as actor_core_lib from acme.agents.jax import actors from acme.agents.jax.mbop import models from acme.agents.jax.mbop import mppi from acme.agents.jax.mbop import networks as mbop_networks from acme.jax import networks as networks_lib from acme.jax import running_statistics from acme.jax import variable_utils import jax from jax import numpy as jnp # Recurrent state is the trajectory. Trajectory = jnp.ndarray ActorCore = actor_core_lib.ActorCore[ actor_core_lib.SimpleActorCoreRecurrentState[Trajectory], Mapping[str, jnp.ndarray]] def make_actor_core( mppi_config: mppi.MPPIConfig, world_model: models.WorldModel, policy_prior: models.PolicyPrior, n_step_return: models.NStepReturn, environment_spec: specs.EnvironmentSpec, mean_std: Optional[running_statistics.NestedMeanStd] = None, ) -> ActorCore: """Creates an actor core wrapping the MBOP-configured MPPI planner. Args: mppi_config: Planner hyperparameters. world_model: A world model. policy_prior: A policy prior. n_step_return: An n-step return. environment_spec: Used to initialize the initial trajectory data structure. mean_std: Used to undo normalization if the networks trained normalized. Returns: A recurrent actor core. """ if mean_std is not None: mean_std_observation = running_statistics.NestedMeanStd( mean=mean_std.mean.observation, std=mean_std.std.observation) mean_std_action = running_statistics.NestedMeanStd( mean=mean_std.mean.action, std=mean_std.std.action) mean_std_reward = running_statistics.NestedMeanStd( mean=mean_std.mean.reward, std=mean_std.std.reward) mean_std_n_step_return = running_statistics.NestedMeanStd( mean=mean_std.mean.extras['n_step_return'], std=mean_std.std.extras['n_step_return']) def denormalized_world_model( params: networks_lib.Params, observation_t: networks_lib.Observation, action_t: networks_lib.Action ) -> Tuple[networks_lib.Observation, networks_lib.Value]: """Denormalizes the reward for proper weighting in the planner.""" observation_tp1, normalized_reward_t = world_model( params, observation_t, action_t) reward_t = running_statistics.denormalize(normalized_reward_t, mean_std_reward) return observation_tp1, reward_t planner_world_model = denormalized_world_model def denormalized_n_step_return( params: networks_lib.Params, observation_t: networks_lib.Observation, action_t: networks_lib.Action) -> networks_lib.Value: """Denormalize the n-step return for proper weighting in the planner.""" normalized_n_step_return_t = n_step_return(params, observation_t, action_t) return running_statistics.denormalize(normalized_n_step_return_t, mean_std_n_step_return) planner_n_step_return = denormalized_n_step_return else: planner_world_model = world_model planner_n_step_return = n_step_return def recurrent_policy( params_list: List[networks_lib.Params], random_key: networks_lib.PRNGKey, observation: networks_lib.Observation, previous_trajectory: Trajectory, ) -> Tuple[networks_lib.Action, Trajectory]: # Note that splitting the random key is handled by GenericActor. if mean_std is not None: observation = running_statistics.normalize( observation, mean_std=mean_std_observation) trajectory = mppi.mppi_planner( config=mppi_config, world_model=planner_world_model, policy_prior=policy_prior, n_step_return=planner_n_step_return, world_model_params=params_list[0], policy_prior_params=params_list[1], n_step_return_params=params_list[2], random_key=random_key, observation=observation, previous_trajectory=previous_trajectory) action = trajectory[0, ...] if mean_std is not None: action = running_statistics.denormalize(action, mean_std=mean_std_action) return (action, trajectory) batched_policy = jax.vmap(recurrent_policy, in_axes=(None, None, 0, 0)) batched_policy = jax.jit(batched_policy) initial_trajectory = mppi.get_initial_trajectory( config=mppi_config, env_spec=environment_spec) initial_trajectory = jnp.expand_dims(initial_trajectory, axis=0) return actor_core_lib.batched_recurrent_to_actor_core(batched_policy, initial_trajectory) def make_ensemble_actor_core( networks: mbop_networks.MBOPNetworks, mppi_config: mppi.MPPIConfig, environment_spec: specs.EnvironmentSpec, mean_std: Optional[running_statistics.NestedMeanStd] = None, use_round_robin: bool = True, ) -> ActorCore: """Creates an actor core that uses ensemble models. Args: networks: MBOP networks. mppi_config: Planner hyperparameters. environment_spec: Used to initialize the initial trajectory data structure. mean_std: Used to undo normalization if the networks trained normalized. use_round_robin: Whether to use round robin or mean to calculate the policy prior over the ensemble members. Returns: A recurrent actor core. """ world_model = models.make_ensemble_world_model(networks.world_model_network) policy_prior = models.make_ensemble_policy_prior( networks.policy_prior_network, environment_spec, use_round_robin=use_round_robin) n_step_return = models.make_ensemble_n_step_return( networks.n_step_return_network) return make_actor_core(mppi_config, world_model, policy_prior, n_step_return, environment_spec, mean_std) def make_actor(actor_core: ActorCore, random_key: networks_lib.PRNGKey, variable_source: core.VariableSource, adder: Optional[adders.Adder] = None) -> core.Actor: """Creates an MBOP actor from an actor core. Args: actor_core: An MBOP actor core. random_key: JAX Random key. variable_source: The source to get networks parameters from. adder: An adder to add experiences to. The `extras` of the adder holds the state of the recurrent policy. If `has_extras=True` then the `extras` part returned from the recurrent policy is appended to the state before added to the adder. Returns: A recurrent actor. """ variable_client = variable_utils.VariableClient( client=variable_source, key=['world_model-policy', 'policy_prior-policy', 'n_step_return-policy']) return actors.GenericActor( actor_core, random_key, variable_client, adder, backend=None)
acme-master
acme/agents/jax/mbop/acting.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """PPO config.""" import dataclasses from acme.adders import reverb as adders_reverb import rlax @dataclasses.dataclass class R2D2Config: """Configuration options for R2D2 agent.""" discount: float = 0.997 target_update_period: int = 2500 evaluation_epsilon: float = 0. num_epsilons: int = 256 variable_update_period: int = 400 # Learner options burn_in_length: int = 40 trace_length: int = 80 sequence_period: int = 40 learning_rate: float = 1e-3 bootstrap_n: int = 5 clip_rewards: bool = False tx_pair: rlax.TxPair = rlax.SIGNED_HYPERBOLIC_PAIR # Replay options samples_per_insert_tolerance_rate: float = 0.1 samples_per_insert: float = 4.0 min_replay_size: int = 50_000 max_replay_size: int = 100_000 batch_size: int = 64 prefetch_size: int = 2 num_parallel_calls: int = 16 replay_table_name: str = adders_reverb.DEFAULT_PRIORITY_TABLE # Priority options importance_sampling_exponent: float = 0.6 priority_exponent: float = 0.9 max_priority_weight: float = 0.9
acme-master
acme/agents/jax/r2d2/config.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Implementation of an R2D2 agent.""" from acme.agents.jax.r2d2.actor import EpsilonRecurrentPolicy from acme.agents.jax.r2d2.actor import make_behavior_policy from acme.agents.jax.r2d2.builder import R2D2Builder from acme.agents.jax.r2d2.config import R2D2Config from acme.agents.jax.r2d2.learning import R2D2Learner from acme.agents.jax.r2d2.learning import R2D2ReplaySample from acme.agents.jax.r2d2.networks import make_atari_networks from acme.agents.jax.r2d2.networks import R2D2Networks
acme-master
acme/agents/jax/r2d2/__init__.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """R2D2 Builder.""" from typing import Generic, Iterator, List, Optional import acme from acme import adders from acme import core from acme import specs from acme.adders import reverb as adders_reverb from acme.adders.reverb import base as reverb_base from acme.adders.reverb import structured from acme.agents.jax import actor_core as actor_core_lib from acme.agents.jax import actors from acme.agents.jax import builders from acme.agents.jax.r2d2 import actor as r2d2_actor from acme.agents.jax.r2d2 import config as r2d2_config from acme.agents.jax.r2d2 import learning as r2d2_learning from acme.agents.jax.r2d2 import networks as r2d2_networks from acme.datasets import reverb as datasets from acme.jax import networks as networks_lib from acme.jax import utils from acme.jax import variable_utils from acme.utils import counting from acme.utils import loggers import jax import optax import reverb from reverb import structured_writer as sw import tensorflow as tf import tree # TODO(b/450949030): extrac the private functions to a library once other agents # reuse them. # TODO(b/450949030): add support to add all the final subsequences of # length < sequence_lenght at the end of the episode and pad them with zeros. # We have to check if this requires moving _zero_pad to the adder. def _build_sequence(length: int, step_spec: reverb_base.Step) -> reverb_base.Trajectory: """Constructs the sequence using only the first value of core_state.""" step_dict = step_spec._asdict() extras_dict = step_dict.pop('extras') return reverb_base.Trajectory( **tree.map_structure(lambda x: x[-length:], step_dict), extras=tree.map_structure(lambda x: x[-length], extras_dict)) def _zero_pad(sequence_length: int) -> datasets.Transform: """Adds zero padding to the right so all samples have the same length.""" def _zero_pad_transform(sample: reverb.ReplaySample) -> reverb.ReplaySample: trajectory: reverb_base.Trajectory = sample.data # Split steps and extras data (the extras won't be padded as they only # contain one element) trajectory_steps = trajectory._asdict() trajectory_extras = trajectory_steps.pop('extras') unpadded_length = len(tree.flatten(trajectory_steps)[0]) # Do nothing if the sequence is already full. if unpadded_length != sequence_length: to_pad = sequence_length - unpadded_length pad = lambda x: tf.pad(x, [[0, to_pad]] + [[0, 0]] * (len(x.shape) - 1)) trajectory_steps = tree.map_structure(pad, trajectory_steps) # Set the shape to be statically known, and checks it at runtime. def _ensure_shape(x): shape = tf.TensorShape([sequence_length]).concatenate(x.shape[1:]) return tf.ensure_shape(x, shape) trajectory_steps = tree.map_structure(_ensure_shape, trajectory_steps) return reverb.ReplaySample( info=sample.info, data=reverb_base.Trajectory( **trajectory_steps, extras=trajectory_extras)) return _zero_pad_transform def _make_adder_config(step_spec: reverb_base.Step, seq_len: int, seq_period: int) -> List[sw.Config]: return structured.create_sequence_config( step_spec=step_spec, sequence_length=seq_len, period=seq_period, end_of_episode_behavior=adders_reverb.EndBehavior.TRUNCATE, sequence_pattern=_build_sequence) class R2D2Builder(Generic[actor_core_lib.RecurrentState], builders.ActorLearnerBuilder[r2d2_networks.R2D2Networks, r2d2_actor.R2D2Policy, r2d2_learning.R2D2ReplaySample]): """R2D2 Builder. This is constructs all of the components for Recurrent Experience Replay in Distributed Reinforcement Learning (Kapturowski et al.) https://openreview.net/pdf?id=r1lyTjAqYX. """ def __init__(self, config: r2d2_config.R2D2Config): """Creates a R2D2 learner, a behavior policy and an eval actor.""" self._config = config self._sequence_length = ( self._config.burn_in_length + self._config.trace_length + 1) @property def _batch_size_per_device(self) -> int: """Splits batch size across all learner devices evenly.""" # TODO(bshahr): Using jax.device_count will not be valid when colocating # learning and inference. return self._config.batch_size // jax.device_count() def make_learner( self, random_key: networks_lib.PRNGKey, networks: r2d2_networks.R2D2Networks, dataset: Iterator[r2d2_learning.R2D2ReplaySample], logger_fn: loggers.LoggerFactory, environment_spec: specs.EnvironmentSpec, replay_client: Optional[reverb.Client] = None, counter: Optional[counting.Counter] = None, ) -> core.Learner: del environment_spec # The learner updates the parameters (and initializes them). return r2d2_learning.R2D2Learner( networks=networks, batch_size=self._batch_size_per_device, random_key=random_key, burn_in_length=self._config.burn_in_length, discount=self._config.discount, importance_sampling_exponent=( self._config.importance_sampling_exponent), max_priority_weight=self._config.max_priority_weight, target_update_period=self._config.target_update_period, iterator=dataset, optimizer=optax.adam(self._config.learning_rate), bootstrap_n=self._config.bootstrap_n, tx_pair=self._config.tx_pair, clip_rewards=self._config.clip_rewards, replay_client=replay_client, counter=counter, logger=logger_fn('learner')) def make_replay_tables( self, environment_spec: specs.EnvironmentSpec, policy: r2d2_actor.R2D2Policy, ) -> List[reverb.Table]: """Create tables to insert data into.""" dummy_actor_state = policy.init(jax.random.PRNGKey(0)) extras_spec = policy.get_extras(dummy_actor_state) step_spec = structured.create_step_spec( environment_spec=environment_spec, extras_spec=extras_spec) if self._config.samples_per_insert: samples_per_insert_tolerance = ( self._config.samples_per_insert_tolerance_rate * self._config.samples_per_insert) error_buffer = self._config.min_replay_size * samples_per_insert_tolerance limiter = reverb.rate_limiters.SampleToInsertRatio( min_size_to_sample=self._config.min_replay_size, samples_per_insert=self._config.samples_per_insert, error_buffer=error_buffer) else: limiter = reverb.rate_limiters.MinSize(self._config.min_replay_size) return [ reverb.Table( name=self._config.replay_table_name, sampler=reverb.selectors.Prioritized( self._config.priority_exponent), remover=reverb.selectors.Fifo(), max_size=self._config.max_replay_size, rate_limiter=limiter, signature=sw.infer_signature( configs=_make_adder_config(step_spec, self._sequence_length, self._config.sequence_period), step_spec=step_spec)) ] def make_dataset_iterator( self, replay_client: reverb.Client) -> Iterator[r2d2_learning.R2D2ReplaySample]: """Create a dataset iterator to use for learning/updating the agent.""" batch_size_per_learner = self._config.batch_size // jax.process_count() dataset = datasets.make_reverb_dataset( table=self._config.replay_table_name, server_address=replay_client.server_address, batch_size=self._batch_size_per_device, num_parallel_calls=None, max_in_flight_samples_per_worker=2 * batch_size_per_learner, postprocess=_zero_pad(self._sequence_length), ) return utils.multi_device_put( dataset.as_numpy_iterator(), devices=jax.local_devices(), split_fn=utils.keep_key_on_host) def make_adder( self, replay_client: reverb.Client, environment_spec: Optional[specs.EnvironmentSpec], policy: Optional[r2d2_actor.R2D2Policy]) -> Optional[adders.Adder]: """Create an adder which records data generated by the actor/environment.""" if environment_spec is None or policy is None: raise ValueError('`environment_spec` and `policy` cannot be None.') dummy_actor_state = policy.init(jax.random.PRNGKey(0)) extras_spec = policy.get_extras(dummy_actor_state) step_spec = structured.create_step_spec( environment_spec=environment_spec, extras_spec=extras_spec) return structured.StructuredAdder( client=replay_client, max_in_flight_items=5, configs=_make_adder_config(step_spec, self._sequence_length, self._config.sequence_period), step_spec=step_spec) def make_actor( self, random_key: networks_lib.PRNGKey, policy: r2d2_actor.R2D2Policy, environment_spec: specs.EnvironmentSpec, variable_source: Optional[core.VariableSource] = None, adder: Optional[adders.Adder] = None, ) -> acme.Actor: del environment_spec # Create variable client. variable_client = variable_utils.VariableClient( variable_source, key='actor_variables', update_period=self._config.variable_update_period) return actors.GenericActor( policy, random_key, variable_client, adder, backend='cpu') def make_policy(self, networks: r2d2_networks.R2D2Networks, environment_spec: specs.EnvironmentSpec, evaluation: bool = False) -> r2d2_actor.R2D2Policy: if evaluation: return r2d2_actor.get_actor_core( networks, num_epsilons=None, evaluation_epsilon=self._config.evaluation_epsilon) else: return r2d2_actor.get_actor_core(networks, self._config.num_epsilons)
acme-master
acme/agents/jax/r2d2/builder.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """R2D2 Networks.""" from acme import specs from acme.jax import networks as networks_lib R2D2Networks = networks_lib.UnrollableNetwork def make_atari_networks(env_spec: specs.EnvironmentSpec) -> R2D2Networks: """Builds default R2D2 networks for Atari games.""" def make_core_module() -> networks_lib.R2D2AtariNetwork: return networks_lib.R2D2AtariNetwork(env_spec.actions.num_values) return networks_lib.make_unrollable_network(env_spec, make_core_module)
acme-master
acme/agents/jax/r2d2/networks.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """R2D2 learner implementation.""" import functools import time from typing import Dict, Iterator, List, NamedTuple, Optional, Tuple from absl import logging import acme from acme.adders import reverb as adders from acme.agents.jax.r2d2 import networks as r2d2_networks from acme.jax import networks as networks_lib from acme.jax import utils from acme.utils import async_utils from acme.utils import counting from acme.utils import loggers import jax import jax.numpy as jnp import optax import reverb import rlax import tree _PMAP_AXIS_NAME = 'data' # This type allows splitting a sample between the host and device, which avoids # putting item keys (uint64) on device for the purposes of priority updating. R2D2ReplaySample = utils.PrefetchingSplit class TrainingState(NamedTuple): """Holds the agent's training state.""" params: networks_lib.Params target_params: networks_lib.Params opt_state: optax.OptState steps: int random_key: networks_lib.PRNGKey class R2D2Learner(acme.Learner): """R2D2 learner.""" def __init__(self, networks: r2d2_networks.R2D2Networks, batch_size: int, random_key: networks_lib.PRNGKey, burn_in_length: int, discount: float, importance_sampling_exponent: float, max_priority_weight: float, target_update_period: int, iterator: Iterator[R2D2ReplaySample], optimizer: optax.GradientTransformation, bootstrap_n: int = 5, tx_pair: rlax.TxPair = rlax.SIGNED_HYPERBOLIC_PAIR, clip_rewards: bool = False, max_abs_reward: float = 1., use_core_state: bool = True, prefetch_size: int = 2, replay_client: Optional[reverb.Client] = None, counter: Optional[counting.Counter] = None, logger: Optional[loggers.Logger] = None): """Initializes the learner.""" def loss( params: networks_lib.Params, target_params: networks_lib.Params, key_grad: networks_lib.PRNGKey, sample: reverb.ReplaySample, ) -> Tuple[jnp.ndarray, jnp.ndarray]: """Computes mean transformed N-step loss for a batch of sequences.""" # Get core state & warm it up on observations for a burn-in period. if use_core_state: # Replay core state. # NOTE: We may need to recover the type of the hk.LSTMState if the user # specifies a dynamically unrolled RNN as it will strictly enforce the # match between input/output state types. online_state = utils.maybe_recover_lstm_type( sample.data.extras.get('core_state')) else: key_grad, initial_state_rng = jax.random.split(key_grad) online_state = networks.init_recurrent_state(initial_state_rng, batch_size) target_state = online_state # Convert sample data to sequence-major format [T, B, ...]. data = utils.batch_to_sequence(sample.data) # Maybe burn the core state in. if burn_in_length: burn_obs = jax.tree_map(lambda x: x[:burn_in_length], data.observation) key_grad, key1, key2 = jax.random.split(key_grad, 3) _, online_state = networks.unroll(params, key1, burn_obs, online_state) _, target_state = networks.unroll(target_params, key2, burn_obs, target_state) # Only get data to learn on from after the end of the burn in period. data = jax.tree_map(lambda seq: seq[burn_in_length:], data) # Unroll on sequences to get online and target Q-Values. key1, key2 = jax.random.split(key_grad) online_q, _ = networks.unroll(params, key1, data.observation, online_state) target_q, _ = networks.unroll(target_params, key2, data.observation, target_state) # Get value-selector actions from online Q-values for double Q-learning. selector_actions = jnp.argmax(online_q, axis=-1) # Preprocess discounts & rewards. discounts = (data.discount * discount).astype(online_q.dtype) rewards = data.reward if clip_rewards: rewards = jnp.clip(rewards, -max_abs_reward, max_abs_reward) rewards = rewards.astype(online_q.dtype) # Get N-step transformed TD error and loss. batch_td_error_fn = jax.vmap( functools.partial( rlax.transformed_n_step_q_learning, n=bootstrap_n, tx_pair=tx_pair), in_axes=1, out_axes=1) batch_td_error = batch_td_error_fn( online_q[:-1], data.action[:-1], target_q[1:], selector_actions[1:], rewards[:-1], discounts[:-1]) batch_loss = 0.5 * jnp.square(batch_td_error).sum(axis=0) # Importance weighting. probs = sample.info.probability importance_weights = (1. / (probs + 1e-6)).astype(online_q.dtype) importance_weights **= importance_sampling_exponent importance_weights /= jnp.max(importance_weights) mean_loss = jnp.mean(importance_weights * batch_loss) # Calculate priorities as a mixture of max and mean sequence errors. abs_td_error = jnp.abs(batch_td_error).astype(online_q.dtype) max_priority = max_priority_weight * jnp.max(abs_td_error, axis=0) mean_priority = (1 - max_priority_weight) * jnp.mean(abs_td_error, axis=0) priorities = (max_priority + mean_priority) return mean_loss, priorities def sgd_step( state: TrainingState, samples: reverb.ReplaySample ) -> Tuple[TrainingState, jnp.ndarray, Dict[str, jnp.ndarray]]: """Performs an update step, averaging over pmap replicas.""" # Compute loss and gradients. grad_fn = jax.value_and_grad(loss, has_aux=True) key, key_grad = jax.random.split(state.random_key) (loss_value, priorities), gradients = grad_fn(state.params, state.target_params, key_grad, samples) # Average gradients over pmap replicas before optimizer update. gradients = jax.lax.pmean(gradients, _PMAP_AXIS_NAME) # Apply optimizer updates. updates, new_opt_state = optimizer.update(gradients, state.opt_state) new_params = optax.apply_updates(state.params, updates) # Periodically update target networks. steps = state.steps + 1 target_params = optax.periodic_update(new_params, state.target_params, # pytype: disable=wrong-arg-types # numpy-scalars steps, self._target_update_period) new_state = TrainingState( params=new_params, target_params=target_params, opt_state=new_opt_state, steps=steps, random_key=key) return new_state, priorities, {'loss': loss_value} def update_priorities( keys_and_priorities: Tuple[jnp.ndarray, jnp.ndarray]): keys, priorities = keys_and_priorities keys, priorities = tree.map_structure( # Fetch array and combine device and batch dimensions. lambda x: utils.fetch_devicearray(x).reshape((-1,) + x.shape[2:]), (keys, priorities)) replay_client.mutate_priorities( # pytype: disable=attribute-error table=adders.DEFAULT_PRIORITY_TABLE, updates=dict(zip(keys, priorities))) # Internalise components, hyperparameters, logger, counter, and methods. self._iterator = iterator self._replay_client = replay_client self._target_update_period = target_update_period self._counter = counter or counting.Counter() self._logger = logger or loggers.make_default_logger( 'learner', asynchronous=True, serialize_fn=utils.fetch_devicearray, time_delta=1., steps_key=self._counter.get_steps_key()) self._sgd_step = jax.pmap(sgd_step, axis_name=_PMAP_AXIS_NAME) self._async_priority_updater = async_utils.AsyncExecutor(update_priorities) # Initialise and internalise training state (parameters/optimiser state). random_key, key_init = jax.random.split(random_key) initial_params = networks.init(key_init) opt_state = optimizer.init(initial_params) # Log how many parameters the network has. sizes = tree.map_structure(jnp.size, initial_params) logging.info('Total number of params: %d', sum(tree.flatten(sizes.values()))) state = TrainingState( params=initial_params, target_params=initial_params, opt_state=opt_state, steps=jnp.array(0), random_key=random_key) # Replicate parameters. self._state = utils.replicate_in_all_devices(state) def step(self): prefetching_split = next(self._iterator) # The split_sample method passed to utils.sharded_prefetch specifies what # parts of the objects returned by the original iterator are kept in the # host and what parts are prefetched on-device. # In this case the host property of the prefetching split contains only the # replay keys and the device property is the prefetched full original # sample. keys = prefetching_split.host samples: reverb.ReplaySample = prefetching_split.device # Do a batch of SGD. start = time.time() self._state, priorities, metrics = self._sgd_step(self._state, samples) # Take metrics from first replica. metrics = utils.get_from_first_device(metrics) # Update our counts and record it. counts = self._counter.increment(steps=1, time_elapsed=time.time() - start) # Update priorities in replay. if self._replay_client: self._async_priority_updater.put((keys, priorities)) # Attempt to write logs. self._logger.write({**metrics, **counts}) def get_variables(self, names: List[str]) -> List[networks_lib.Params]: del names # There's only one available set of params in this agent. # Return first replica of parameters. return utils.get_from_first_device([self._state.params]) def save(self) -> TrainingState: # Serialize only the first replica of parameters and optimizer state. return utils.get_from_first_device(self._state) def restore(self, state: TrainingState): self._state = utils.replicate_in_all_devices(state)
acme-master
acme/agents/jax/r2d2/learning.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """R2D2 actor.""" from typing import Callable, Generic, Mapping, Optional, Tuple from acme import types from acme.agents.jax import actor_core as actor_core_lib from acme.agents.jax.r2d2 import config as r2d2_config from acme.agents.jax.r2d2 import networks as r2d2_networks from acme.jax import networks as networks_lib import chex import jax import jax.numpy as jnp import numpy as np import rlax Epsilon = float R2D2Extras = Mapping[str, jnp.ndarray] EpsilonRecurrentPolicy = Callable[[ networks_lib.Params, networks_lib.PRNGKey, networks_lib .Observation, actor_core_lib.RecurrentState, Epsilon ], Tuple[networks_lib.Action, actor_core_lib.RecurrentState]] @chex.dataclass(frozen=True, mappable_dataclass=False) class R2D2ActorState(Generic[actor_core_lib.RecurrentState]): rng: networks_lib.PRNGKey epsilon: jnp.ndarray recurrent_state: actor_core_lib.RecurrentState prev_recurrent_state: actor_core_lib.RecurrentState R2D2Policy = actor_core_lib.ActorCore[ R2D2ActorState[actor_core_lib.RecurrentState], R2D2Extras] def get_actor_core( networks: r2d2_networks.R2D2Networks, num_epsilons: Optional[int], evaluation_epsilon: Optional[float] = None, ) -> R2D2Policy: """Returns ActorCore for R2D2.""" if (not num_epsilons and evaluation_epsilon is None) or (num_epsilons and evaluation_epsilon): raise ValueError( 'Exactly one of `num_epsilons` or `evaluation_epsilon` must be ' f'specified. Received num_epsilon={num_epsilons} and ' f'evaluation_epsilon={evaluation_epsilon}.') def select_action(params: networks_lib.Params, observation: networks_lib.Observation, state: R2D2ActorState[actor_core_lib.RecurrentState]): rng, policy_rng = jax.random.split(state.rng) q_values, recurrent_state = networks.apply(params, policy_rng, observation, state.recurrent_state) action = rlax.epsilon_greedy(state.epsilon).sample(policy_rng, q_values) return action, R2D2ActorState( rng=rng, epsilon=state.epsilon, recurrent_state=recurrent_state, prev_recurrent_state=state.recurrent_state) def init( rng: networks_lib.PRNGKey ) -> R2D2ActorState[actor_core_lib.RecurrentState]: rng, epsilon_rng, state_rng = jax.random.split(rng, 3) if num_epsilons: epsilon = jax.random.choice(epsilon_rng, np.logspace(1, 3, num_epsilons, base=0.1)) else: epsilon = evaluation_epsilon initial_core_state = networks.init_recurrent_state(state_rng, None) return R2D2ActorState( rng=rng, epsilon=epsilon, recurrent_state=initial_core_state, prev_recurrent_state=initial_core_state) def get_extras( state: R2D2ActorState[actor_core_lib.RecurrentState]) -> R2D2Extras: return {'core_state': state.prev_recurrent_state} return actor_core_lib.ActorCore(init=init, select_action=select_action, get_extras=get_extras) # TODO(bshahr): Deprecate this in favour of R2D2Builder.make_policy. def make_behavior_policy(networks: r2d2_networks.R2D2Networks, config: r2d2_config.R2D2Config, evaluation: bool = False) -> EpsilonRecurrentPolicy: """Selects action according to the policy.""" def behavior_policy(params: networks_lib.Params, key: networks_lib.PRNGKey, observation: types.NestedArray, core_state: types.NestedArray, epsilon: float): q_values, core_state = networks.apply(params, key, observation, core_state) epsilon = config.evaluation_epsilon if evaluation else epsilon return rlax.epsilon_greedy(epsilon).sample(key, q_values), core_state return behavior_policy
acme-master
acme/agents/jax/r2d2/actor.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """ValueDice config.""" import dataclasses from acme.adders import reverb as adders_reverb @dataclasses.dataclass class ValueDiceConfig: """Configuration options for ValueDice.""" policy_learning_rate: float = 1e-5 nu_learning_rate: float = 1e-3 discount: float = .99 batch_size: int = 256 alpha: float = 0.05 policy_reg_scale: float = 1e-4 nu_reg_scale: float = 10.0 # Replay options replay_table_name: str = adders_reverb.DEFAULT_PRIORITY_TABLE samples_per_insert: float = 256 * 4 # Rate to be used for the SampleToInsertRatio rate limitter tolerance. # See a formula in make_replay_tables for more details. samples_per_insert_tolerance_rate: float = 0.1 min_replay_size: int = 1000 max_replay_size: int = 1000000 prefetch_size: int = 4 # How many gradient updates to perform per step. num_sgd_steps_per_step: int = 1
acme-master
acme/agents/jax/value_dice/config.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """ValueDice agent.""" from acme.agents.jax.value_dice.builder import ValueDiceBuilder from acme.agents.jax.value_dice.config import ValueDiceConfig from acme.agents.jax.value_dice.learning import ValueDiceLearner from acme.agents.jax.value_dice.networks import apply_policy_and_sample from acme.agents.jax.value_dice.networks import make_networks from acme.agents.jax.value_dice.networks import ValueDiceNetworks
acme-master
acme/agents/jax/value_dice/__init__.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """ValueDice agent implementation, using JAX.""" from typing import Callable, Iterator, List, Optional from acme import adders from acme import core from acme import specs from acme import types from acme.adders import reverb as adders_reverb from acme.agents.jax import actor_core as actor_core_lib from acme.agents.jax import actors from acme.agents.jax import builders from acme.agents.jax.value_dice import config as value_dice_config from acme.agents.jax.value_dice import learning from acme.agents.jax.value_dice import networks as value_dice_networks from acme.datasets import reverb as datasets from acme.jax import networks as networks_lib from acme.jax import utils from acme.jax import variable_utils from acme.utils import counting from acme.utils import loggers import jax import optax import reverb from reverb import rate_limiters class ValueDiceBuilder( builders.ActorLearnerBuilder[value_dice_networks.ValueDiceNetworks, actor_core_lib.FeedForwardPolicy, reverb.ReplaySample]): """ValueDice Builder. This builder is an entry point for online version of ValueDice. For offline please use the ValueDiceLearner directly. """ def __init__(self, config: value_dice_config.ValueDiceConfig, make_demonstrations: Callable[[int], Iterator[types.Transition]]): self._make_demonstrations = make_demonstrations self._config = config def make_learner( self, random_key: networks_lib.PRNGKey, networks: value_dice_networks.ValueDiceNetworks, dataset: Iterator[reverb.ReplaySample], logger_fn: loggers.LoggerFactory, environment_spec: specs.EnvironmentSpec, replay_client: Optional[reverb.Client] = None, counter: Optional[counting.Counter] = None, ) -> core.Learner: del environment_spec, replay_client iterator_demonstration = self._make_demonstrations( self._config.batch_size * self._config.num_sgd_steps_per_step) policy_optimizer = optax.adam( learning_rate=self._config.policy_learning_rate) nu_optimizer = optax.adam(learning_rate=self._config.nu_learning_rate) return learning.ValueDiceLearner( networks=networks, policy_optimizer=policy_optimizer, nu_optimizer=nu_optimizer, discount=self._config.discount, rng=random_key, alpha=self._config.alpha, policy_reg_scale=self._config.policy_reg_scale, nu_reg_scale=self._config.nu_reg_scale, num_sgd_steps_per_step=self._config.num_sgd_steps_per_step, iterator_replay=dataset, iterator_demonstrations=iterator_demonstration, logger=logger_fn('learner'), counter=counter, ) def make_replay_tables( self, environment_spec: specs.EnvironmentSpec, policy: actor_core_lib.FeedForwardPolicy, ) -> List[reverb.Table]: del policy samples_per_insert_tolerance = ( self._config.samples_per_insert_tolerance_rate * self._config.samples_per_insert) error_buffer = self._config.min_replay_size * samples_per_insert_tolerance limiter = rate_limiters.SampleToInsertRatio( min_size_to_sample=self._config.min_replay_size, samples_per_insert=self._config.samples_per_insert, error_buffer=error_buffer) return [reverb.Table( name=self._config.replay_table_name, sampler=reverb.selectors.Uniform(), remover=reverb.selectors.Fifo(), max_size=self._config.max_replay_size, rate_limiter=limiter, signature=adders_reverb.NStepTransitionAdder.signature( environment_spec))] def make_dataset_iterator( self, replay_client: reverb.Client) -> Iterator[reverb.ReplaySample]: """Creates a dataset iterator to use for learning.""" dataset = datasets.make_reverb_dataset( table=self._config.replay_table_name, server_address=replay_client.server_address, batch_size=( self._config.batch_size * self._config.num_sgd_steps_per_step), prefetch_size=self._config.prefetch_size) return utils.device_put(dataset.as_numpy_iterator(), jax.devices()[0]) def make_adder( self, replay_client: reverb.Client, environment_spec: Optional[specs.EnvironmentSpec], policy: Optional[actor_core_lib.FeedForwardPolicy] ) -> Optional[adders.Adder]: del environment_spec, policy return adders_reverb.NStepTransitionAdder( priority_fns={self._config.replay_table_name: None}, client=replay_client, n_step=1, discount=self._config.discount) def make_actor( self, random_key: networks_lib.PRNGKey, policy: actor_core_lib.FeedForwardPolicy, environment_spec: specs.EnvironmentSpec, variable_source: Optional[core.VariableSource] = None, adder: Optional[adders.Adder] = None, ) -> core.Actor: del environment_spec assert variable_source is not None actor_core = actor_core_lib.batched_feed_forward_to_actor_core(policy) # Inference happens on CPU, so it's better to move variables there too. variable_client = variable_utils.VariableClient(variable_source, 'policy', device='cpu') return actors.GenericActor( actor_core, random_key, variable_client, adder, backend='cpu') def make_policy(self, networks: value_dice_networks.ValueDiceNetworks, environment_spec: specs.EnvironmentSpec, evaluation: bool = False) -> actor_core_lib.FeedForwardPolicy: del environment_spec return value_dice_networks.apply_policy_and_sample( networks, eval_mode=evaluation)
acme-master
acme/agents/jax/value_dice/builder.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """ValueDice networks definition.""" import dataclasses from typing import Callable, Optional, Tuple from acme import specs from acme.agents.jax import actor_core as actor_core_lib from acme.jax import networks as networks_lib from acme.jax import utils import haiku as hk import jax import jax.numpy as jnp import numpy as np @dataclasses.dataclass class ValueDiceNetworks: """ValueDice networks.""" policy_network: networks_lib.FeedForwardNetwork nu_network: networks_lib.FeedForwardNetwork # Functions for actors and evaluators, resp., to sample actions. sample: networks_lib.SampleFn sample_eval: Optional[networks_lib.SampleFn] = None # Function that transforms an action before a mixture is applied, typically # the identity for continuous actions and one-hot encoding for discrete # actions. encode_action: Callable[[networks_lib.Action], jnp.ndarray] = lambda x: x def apply_policy_and_sample( networks: ValueDiceNetworks, eval_mode: bool = False) -> actor_core_lib.FeedForwardPolicy: """Returns a function that computes actions.""" sample_fn = networks.sample if not eval_mode else networks.sample_eval if not sample_fn: raise ValueError('sample function is not provided') def apply_and_sample(params, key, obs): return sample_fn(networks.policy_network.apply(params, obs), key) return apply_and_sample def make_networks( spec: specs.EnvironmentSpec, hidden_layer_sizes: Tuple[int, ...] = (256, 256)) -> ValueDiceNetworks: """Creates networks used by the agent.""" num_dimensions = np.prod(spec.actions.shape, dtype=int) def _actor_fn(obs): network = hk.Sequential([ hk.nets.MLP( list(hidden_layer_sizes), w_init=hk.initializers.VarianceScaling(1.0, 'fan_in', 'uniform'), activation=jax.nn.relu, activate_final=True), networks_lib.NormalTanhDistribution(num_dimensions), ]) return network(obs) def _nu_fn(obs, action): network = hk.Sequential([ hk.nets.MLP( list(hidden_layer_sizes) + [1], w_init=hk.initializers.VarianceScaling(1.0, 'fan_in', 'uniform'), activation=jax.nn.relu), ]) return network(jnp.concatenate([obs, action], axis=-1)) policy = hk.without_apply_rng(hk.transform(_actor_fn)) nu = hk.without_apply_rng(hk.transform(_nu_fn)) # Create dummy observations and actions to create network parameters. dummy_action = utils.zeros_like(spec.actions) dummy_obs = utils.zeros_like(spec.observations) dummy_action = utils.add_batch_dim(dummy_action) dummy_obs = utils.add_batch_dim(dummy_obs) return ValueDiceNetworks( policy_network=networks_lib.FeedForwardNetwork( lambda key: policy.init(key, dummy_obs), policy.apply), nu_network=networks_lib.FeedForwardNetwork( lambda key: nu.init(key, dummy_obs, dummy_action), nu.apply), sample=lambda params, key: params.sample(seed=key), sample_eval=lambda params, key: params.mode())
acme-master
acme/agents/jax/value_dice/networks.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """ValueDice learner implementation.""" import functools import time from typing import Any, Dict, Iterator, List, Mapping, NamedTuple, Optional, Tuple import acme from acme import types from acme.agents.jax.value_dice import networks as value_dice_networks from acme.jax import networks as networks_lib from acme.jax import utils from acme.utils import counting from acme.utils import loggers import jax import jax.numpy as jnp import optax import reverb class TrainingState(NamedTuple): """Contains training state for the learner.""" policy_optimizer_state: optax.OptState policy_params: networks_lib.Params nu_optimizer_state: optax.OptState nu_params: networks_lib.Params key: jnp.ndarray steps: int def _orthogonal_regularization_loss(params: networks_lib.Params): """Orthogonal regularization. See equation (3) in https://arxiv.org/abs/1809.11096. Args: params: Dictionary of parameters to apply regualization for. Returns: A regularization loss term. """ reg_loss = 0 for key in params: if isinstance(params[key], Mapping): reg_loss += _orthogonal_regularization_loss(params[key]) continue variable = params[key] assert len(variable.shape) in [1, 2, 4] if len(variable.shape) == 1: # This is a bias so do not apply regularization. continue if len(variable.shape) == 4: # CNN variable = jnp.reshape(variable, (-1, variable.shape[-1])) prod = jnp.matmul(jnp.transpose(variable), variable) reg_loss += jnp.sum(jnp.square(prod * (1 - jnp.eye(prod.shape[0])))) return reg_loss class ValueDiceLearner(acme.Learner): """ValueDice learner.""" _state: TrainingState def __init__(self, networks: value_dice_networks.ValueDiceNetworks, policy_optimizer: optax.GradientTransformation, nu_optimizer: optax.GradientTransformation, discount: float, rng: jnp.ndarray, iterator_replay: Iterator[reverb.ReplaySample], iterator_demonstrations: Iterator[types.Transition], alpha: float = 0.05, policy_reg_scale: float = 1e-4, nu_reg_scale: float = 10.0, num_sgd_steps_per_step: int = 1, counter: Optional[counting.Counter] = None, logger: Optional[loggers.Logger] = None): rng, policy_key, nu_key = jax.random.split(rng, 3) policy_init_params = networks.policy_network.init(policy_key) policy_optimizer_state = policy_optimizer.init(policy_init_params) nu_init_params = networks.nu_network.init(nu_key) nu_optimizer_state = nu_optimizer.init(nu_init_params) def compute_losses( policy_params: networks_lib.Params, nu_params: networks_lib.Params, key: jnp.ndarray, replay_o_tm1: types.NestedArray, replay_a_tm1: types.NestedArray, replay_o_t: types.NestedArray, demo_o_tm1: types.NestedArray, demo_a_tm1: types.NestedArray, demo_o_t: types.NestedArray, ) -> jnp.ndarray: # TODO(damienv, hussenot): what to do with the discounts ? def policy(obs, key): dist_params = networks.policy_network.apply(policy_params, obs) return networks.sample(dist_params, key) key1, key2, key3, key4 = jax.random.split(key, 4) # Predicted actions. demo_o_t0 = demo_o_tm1 policy_demo_a_t0 = policy(demo_o_t0, key1) policy_demo_a_t = policy(demo_o_t, key2) policy_replay_a_t = policy(replay_o_t, key3) replay_a_tm1 = networks.encode_action(replay_a_tm1) demo_a_tm1 = networks.encode_action(demo_a_tm1) policy_demo_a_t0 = networks.encode_action(policy_demo_a_t0) policy_demo_a_t = networks.encode_action(policy_demo_a_t) policy_replay_a_t = networks.encode_action(policy_replay_a_t) # "Value function" nu over the expert states. nu_demo_t0 = networks.nu_network.apply(nu_params, demo_o_t0, policy_demo_a_t0) nu_demo_tm1 = networks.nu_network.apply(nu_params, demo_o_tm1, demo_a_tm1) nu_demo_t = networks.nu_network.apply(nu_params, demo_o_t, policy_demo_a_t) nu_demo_diff = nu_demo_tm1 - discount * nu_demo_t # "Value function" nu over the replay buffer states. nu_replay_tm1 = networks.nu_network.apply(nu_params, replay_o_tm1, replay_a_tm1) nu_replay_t = networks.nu_network.apply(nu_params, replay_o_t, policy_replay_a_t) nu_replay_diff = nu_replay_tm1 - discount * nu_replay_t # Linear part of the loss. linear_loss_demo = jnp.mean(nu_demo_t0 * (1.0 - discount)) linear_loss_rb = jnp.mean(nu_replay_diff) linear_loss = (linear_loss_demo * (1 - alpha) + linear_loss_rb * alpha) # Non linear part of the loss. nu_replay_demo_diff = jnp.concatenate([nu_demo_diff, nu_replay_diff], axis=0) replay_demo_weights = jnp.concatenate([ jnp.ones_like(nu_demo_diff) * (1 - alpha), jnp.ones_like(nu_replay_diff) * alpha ], axis=0) replay_demo_weights /= jnp.mean(replay_demo_weights) non_linear_loss = jnp.sum( jax.lax.stop_gradient( utils.weighted_softmax(nu_replay_demo_diff, replay_demo_weights, axis=0)) * nu_replay_demo_diff) # Final loss. loss = (non_linear_loss - linear_loss) # Regularized policy loss. if policy_reg_scale > 0.: policy_reg = _orthogonal_regularization_loss(policy_params) else: policy_reg = 0. # Gradient penality on nu if nu_reg_scale > 0.0: batch_size = demo_o_tm1.shape[0] c = jax.random.uniform(key4, shape=(batch_size,)) shape_o = [ dim if i == 0 else 1 for i, dim in enumerate(replay_o_tm1.shape) ] shape_a = [ dim if i == 0 else 1 for i, dim in enumerate(replay_a_tm1.shape) ] c_o = jnp.reshape(c, shape_o) c_a = jnp.reshape(c, shape_a) mixed_o_tm1 = c_o * demo_o_tm1 + (1 - c_o) * replay_o_tm1 mixed_a_tm1 = c_a * demo_a_tm1 + (1 - c_a) * replay_a_tm1 mixed_o_t = c_o * demo_o_t + (1 - c_o) * replay_o_t mixed_policy_a_t = c_a * policy_demo_a_t + (1 - c_a) * policy_replay_a_t mixed_o = jnp.concatenate([mixed_o_tm1, mixed_o_t], axis=0) mixed_a = jnp.concatenate([mixed_a_tm1, mixed_policy_a_t], axis=0) def sum_nu(o, a): return jnp.sum(networks.nu_network.apply(nu_params, o, a)) nu_grad_o_fn = jax.grad(sum_nu, argnums=0) nu_grad_a_fn = jax.grad(sum_nu, argnums=1) nu_grad_o = nu_grad_o_fn(mixed_o, mixed_a) nu_grad_a = nu_grad_a_fn(mixed_o, mixed_a) nu_grad = jnp.concatenate([ jnp.reshape(nu_grad_o, [batch_size, -1]), jnp.reshape(nu_grad_a, [batch_size, -1])], axis=-1) # TODO(damienv, hussenot): check for the need of eps # (like in the original value dice code). nu_grad_penalty = jnp.mean( jnp.square( jnp.linalg.norm(nu_grad + 1e-8, axis=-1, keepdims=True) - 1)) else: nu_grad_penalty = 0.0 policy_loss = -loss + policy_reg_scale * policy_reg nu_loss = loss + nu_reg_scale * nu_grad_penalty return policy_loss, nu_loss # pytype: disable=bad-return-type # jax-ndarray def sgd_step( state: TrainingState, data: Tuple[types.Transition, types.Transition] ) -> Tuple[TrainingState, Dict[str, jnp.ndarray]]: replay_transitions, demo_transitions = data key, key_loss = jax.random.split(state.key) compute_losses_with_input = functools.partial( compute_losses, replay_o_tm1=replay_transitions.observation, replay_a_tm1=replay_transitions.action, replay_o_t=replay_transitions.next_observation, demo_o_tm1=demo_transitions.observation, demo_a_tm1=demo_transitions.action, demo_o_t=demo_transitions.next_observation, key=key_loss) (policy_loss_value, nu_loss_value), vjpfun = jax.vjp( compute_losses_with_input, state.policy_params, state.nu_params) policy_gradients, _ = vjpfun((1.0, 0.0)) _, nu_gradients = vjpfun((0.0, 1.0)) # Update optimizers. policy_update, policy_optimizer_state = policy_optimizer.update( policy_gradients, state.policy_optimizer_state) policy_params = optax.apply_updates(state.policy_params, policy_update) nu_update, nu_optimizer_state = nu_optimizer.update( nu_gradients, state.nu_optimizer_state) nu_params = optax.apply_updates(state.nu_params, nu_update) new_state = TrainingState( policy_optimizer_state=policy_optimizer_state, policy_params=policy_params, nu_optimizer_state=nu_optimizer_state, nu_params=nu_params, key=key, steps=state.steps + 1, ) metrics = { 'policy_loss': policy_loss_value, 'nu_loss': nu_loss_value, } return new_state, metrics # General learner book-keeping and loggers. self._counter = counter or counting.Counter() self._logger = logger or loggers.make_default_logger( 'learner', asynchronous=True, serialize_fn=utils.fetch_devicearray, steps_key=self._counter.get_steps_key()) # Iterator on demonstration transitions. self._iterator_demonstrations = iterator_demonstrations self._iterator_replay = iterator_replay self._sgd_step = jax.jit(utils.process_multiple_batches( sgd_step, num_sgd_steps_per_step)) # Create initial state. self._state = TrainingState( policy_optimizer_state=policy_optimizer_state, policy_params=policy_init_params, nu_optimizer_state=nu_optimizer_state, nu_params=nu_init_params, key=rng, steps=0, ) # Do not record timestamps until after the first learning step is done. # This is to avoid including the time it takes for actors to come online and # fill the replay buffer. self._timestamp = None def step(self): # Get data from replay (dropping extras if any). Note there is no # extra data here because we do not insert any into Reverb. # TODO(raveman): Add a support for offline training, where we do not consume # data from the replay buffer. sample = next(self._iterator_replay) replay_transitions = types.Transition(*sample.data) # Get a batch of Transitions from the demonstration. demonstration_transitions = next(self._iterator_demonstrations) self._state, metrics = self._sgd_step( self._state, (replay_transitions, demonstration_transitions)) # Compute elapsed time. timestamp = time.time() elapsed_time = timestamp - self._timestamp if self._timestamp else 0 self._timestamp = timestamp # Increment counts and record the current time counts = self._counter.increment(steps=1, walltime=elapsed_time) # Attempts to write the logs. self._logger.write({**metrics, **counts}) def get_variables(self, names: List[str]) -> List[Any]: variables = { 'policy': self._state.policy_params, 'nu': self._state.nu_params, } return [variables[name] for name in names] def save(self) -> TrainingState: return self._state def restore(self, state: TrainingState): self._state = state
acme-master
acme/agents/jax/value_dice/learning.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """PPO config.""" import dataclasses from typing import Callable, Union, Optional from acme import types from acme.adders import reverb as adders_reverb from acme.agents.jax.ppo import normalization @dataclasses.dataclass class PPOConfig: """Configuration options for PPO. Attributes: unroll_length: Length of sequences added to the replay buffer. num_minibatches: The number of minibatches to split an epoch into. i.e. minibatch size = batch_size * unroll_length / num_minibatches. num_epochs: How many times to loop over the set of minibatches. batch_size: Number of trajectory segments of length unroll_length to gather for use in a call to the learner's step function. replay_table_name: Replay table name. ppo_clipping_epsilon: PPO clipping epsilon. normalize_advantage: Whether to normalize the advantages in the batch. normalize_value: Whether the critic should predict normalized values. normalization_ema_tau: Float tau for the exponential moving average used to maintain statistics for normalizing advantages and values. clip_value: Whether to clip the values as described in "What Matters in On-Policy Reinforcement Learning?". value_clipping_epsilon: Epsilon for value clipping. max_abs_reward: If provided clips the rewards in the trajectory to have absolute value less than or equal to max_abs_reward. gae_lambda: Lambda parameter in Generalized Advantage Estimation. discount: Discount factor. learning_rate: Learning rate for updating the policy and critic networks. adam_epsilon: Adam epsilon parameter. entropy_cost: Weight of the entropy regularizer term in policy optimization. value_cost: Weight of the value loss term in optimization. max_gradient_norm: Threshold for clipping the gradient norm. variable_update_period: Determines how frequently actors pull the parameters from the learner. log_global_norm_metrics: Whether to log global norm of gradients and updates. metrics_logging_period: How often metrics should be aggregated to host and logged. pmap_axis_name: The name of the axis used for pmapping obs_normalization_fns_factory: The factory used for create observation normalization functions. Setting to None (default) disables observation normalization. """ unroll_length: int = 8 num_minibatches: int = 8 num_epochs: int = 2 batch_size: int = 256 replay_table_name: str = adders_reverb.DEFAULT_PRIORITY_TABLE ppo_clipping_epsilon: float = 0.2 normalize_advantage: bool = False normalize_value: bool = False normalization_ema_tau: float = 0.995 clip_value: bool = False value_clipping_epsilon: float = 0.2 max_abs_reward: Optional[float] = None gae_lambda: float = 0.95 discount: float = 0.99 learning_rate: Union[float, Callable[[int], float]] = 3e-4 adam_epsilon: float = 1e-7 entropy_cost: float = 3e-4 value_cost: float = 1. max_gradient_norm: float = 0.5 variable_update_period: int = 1 log_global_norm_metrics: bool = False metrics_logging_period: int = 100 pmap_axis_name: str = 'devices' obs_normalization_fns_factory: Optional[Callable[ [types.NestedSpec], normalization.NormalizationFns]] = None
acme-master
acme/agents/jax/ppo/config.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """PPO agent.""" from acme.agents.jax.ppo.builder import PPOBuilder from acme.agents.jax.ppo.config import PPOConfig from acme.agents.jax.ppo.learning import PPOLearner from acme.agents.jax.ppo.networks import EntropyFn from acme.agents.jax.ppo.networks import make_categorical_ppo_networks from acme.agents.jax.ppo.networks import make_continuous_networks from acme.agents.jax.ppo.networks import make_discrete_networks from acme.agents.jax.ppo.networks import make_inference_fn from acme.agents.jax.ppo.networks import make_mvn_diag_ppo_networks from acme.agents.jax.ppo.networks import make_networks from acme.agents.jax.ppo.networks import make_ppo_networks from acme.agents.jax.ppo.networks import make_tanh_normal_ppo_networks from acme.agents.jax.ppo.networks import PPONetworks from acme.agents.jax.ppo.normalization import build_ema_mean_std_normalizer from acme.agents.jax.ppo.normalization import build_mean_std_normalizer from acme.agents.jax.ppo.normalization import NormalizationFns from acme.agents.jax.ppo.normalization import NormalizedGenericActor
acme-master
acme/agents/jax/ppo/__init__.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """PPO Builder.""" from typing import Iterator, List, Optional from acme import adders from acme import core from acme import specs from acme.adders import reverb as adders_reverb from acme.agents.jax import actor_core as actor_core_lib from acme.agents.jax import actors from acme.agents.jax import builders from acme.agents.jax.ppo import config as ppo_config from acme.agents.jax.ppo import learning from acme.agents.jax.ppo import networks as ppo_networks from acme.agents.jax.ppo import normalization from acme.jax import networks as networks_lib from acme.jax import utils from acme.jax import variable_utils from acme.utils import counting from acme.utils import loggers import jax import numpy as np import optax import reverb class PPOBuilder( builders.ActorLearnerBuilder[ppo_networks.PPONetworks, actor_core_lib.FeedForwardPolicyWithExtra, reverb.ReplaySample]): """PPO Builder.""" def __init__( self, config: ppo_config.PPOConfig, ): """Creates PPO builder.""" self._config = config # An extra step is used for bootstrapping when computing advantages. self._sequence_length = config.unroll_length + 1 def make_replay_tables( self, environment_spec: specs.EnvironmentSpec, policy: actor_core_lib.FeedForwardPolicyWithExtra, ) -> List[reverb.Table]: """Creates reverb tables for the algorithm.""" del policy # params_num_sgd_steps is used to track how old the actor parameters are extra_spec = { 'log_prob': np.ones(shape=(), dtype=np.float32), 'params_num_sgd_steps': np.ones(shape=(), dtype=np.float32), } signature = adders_reverb.SequenceAdder.signature( environment_spec, extra_spec, sequence_length=self._sequence_length) return [ reverb.Table.queue( name=self._config.replay_table_name, max_size=self._config.batch_size, signature=signature) ] def make_dataset_iterator( self, replay_client: reverb.Client) -> Iterator[reverb.ReplaySample]: """Creates a dataset. The iterator batch size is computed as follows: Let: B := learner batch size (config.batch_size) H := number of hosts (jax.process_count()) D := number of local devices per host The Reverb iterator will load batches of size B // (H * D). After wrapping the iterator with utils.multi_device_put, this will result in an iterable that provides B // H samples per item, with B // (H * D) samples placed on each local device. In a multi-host setup, each host has its own learner node and builds its own instance of the iterator. This will result in a total batch size of H * (B // H) == B being consumed per learner step (since the learner is pmapped across all devices). Note that jax.device_count() returns the total number of devices across hosts, i.e. H * D. Args: replay_client: the reverb replay client Returns: A replay buffer iterator to be used by the local devices. """ iterator_batch_size, ragged = divmod(self._config.batch_size, jax.device_count()) if ragged: raise ValueError( 'Learner batch size must be divisible by total number of devices!') # We don't use datasets.make_reverb_dataset() here to avoid interleaving # and prefetching, that doesn't work well with can_sample() check on update. # NOTE: Value for max_in_flight_samples_per_worker comes from a # recommendation here: https://git.io/JYzXB dataset = reverb.TrajectoryDataset.from_table_signature( server_address=replay_client.server_address, table=self._config.replay_table_name, max_in_flight_samples_per_worker=( 2 * self._config.batch_size // jax.process_count() ), ) dataset = dataset.batch(iterator_batch_size, drop_remainder=True) dataset = dataset.as_numpy_iterator() return utils.multi_device_put(iterable=dataset, devices=jax.local_devices()) def make_adder( self, replay_client: reverb.Client, environment_spec: Optional[specs.EnvironmentSpec], policy: Optional[actor_core_lib.FeedForwardPolicyWithExtra], ) -> Optional[adders.Adder]: """Creates an adder which handles observations.""" del environment_spec, policy # Note that the last transition in the sequence is used for bootstrapping # only and is ignored otherwise. So we need to make sure that sequences # overlap on one transition, thus "-1" in the period length computation. return adders_reverb.SequenceAdder( client=replay_client, priority_fns={self._config.replay_table_name: None}, period=self._sequence_length - 1, sequence_length=self._sequence_length, ) def make_learner( self, random_key: networks_lib.PRNGKey, networks: ppo_networks.PPONetworks, dataset: Iterator[reverb.ReplaySample], logger_fn: loggers.LoggerFactory, environment_spec: specs.EnvironmentSpec, replay_client: Optional[reverb.Client] = None, counter: Optional[counting.Counter] = None, ) -> core.Learner: del replay_client if callable(self._config.learning_rate): optimizer = optax.chain( optax.clip_by_global_norm(self._config.max_gradient_norm), optax.scale_by_adam(eps=self._config.adam_epsilon), optax.scale_by_schedule(self._config.learning_rate), optax.scale(-1)) # pytype: disable=wrong-arg-types # numpy-scalars else: optimizer = optax.chain( optax.clip_by_global_norm(self._config.max_gradient_norm), optax.scale_by_adam(eps=self._config.adam_epsilon), optax.scale(-self._config.learning_rate)) obs_normalization_fns = None if self._config.obs_normalization_fns_factory is not None: obs_normalization_fns = self._config.obs_normalization_fns_factory( environment_spec.observations) return learning.PPOLearner( ppo_networks=networks, iterator=dataset, discount=self._config.discount, entropy_cost=self._config.entropy_cost, value_cost=self._config.value_cost, ppo_clipping_epsilon=self._config.ppo_clipping_epsilon, normalize_advantage=self._config.normalize_advantage, normalize_value=self._config.normalize_value, normalization_ema_tau=self._config.normalization_ema_tau, clip_value=self._config.clip_value, value_clipping_epsilon=self._config.value_clipping_epsilon, max_abs_reward=self._config.max_abs_reward, gae_lambda=self._config.gae_lambda, counter=counter, random_key=random_key, optimizer=optimizer, num_epochs=self._config.num_epochs, num_minibatches=self._config.num_minibatches, logger=logger_fn('learner'), log_global_norm_metrics=self._config.log_global_norm_metrics, metrics_logging_period=self._config.metrics_logging_period, pmap_axis_name=self._config.pmap_axis_name, obs_normalization_fns=obs_normalization_fns, ) def make_actor( self, random_key: networks_lib.PRNGKey, policy: actor_core_lib.FeedForwardPolicyWithExtra, environment_spec: specs.EnvironmentSpec, variable_source: Optional[core.VariableSource] = None, adder: Optional[adders.Adder] = None, ) -> core.Actor: assert variable_source is not None actor_core = actor_core_lib.batched_feed_forward_with_extras_to_actor_core( policy) if self._config.obs_normalization_fns_factory is not None: variable_client = variable_utils.VariableClient( variable_source, ['params', 'obs_normalization_params'], device='cpu', update_period=self._config.variable_update_period) obs_normalization_fns = self._config.obs_normalization_fns_factory( environment_spec.observations) actor = normalization.NormalizedGenericActor( actor_core, obs_normalization_fns, random_key, variable_client, adder, jit=True, backend='cpu', per_episode_update=False, ) else: variable_client = variable_utils.VariableClient( variable_source, 'params', device='cpu', update_period=self._config.variable_update_period) actor = actors.GenericActor( actor_core, random_key, variable_client, adder, backend='cpu') return actor def make_policy( self, networks: ppo_networks.PPONetworks, environment_spec: specs.EnvironmentSpec, evaluation: bool = False) -> actor_core_lib.FeedForwardPolicyWithExtra: del environment_spec return ppo_networks.make_inference_fn(networks, evaluation)
acme-master
acme/agents/jax/ppo/builder.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """PPO network definitions.""" import dataclasses from typing import Callable, NamedTuple, Optional, Sequence from acme import specs from acme.agents.jax import actor_core as actor_core_lib from acme.jax import networks as networks_lib from acme.jax import utils import haiku as hk import jax import jax.numpy as jnp import numpy as np import tensorflow_probability tfp = tensorflow_probability.substrates.jax tfd = tfp.distributions EntropyFn = Callable[ [networks_lib.Params, networks_lib.PRNGKey], networks_lib.Entropy ] class MVNDiagParams(NamedTuple): """Parameters for a diagonal multi-variate normal distribution.""" loc: jnp.ndarray scale_diag: jnp.ndarray class TanhNormalParams(NamedTuple): """Parameters for a tanh squashed diagonal MVN distribution.""" loc: jnp.ndarray scale: jnp.ndarray class CategoricalParams(NamedTuple): """Parameters for a categorical distribution.""" logits: jnp.ndarray class PPOParams(NamedTuple): model_params: networks_lib.Params # Using float32 as it covers a larger range than int32. If using int64 we # would need to do jax_enable_x64. num_sgd_steps: jnp.float32 @dataclasses.dataclass class PPONetworks: """Network and pure functions for the PPO agent. If 'network' returns tfd.Distribution, you can use make_ppo_networks() to create this object properly. If one is building this object manually, one has a freedom to make 'network' object return anything that is later being passed as input to log_prob/entropy/sample functions to perform the corresponding computations. An example scenario where you would want to do this due to tfd.Distribution not playing nice with jax.vmap. Please refer to the make_continuous_networks() for an example where the network does not return a tfd.Distribution object. """ network: networks_lib.FeedForwardNetwork log_prob: networks_lib.LogProbFn entropy: EntropyFn sample: networks_lib.SampleFn sample_eval: Optional[networks_lib.SampleFn] = None def make_inference_fn( ppo_networks: PPONetworks, evaluation: bool = False) -> actor_core_lib.FeedForwardPolicyWithExtra: """Returns a function to be used for inference by a PPO actor.""" def inference( params: networks_lib.Params, key: networks_lib.PRNGKey, observations: networks_lib.Observation, ): dist_params, _ = ppo_networks.network.apply(params.model_params, observations) if evaluation and ppo_networks.sample_eval: actions = ppo_networks.sample_eval(dist_params, key) else: actions = ppo_networks.sample(dist_params, key) if evaluation: return actions, {} log_prob = ppo_networks.log_prob(dist_params, actions) extras = { 'log_prob': log_prob, # Add batch dimension. 'params_num_sgd_steps': params.num_sgd_steps[None, ...] } return actions, extras return inference def make_networks( spec: specs.EnvironmentSpec, hidden_layer_sizes: Sequence[int] = (256, 256) ) -> PPONetworks: if isinstance(spec.actions, specs.DiscreteArray): return make_discrete_networks(spec, hidden_layer_sizes) else: return make_continuous_networks( spec, policy_layer_sizes=hidden_layer_sizes, value_layer_sizes=hidden_layer_sizes) def make_ppo_networks(network: networks_lib.FeedForwardNetwork) -> PPONetworks: """Constructs a PPONetworks instance from the given FeedForwardNetwork. This method assumes that the network returns a tfd.Distribution. Sometimes it may be preferable to have networks that do not return tfd.Distribution objects, for example, due to tfd.Distribution not playing nice with jax.vmap. Please refer to the make_continuous_networks() for an example where the network does not return a tfd.Distribution object. Args: network: a transformed Haiku network that takes in observations and returns the action distribution and value. Returns: A PPONetworks instance with pure functions wrapping the input network. """ return PPONetworks( network=network, log_prob=lambda distribution, action: distribution.log_prob(action), entropy=lambda distribution, key=None: distribution.entropy(), sample=lambda distribution, key: distribution.sample(seed=key), sample_eval=lambda distribution, key: distribution.mode()) def make_mvn_diag_ppo_networks( network: networks_lib.FeedForwardNetwork) -> PPONetworks: """Constructs a PPONetworks for MVN Diag policy from the FeedForwardNetwork. Args: network: a transformed Haiku network (or equivalent in other libraries) that takes in observations and returns the action distribution and value. Returns: A PPONetworks instance with pure functions wrapping the input network. """ def log_prob(params: MVNDiagParams, action): return tfd.MultivariateNormalDiag( loc=params.loc, scale_diag=params.scale_diag).log_prob(action) def entropy( params: MVNDiagParams, key: networks_lib.PRNGKey ) -> networks_lib.Entropy: del key return tfd.MultivariateNormalDiag( loc=params.loc, scale_diag=params.scale_diag).entropy() def sample(params: MVNDiagParams, key: networks_lib.PRNGKey): return tfd.MultivariateNormalDiag( loc=params.loc, scale_diag=params.scale_diag).sample(seed=key) def sample_eval(params: MVNDiagParams, key: networks_lib.PRNGKey): del key return tfd.MultivariateNormalDiag( loc=params.loc, scale_diag=params.scale_diag).mode() return PPONetworks( network=network, log_prob=log_prob, entropy=entropy, sample=sample, sample_eval=sample_eval) def make_tanh_normal_ppo_networks( network: networks_lib.FeedForwardNetwork) -> PPONetworks: """Constructs a PPONetworks for Tanh MVN Diag policy from the FeedForwardNetwork. Args: network: a transformed Haiku network (or equivalent in other libraries) that takes in observations and returns the action distribution and value. Returns: A PPONetworks instance with pure functions wrapping the input network. """ def build_distribution(params: TanhNormalParams): distribution = tfd.Normal(loc=params.loc, scale=params.scale) distribution = tfd.Independent( networks_lib.TanhTransformedDistribution(distribution), reinterpreted_batch_ndims=1) return distribution def log_prob(params: TanhNormalParams, action): distribution = build_distribution(params) return distribution.log_prob(action) def entropy( params: TanhNormalParams, key: networks_lib.PRNGKey ) -> networks_lib.Entropy: distribution = build_distribution(params) return distribution.entropy(seed=key) def sample(params: TanhNormalParams, key: networks_lib.PRNGKey): distribution = build_distribution(params) return distribution.sample(seed=key) def sample_eval(params: TanhNormalParams, key: networks_lib.PRNGKey): del key distribution = build_distribution(params) return distribution.mode() return PPONetworks( network=network, log_prob=log_prob, entropy=entropy, sample=sample, sample_eval=sample_eval) def make_discrete_networks( environment_spec: specs.EnvironmentSpec, hidden_layer_sizes: Sequence[int] = (512,), use_conv: bool = True, ) -> PPONetworks: """Creates networks used by the agent for discrete action environments. Args: environment_spec: Environment spec used to define number of actions. hidden_layer_sizes: Network definition. use_conv: Whether to use a conv or MLP feature extractor. Returns: PPONetworks """ num_actions = environment_spec.actions.num_values def forward_fn(inputs): layers = [] if use_conv: layers.extend([networks_lib.AtariTorso()]) layers.extend([hk.nets.MLP(hidden_layer_sizes, activate_final=True)]) trunk = hk.Sequential(layers) h = utils.batch_concat(inputs) h = trunk(h) logits = hk.Linear(num_actions)(h) values = hk.Linear(1)(h) values = jnp.squeeze(values, axis=-1) return (CategoricalParams(logits=logits), values) forward_fn = hk.without_apply_rng(hk.transform(forward_fn)) dummy_obs = utils.zeros_like(environment_spec.observations) dummy_obs = utils.add_batch_dim(dummy_obs) # Dummy 'sequence' dim. network = networks_lib.FeedForwardNetwork( lambda rng: forward_fn.init(rng, dummy_obs), forward_fn.apply) # Create PPONetworks to add functionality required by the agent. return make_categorical_ppo_networks(network) # pylint:disable=undefined-variable def make_categorical_ppo_networks( network: networks_lib.FeedForwardNetwork) -> PPONetworks: """Constructs a PPONetworks for Categorical Policy from FeedForwardNetwork. Args: network: a transformed Haiku network (or equivalent in other libraries) that takes in observations and returns the action distribution and value. Returns: A PPONetworks instance with pure functions wrapping the input network. """ def log_prob(params: CategoricalParams, action): return tfd.Categorical(logits=params.logits).log_prob(action) def entropy( params: CategoricalParams, key: networks_lib.PRNGKey ) -> networks_lib.Entropy: del key return tfd.Categorical(logits=params.logits).entropy() def sample(params: CategoricalParams, key: networks_lib.PRNGKey): return tfd.Categorical(logits=params.logits).sample(seed=key) def sample_eval(params: CategoricalParams, key: networks_lib.PRNGKey): del key return tfd.Categorical(logits=params.logits).mode() return PPONetworks( network=network, log_prob=log_prob, entropy=entropy, sample=sample, sample_eval=sample_eval) def make_continuous_networks( environment_spec: specs.EnvironmentSpec, policy_layer_sizes: Sequence[int] = (64, 64), value_layer_sizes: Sequence[int] = (64, 64), use_tanh_gaussian_policy: bool = True, ) -> PPONetworks: """Creates PPONetworks to be used for continuous action environments.""" # Get total number of action dimensions from action spec. num_dimensions = np.prod(environment_spec.actions.shape, dtype=int) def forward_fn(inputs: networks_lib.Observation): def _policy_network(obs: networks_lib.Observation): h = utils.batch_concat(obs) h = hk.nets.MLP(policy_layer_sizes, activate_final=True)(h) # tfd distributions have a weird bug in jax when vmapping is used, so the # safer implementation in general is for the policy network to output the # distribution parameters, and for the distribution to be constructed # in a method such as make_ppo_networks above if not use_tanh_gaussian_policy: # Following networks_lib.MultivariateNormalDiagHead init_scale = 0.3 min_scale = 1e-6 w_init = hk.initializers.VarianceScaling(1e-4) b_init = hk.initializers.Constant(0.) loc_layer = hk.Linear(num_dimensions, w_init=w_init, b_init=b_init) scale_layer = hk.Linear(num_dimensions, w_init=w_init, b_init=b_init) loc = loc_layer(h) scale = jax.nn.softplus(scale_layer(h)) scale *= init_scale / jax.nn.softplus(0.) scale += min_scale return MVNDiagParams(loc=loc, scale_diag=scale) # Following networks_lib.NormalTanhDistribution min_scale = 1e-3 w_init = hk.initializers.VarianceScaling(1.0, 'fan_in', 'uniform') b_init = hk.initializers.Constant(0.) loc_layer = hk.Linear(num_dimensions, w_init=w_init, b_init=b_init) scale_layer = hk.Linear(num_dimensions, w_init=w_init, b_init=b_init) loc = loc_layer(h) scale = scale_layer(h) scale = jax.nn.softplus(scale) + min_scale return TanhNormalParams(loc=loc, scale=scale) value_network = hk.Sequential([ utils.batch_concat, hk.nets.MLP(value_layer_sizes, activate_final=True), hk.Linear(1), lambda x: jnp.squeeze(x, axis=-1) ]) policy_output = _policy_network(inputs) value = value_network(inputs) return (policy_output, value) # Transform into pure functions. forward_fn = hk.without_apply_rng(hk.transform(forward_fn)) dummy_obs = utils.zeros_like(environment_spec.observations) dummy_obs = utils.add_batch_dim(dummy_obs) # Dummy 'sequence' dim. network = networks_lib.FeedForwardNetwork( lambda rng: forward_fn.init(rng, dummy_obs), forward_fn.apply) # Create PPONetworks to add functionality required by the agent. if not use_tanh_gaussian_policy: return make_mvn_diag_ppo_networks(network) return make_tanh_normal_ppo_networks(network)
acme-master
acme/agents/jax/ppo/networks.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Learner for the PPO agent.""" from typing import Dict, Iterator, List, NamedTuple, Optional, Tuple import acme from acme import types from acme.agents.jax.ppo import networks from acme.agents.jax.ppo import normalization from acme.jax import networks as networks_lib from acme.jax.utils import get_from_first_device from acme.utils import counting from acme.utils import loggers import jax import jax.numpy as jnp import optax import reverb import rlax PPOParams = networks.PPOParams class Batch(NamedTuple): """A batch of data; all shapes are expected to be [B, ...].""" observations: types.NestedArray actions: jnp.ndarray advantages: jnp.ndarray # Target value estimate used to bootstrap the value function. target_values: jnp.ndarray # Value estimate and action log-prob at behavior time. behavior_values: jnp.ndarray behavior_log_probs: jnp.ndarray class TrainingState(NamedTuple): """Training state for the PPO learner.""" params: PPOParams opt_state: optax.OptState random_key: networks_lib.PRNGKey # Optional counter used for exponential moving average zero debiasing # Using float32 as it covers a larger range than int32. If using int64 we # would need to do jax_enable_x64. ema_counter: Optional[jnp.float32] = None # Optional parameter for maintaining a running estimate of the scale of # advantage estimates biased_advantage_scale: Optional[networks_lib.Params] = None advantage_scale: Optional[networks_lib.Params] = None # Optional parameter for maintaining a running estimate of the mean and # standard deviation of value estimates biased_value_first_moment: Optional[networks_lib.Params] = None biased_value_second_moment: Optional[networks_lib.Params] = None value_mean: Optional[networks_lib.Params] = None value_std: Optional[networks_lib.Params] = None # Optional parameters for observation normalization obs_normalization_params: Optional[normalization.NormalizationParams] = None class PPOLearner(acme.Learner): """Learner for PPO.""" def __init__( self, ppo_networks: networks.PPONetworks, iterator: Iterator[reverb.ReplaySample], optimizer: optax.GradientTransformation, random_key: networks_lib.PRNGKey, ppo_clipping_epsilon: float = 0.2, normalize_advantage: bool = True, normalize_value: bool = False, normalization_ema_tau: float = 0.995, clip_value: bool = False, value_clipping_epsilon: float = 0.2, max_abs_reward: Optional[float] = None, gae_lambda: float = 0.95, discount: float = 0.99, entropy_cost: float = 0., value_cost: float = 1., num_epochs: int = 4, num_minibatches: int = 1, counter: Optional[counting.Counter] = None, logger: Optional[loggers.Logger] = None, log_global_norm_metrics: bool = False, metrics_logging_period: int = 100, pmap_axis_name: str = 'devices', obs_normalization_fns: Optional[normalization.NormalizationFns] = None, ): self.local_learner_devices = jax.local_devices() self.num_local_learner_devices = jax.local_device_count() self.learner_devices = jax.devices() self.num_epochs = num_epochs self.num_minibatches = num_minibatches self.metrics_logging_period = metrics_logging_period self._num_full_update_steps = 0 self._iterator = iterator normalize_obs = obs_normalization_fns is not None if normalize_obs: assert obs_normalization_fns is not None # Set up logging/counting. self._counter = counter or counting.Counter() self._logger = logger or loggers.make_default_logger('learner') def ppo_loss( params: networks_lib.Params, observations: networks_lib.Observation, actions: networks_lib.Action, advantages: jnp.ndarray, target_values: networks_lib.Value, behavior_values: networks_lib.Value, behavior_log_probs: networks_lib.LogProb, value_mean: jnp.ndarray, value_std: jnp.ndarray, key: networks_lib.PRNGKey, ) -> Tuple[jnp.ndarray, Dict[str, jnp.ndarray]]: """PPO loss for the policy and the critic.""" distribution_params, values = ppo_networks.network.apply( params, observations) if normalize_value: # values = values * jnp.fmax(value_std, 1e-6) + value_mean target_values = (target_values - value_mean) / jnp.fmax(value_std, 1e-6) policy_log_probs = ppo_networks.log_prob(distribution_params, actions) key, sub_key = jax.random.split(key) policy_entropies = ppo_networks.entropy(distribution_params, sub_key) # Compute the policy losses rhos = jnp.exp(policy_log_probs - behavior_log_probs) clipped_ppo_policy_loss = rlax.clipped_surrogate_pg_loss( rhos, advantages, ppo_clipping_epsilon) policy_entropy_loss = -jnp.mean(policy_entropies) total_policy_loss = ( clipped_ppo_policy_loss + entropy_cost * policy_entropy_loss) # Compute the critic losses unclipped_value_loss = (values - target_values)**2 if clip_value: # Clip values to reduce variablility during critic training. clipped_values = behavior_values + jnp.clip(values - behavior_values, -value_clipping_epsilon, value_clipping_epsilon) clipped_value_error = target_values - clipped_values clipped_value_loss = clipped_value_error ** 2 value_loss = jnp.mean(jnp.fmax(unclipped_value_loss, clipped_value_loss)) else: # For Mujoco envs clipping hurts a lot. Evidenced by Figure 43 in # https://arxiv.org/pdf/2006.05990.pdf value_loss = jnp.mean(unclipped_value_loss) total_ppo_loss = total_policy_loss + value_cost * value_loss return total_ppo_loss, { # pytype: disable=bad-return-type # numpy-scalars 'loss_total': total_ppo_loss, 'loss_policy_total': total_policy_loss, 'loss_policy_pg': clipped_ppo_policy_loss, 'loss_policy_entropy': policy_entropy_loss, 'loss_critic': value_loss, } ppo_loss_grad = jax.grad(ppo_loss, has_aux=True) def sgd_step(state: TrainingState, minibatch: Batch): observations = minibatch.observations actions = minibatch.actions advantages = minibatch.advantages target_values = minibatch.target_values behavior_values = minibatch.behavior_values behavior_log_probs = minibatch.behavior_log_probs key, sub_key = jax.random.split(state.random_key) loss_grad, metrics = ppo_loss_grad( state.params.model_params, observations, actions, advantages, target_values, behavior_values, behavior_log_probs, state.value_mean, state.value_std, sub_key, ) # Apply updates loss_grad = jax.lax.pmean(loss_grad, axis_name=pmap_axis_name) updates, opt_state = optimizer.update(loss_grad, state.opt_state) model_params = optax.apply_updates(state.params.model_params, updates) params = PPOParams( model_params=model_params, num_sgd_steps=state.params.num_sgd_steps + 1) if log_global_norm_metrics: metrics['norm_grad'] = optax.global_norm(loss_grad) metrics['norm_updates'] = optax.global_norm(updates) state = state._replace(params=params, opt_state=opt_state, random_key=key) return state, metrics def epoch_update( carry: Tuple[TrainingState, Batch], unused_t: Tuple[()], ): state, carry_batch = carry # Shuffling into minibatches batch_size = carry_batch.advantages.shape[0] key, sub_key = jax.random.split(state.random_key) # TODO(kamyar) For effiency could use same permutation for all epochs permuted_batch = jax.tree_util.tree_map( lambda x: jax.random.permutation( # pylint: disable=g-long-lambda sub_key, x, axis=0, independent=False), carry_batch) state = state._replace(random_key=key) minibatches = jax.tree_util.tree_map( lambda x: jnp.reshape( # pylint: disable=g-long-lambda x, [ # pylint: disable=g-long-lambda num_minibatches, batch_size // num_minibatches ] + list(x.shape[1:])), permuted_batch) # Scan over the minibatches state, metrics = jax.lax.scan( sgd_step, state, minibatches, length=num_minibatches) metrics = jax.tree_util.tree_map(jnp.mean, metrics) return (state, carry_batch), metrics vmapped_network_apply = jax.vmap( ppo_networks.network.apply, in_axes=(None, 0), out_axes=0) def single_device_update( state: TrainingState, trajectories: types.NestedArray, ): params_num_sgd_steps_before_update = state.params.num_sgd_steps # Update the EMA counter and obtain the zero debiasing multiplier if normalize_advantage or normalize_value: ema_counter = state.ema_counter + 1 state = state._replace(ema_counter=ema_counter) zero_debias = 1. / (1. - jnp.power(normalization_ema_tau, ema_counter)) # Extract the data. data = trajectories.data observations, actions, rewards, termination, extra = (data.observation, data.action, data.reward, data.discount, data.extras) if normalize_obs: obs_norm_params = obs_normalization_fns.update( state.obs_normalization_params, observations, pmap_axis_name) state = state._replace(obs_normalization_params=obs_norm_params) observations = obs_normalization_fns.normalize( observations, state.obs_normalization_params) if max_abs_reward is not None: # Apply reward clipping. rewards = jnp.clip(rewards, -1. * max_abs_reward, max_abs_reward) discounts = termination * discount behavior_log_probs = extra['log_prob'] _, behavior_values = vmapped_network_apply(state.params.model_params, observations) if normalize_value: batch_value_first_moment = jnp.mean(behavior_values) batch_value_second_moment = jnp.mean(behavior_values**2) batch_value_first_moment, batch_value_second_moment = jax.lax.pmean( (batch_value_first_moment, batch_value_second_moment), axis_name=pmap_axis_name) biased_value_first_moment = ( normalization_ema_tau * state.biased_value_first_moment + (1. - normalization_ema_tau) * batch_value_first_moment) biased_value_second_moment = ( normalization_ema_tau * state.biased_value_second_moment + (1. - normalization_ema_tau) * batch_value_second_moment) value_mean = biased_value_first_moment * zero_debias value_second_moment = biased_value_second_moment * zero_debias value_std = jnp.sqrt(jax.nn.relu(value_second_moment - value_mean**2)) state = state._replace( biased_value_first_moment=biased_value_first_moment, biased_value_second_moment=biased_value_second_moment, value_mean=value_mean, value_std=value_std, ) behavior_values = behavior_values * jnp.fmax(state.value_std, 1e-6) + state.value_mean behavior_values = jax.lax.stop_gradient(behavior_values) # Compute GAE using rlax vmapped_rlax_truncated_generalized_advantage_estimation = jax.vmap( rlax.truncated_generalized_advantage_estimation, in_axes=(0, 0, None, 0)) advantages = vmapped_rlax_truncated_generalized_advantage_estimation( rewards[:, :-1], discounts[:, :-1], gae_lambda, behavior_values) advantages = jax.lax.stop_gradient(advantages) target_values = behavior_values[:, :-1] + advantages target_values = jax.lax.stop_gradient(target_values) # Exclude the last step - it was only used for bootstrapping. # The shape is [num_sequences, num_steps, ..] (observations, actions, behavior_log_probs, behavior_values) = ( jax.tree_util.tree_map( lambda x: x[:, :-1], (observations, actions, behavior_log_probs, behavior_values), ) ) # Shuffle the data and break into minibatches batch_size = advantages.shape[0] * advantages.shape[1] batch = Batch( observations=observations, actions=actions, advantages=advantages, target_values=target_values, behavior_values=behavior_values, behavior_log_probs=behavior_log_probs) batch = jax.tree_util.tree_map( lambda x: jnp.reshape(x, [batch_size] + list(x.shape[2:])), batch) if normalize_advantage: batch_advantage_scale = jnp.mean(jnp.abs(batch.advantages)) batch_advantage_scale = jax.lax.pmean(batch_advantage_scale, pmap_axis_name) # update the running statistics biased_advantage_scale = ( normalization_ema_tau * state.biased_advantage_scale + (1. - normalization_ema_tau) * batch_advantage_scale) advantage_scale = biased_advantage_scale * zero_debias state = state._replace( biased_advantage_scale=biased_advantage_scale, advantage_scale=advantage_scale) # scale the advantages scaled_advantages = batch.advantages / jnp.fmax(state.advantage_scale, 1e-6) batch = batch._replace(advantages=scaled_advantages) # Scan desired number of epoch updates (state, _), metrics = jax.lax.scan( epoch_update, (state, batch), (), length=num_epochs) metrics = jax.tree_util.tree_map(jnp.mean, metrics) if normalize_advantage: metrics['advantage_scale'] = state.advantage_scale if normalize_value: metrics['value_mean'] = value_mean metrics['value_std'] = value_std delta_params_sgd_steps = ( data.extras['params_num_sgd_steps'][:, 0] - params_num_sgd_steps_before_update) metrics['delta_params_sgd_steps_min'] = jnp.min(delta_params_sgd_steps) metrics['delta_params_sgd_steps_max'] = jnp.max(delta_params_sgd_steps) metrics['delta_params_sgd_steps_mean'] = jnp.mean(delta_params_sgd_steps) metrics['delta_params_sgd_steps_std'] = jnp.std(delta_params_sgd_steps) return state, metrics pmapped_update_step = jax.pmap( single_device_update, axis_name=pmap_axis_name, devices=self.learner_devices) def full_update_step( state: TrainingState, trajectories: types.NestedArray, ): state, metrics = pmapped_update_step(state, trajectories) return state, metrics self._full_update_step = full_update_step def make_initial_state(key: networks_lib.PRNGKey) -> TrainingState: """Initialises the training state (parameters and optimiser state).""" all_keys = jax.random.split(key, num=self.num_local_learner_devices + 1) key_init, key_state = all_keys[0], all_keys[1:] key_state = [key_state[i] for i in range(self.num_local_learner_devices)] key_state = jax.device_put_sharded(key_state, self.local_learner_devices) initial_params = ppo_networks.network.init(key_init) initial_opt_state = optimizer.init(initial_params) # Using float32 as it covers a larger range than int32. If using int64 we # would need to do jax_enable_x64. params_num_sgd_steps = jnp.zeros(shape=(), dtype=jnp.float32) initial_params = jax.device_put_replicated(initial_params, self.local_learner_devices) initial_opt_state = jax.device_put_replicated(initial_opt_state, self.local_learner_devices) params_num_sgd_steps = jax.device_put_replicated( params_num_sgd_steps, self.local_learner_devices) ema_counter = jnp.float32(0) ema_counter = jax.device_put_replicated(ema_counter, self.local_learner_devices) init_state = TrainingState( params=PPOParams( model_params=initial_params, num_sgd_steps=params_num_sgd_steps), opt_state=initial_opt_state, random_key=key_state, ema_counter=ema_counter, ) if normalize_advantage: biased_advantage_scale = jax.device_put_replicated( jnp.zeros([]), self.local_learner_devices) advantage_scale = jax.device_put_replicated( jnp.zeros([]), self.local_learner_devices) init_state = init_state._replace( biased_advantage_scale=biased_advantage_scale, advantage_scale=advantage_scale) if normalize_value: biased_value_first_moment = jax.device_put_replicated( jnp.zeros([]), self.local_learner_devices) value_mean = biased_value_first_moment biased_value_second_moment = jax.device_put_replicated( jnp.zeros([]), self.local_learner_devices) value_second_moment = biased_value_second_moment value_std = jnp.sqrt(jax.nn.relu(value_second_moment - value_mean**2)) init_state = init_state._replace( biased_value_first_moment=biased_value_first_moment, biased_value_second_moment=biased_value_second_moment, value_mean=value_mean, value_std=value_std) if normalize_obs: obs_norm_params = obs_normalization_fns.init() # pytype: disable=attribute-error obs_norm_params = jax.device_put_replicated(obs_norm_params, self.local_learner_devices) init_state = init_state._replace( obs_normalization_params=obs_norm_params) return init_state # Initialise training state (parameters and optimizer state). self._state = make_initial_state(random_key) self._cached_state = get_from_first_device(self._state, as_numpy=True) def step(self): """Does a learner step and logs the results. One learner step consists of (possibly multiple) epochs of PPO updates on a batch of NxT steps collected by the actors. """ sample = next(self._iterator) self._state, results = self._full_update_step(self._state, sample) self._cached_state = get_from_first_device(self._state, as_numpy=True) # Update our counts and record it. counts = self._counter.increment(steps=self.num_epochs * self.num_minibatches) # Snapshot and attempt to write logs. if self._num_full_update_steps % self.metrics_logging_period == 0: results = jax.tree_util.tree_map(jnp.mean, results) self._logger.write({**results, **counts}) self._num_full_update_steps += 1 def get_variables(self, names: List[str]) -> List[networks_lib.Params]: variables = self._cached_state return [getattr(variables, name) for name in names] def save(self) -> TrainingState: return self._cached_state def restore(self, state: TrainingState): # TODO(kamyar) Should the random_key come from self._state instead? random_key = state.random_key random_key = jax.random.split( random_key, num=self.num_local_learner_devices) random_key = jax.device_put_sharded( [random_key[i] for i in range(self.num_local_learner_devices)], self.local_learner_devices) state = jax.device_put_replicated(state, self.local_learner_devices) state = state._replace(random_key=random_key) self._state = state self._cached_state = get_from_first_device(self._state, as_numpy=True)
acme-master
acme/agents/jax/ppo/learning.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Utilities for normalization.""" import dataclasses from typing import Any, Callable, Generic, NamedTuple, Optional from acme import adders from acme import types from acme.agents.jax import actor_core from acme.agents.jax import actors from acme.jax import networks as network_lib from acme.jax import running_statistics from acme.jax import utils from acme.jax import variable_utils import jax import jax.numpy as jnp NormalizationParams = Any RunningStatisticsState = running_statistics.RunningStatisticsState @dataclasses.dataclass class NormalizationFns: """Holds pure functions for normalization. Attributes: init: A pure function: ``params = init()`` normalize: A pure function: ``norm_x = normalize(x, params)`` update: A pure function: ``params = update(params, x, pmap_axis_name)`` """ # Returns the initial parameters for the normalization utility. init: Callable[[], NormalizationParams] # Returns the normalized input nested array. normalize: Callable[[types.NestedArray, NormalizationParams], types.NestedArray] # Returns updates normalization parameters. update: Callable[[NormalizationParams, types.NestedArray, Optional[str]], NormalizationParams] class NormalizedGenericActor(actors.GenericActor[actor_core.State, actor_core.Extras], Generic[actor_core.State, actor_core.Extras]): """A GenericActor that uses observation normalization.""" def __init__(self, actor: actor_core.ActorCore[actor_core.State, actor_core.Extras], normalization_fns: NormalizationFns, random_key: network_lib.PRNGKey, variable_client: Optional[variable_utils.VariableClient], adder: Optional[adders.Adder] = None, jit: bool = True, backend: Optional[str] = 'cpu', per_episode_update: bool = False): """Initializes a feed forward actor. Args: actor: actor core. normalization_fns: Function that are used for normalizing observations. random_key: Random key. variable_client: The variable client to get policy and observation normalization parameters from. The variable client should be defined to provide [policy_params, obs_norm_params]. adder: An adder to add experiences to. jit: Whether or not to jit the ActorCore and normalization functions. backend: Which backend to use when jitting. per_episode_update: if True, updates variable client params once at the beginning of each episode """ super().__init__(actor, random_key, variable_client, adder, jit, backend, per_episode_update) if jit: self._apply_normalization = jax.jit( normalization_fns.normalize, backend=backend) else: self._apply_normalization = normalization_fns.normalize def select_action(self, observation: network_lib.Observation) -> types.NestedArray: policy_params, obs_norm_params = tuple(self._params) observation = self._apply_normalization(observation, obs_norm_params) action, self._state = self._policy(policy_params, observation, self._state) return utils.to_numpy(action) class EMAMeanStdNormalizerParams(NamedTuple): """Using technique form Adam optimizer paper for computing running stats.""" ema_counter: jnp.int32 biased_first_moment: types.NestedArray biased_second_moment: types.NestedArray def build_ema_mean_std_normalizer( nested_spec: types.NestedSpec, tau: float = 0.995, epsilon: float = 1e-6,) -> NormalizationFns: """Builds pure functions used for normalizing based on EMA mean and std. The built normalizer functions can be used to normalize nested arrays that have a structure corresponding to nested_spec. Currently only supports nested_spec where all leafs have float dtype. Arguments: nested_spec: A nested spec where all leaves have float dtype tau: tau parameter for exponential moving average epsilon: epsilon for avoiding division by zero std Returns: NormalizationFns to be used for normalization """ nested_dims = jax.tree_util.tree_map(lambda x: len(x.shape), nested_spec) def init() -> EMAMeanStdNormalizerParams: first_moment = utils.zeros_like(nested_spec) second_moment = utils.zeros_like(nested_spec) return EMAMeanStdNormalizerParams( ema_counter=jnp.int32(0), biased_first_moment=first_moment, biased_second_moment=second_moment, ) def _normalize_leaf( x: jnp.ndarray, ema_counter: jnp.int32, biased_first_moment: jnp.ndarray, biased_second_moment: jnp.ndarray, ) -> jnp.ndarray: zero_debias = 1. / (1. - jnp.power(tau, ema_counter)) mean = biased_first_moment * zero_debias second_moment = biased_second_moment * zero_debias std = jnp.sqrt(jax.nn.relu(second_moment - mean**2)) mean = jnp.broadcast_to(mean, x.shape) std = jnp.broadcast_to(std, x.shape) return (x - mean) / jnp.fmax(std, epsilon) def _normalize(nested_array: types.NestedArray, params: EMAMeanStdNormalizerParams) -> types.NestedArray: ema_counter = params.ema_counter normalized_nested_array = jax.tree_util.tree_map( lambda x, f, s: _normalize_leaf(x, ema_counter, f, s), nested_array, params.biased_first_moment, params.biased_second_moment) return normalized_nested_array def normalize(nested_array: types.NestedArray, params: EMAMeanStdNormalizerParams) -> types.NestedArray: ema_counter = params.ema_counter norm_obs = jax.lax.cond( ema_counter > 0, _normalize, lambda o, p: o, nested_array, params) return norm_obs def _compute_first_moment(x: jnp.ndarray, ndim: int): reduce_axes = tuple(range(len(x.shape) - ndim)) first_moment = jnp.mean(x, axis=reduce_axes) return first_moment def _compute_second_moment(x: jnp.ndarray, ndim: int): reduce_axes = tuple(range(len(x.shape) - ndim)) second_moment = jnp.mean(x**2, axis=reduce_axes) return second_moment def update( params: EMAMeanStdNormalizerParams, nested_array: types.NestedArray, pmap_axis_name: Optional[str] = None) -> EMAMeanStdNormalizerParams: # compute the stats first_moment = jax.tree_util.tree_map( _compute_first_moment, nested_array, nested_dims) second_moment = jax.tree_util.tree_map( _compute_second_moment, nested_array, nested_dims) # propagate across devices if pmap_axis_name is not None: first_moment, second_moment = jax.lax.pmean( (first_moment, second_moment), axis_name=pmap_axis_name) # update running statistics new_first_moment = jax.tree_util.tree_map( lambda x, y: tau * x + # pylint: disable=g-long-lambda (1. - tau) * y, params.biased_first_moment, first_moment) new_second_moment = jax.tree_util.tree_map( lambda x, y: tau * x + # pylint: disable=g-long-lambda (1. - tau) * y, params.biased_second_moment, second_moment) # update ema_counter and return updated params new_params = EMAMeanStdNormalizerParams( ema_counter=params.ema_counter + 1, biased_first_moment=new_first_moment, biased_second_moment=new_second_moment, ) return new_params return NormalizationFns( init=init, normalize=normalize, update=update, ) def build_mean_std_normalizer( nested_spec: types.NestedSpec, max_abs_value: Optional[float] = None) -> NormalizationFns: """Builds pure functions used for normalizing based on mean and std. Arguments: nested_spec: A nested spec where all leaves have float dtype max_abs_value: Normalized nested arrays will be clipped so that all values will be between -max_abs_value and +max_abs_value. Setting to None (default) does not perform this clipping. Returns: NormalizationFns to be used for normalization """ def init() -> RunningStatisticsState: return running_statistics.init_state(nested_spec) def normalize( nested_array: types.NestedArray, params: RunningStatisticsState) -> types.NestedArray: return running_statistics.normalize( nested_array, params, max_abs_value=max_abs_value) def update( params: RunningStatisticsState, nested_array: types.NestedArray, pmap_axis_name: Optional[str]) -> RunningStatisticsState: return running_statistics.update( params, nested_array, pmap_axis_name=pmap_axis_name) return NormalizationFns( init=init, normalize=normalize, update=update)
acme-master
acme/agents/jax/ppo/normalization.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Config classes for CRR.""" import dataclasses @dataclasses.dataclass class CRRConfig: """Configuration options for CRR. Attributes: learning_rate: Learning rate. discount: discount to use for TD updates. target_update_period: period to update target's parameters. use_sarsa_target: compute on-policy target using iterator's actions rather than sampled actions. Useful for 1-step offline RL (https://arxiv.org/pdf/2106.08909.pdf). """ learning_rate: float = 3e-4 discount: float = 0.99 target_update_period: int = 100 use_sarsa_target: bool = False
acme-master
acme/agents/jax/crr/config.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Implementation of the Critic Regularized Regression (CRR) agent.""" from acme.agents.jax.crr.builder import CRRBuilder from acme.agents.jax.crr.config import CRRConfig from acme.agents.jax.crr.learning import CRRLearner from acme.agents.jax.crr.losses import policy_loss_coeff_advantage_exp from acme.agents.jax.crr.losses import policy_loss_coeff_advantage_indicator from acme.agents.jax.crr.losses import policy_loss_coeff_constant from acme.agents.jax.crr.losses import PolicyLossCoeff from acme.agents.jax.crr.networks import CRRNetworks from acme.agents.jax.crr.networks import make_networks
acme-master
acme/agents/jax/crr/__init__.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """CRR Builder.""" from typing import Iterator, Optional from acme import core from acme import specs from acme import types from acme.agents.jax import actor_core as actor_core_lib from acme.agents.jax import actors from acme.agents.jax import builders from acme.agents.jax.crr import config as crr_config from acme.agents.jax.crr import learning from acme.agents.jax.crr import losses from acme.agents.jax.crr import networks as crr_networks from acme.jax import networks as networks_lib from acme.jax import variable_utils from acme.utils import counting from acme.utils import loggers import optax class CRRBuilder(builders.OfflineBuilder[crr_networks.CRRNetworks, actor_core_lib.FeedForwardPolicy, types.Transition]): """CRR Builder.""" def __init__( self, config: crr_config.CRRConfig, policy_loss_coeff_fn: losses.PolicyLossCoeff, ): """Creates a CRR learner, an evaluation policy and an eval actor. Args: config: a config with CRR hps. policy_loss_coeff_fn: set the loss function for the policy. """ self._config = config self._policy_loss_coeff_fn = policy_loss_coeff_fn def make_learner( self, random_key: networks_lib.PRNGKey, networks: crr_networks.CRRNetworks, dataset: Iterator[types.Transition], logger_fn: loggers.LoggerFactory, environment_spec: specs.EnvironmentSpec, *, counter: Optional[counting.Counter] = None, ) -> core.Learner: del environment_spec return learning.CRRLearner( networks=networks, random_key=random_key, discount=self._config.discount, target_update_period=self._config.target_update_period, iterator=dataset, policy_loss_coeff_fn=self._policy_loss_coeff_fn, policy_optimizer=optax.adam(self._config.learning_rate), critic_optimizer=optax.adam(self._config.learning_rate), use_sarsa_target=self._config.use_sarsa_target, logger=logger_fn('learner'), counter=counter) def make_actor( self, random_key: networks_lib.PRNGKey, policy: actor_core_lib.FeedForwardPolicy, environment_spec: specs.EnvironmentSpec, variable_source: Optional[core.VariableSource] = None, ) -> core.Actor: del environment_spec assert variable_source is not None actor_core = actor_core_lib.batched_feed_forward_to_actor_core(policy) variable_client = variable_utils.VariableClient( variable_source, 'policy', device='cpu') return actors.GenericActor( actor_core, random_key, variable_client, backend='cpu') def make_policy(self, networks: crr_networks.CRRNetworks, environment_spec: specs.EnvironmentSpec, evaluation: bool) -> actor_core_lib.FeedForwardPolicy: """Construct the policy.""" del environment_spec, evaluation def evaluation_policy( params: networks_lib.Params, key: networks_lib.PRNGKey, observation: networks_lib.Observation) -> networks_lib.Action: dist_params = networks.policy_network.apply(params, observation) return networks.sample_eval(dist_params, key) return evaluation_policy
acme-master
acme/agents/jax/crr/builder.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Networks definition for CRR.""" import dataclasses from typing import Callable, Tuple from acme import specs from acme.jax import networks as networks_lib from acme.jax import utils import haiku as hk import jax import jax.numpy as jnp import numpy as np @dataclasses.dataclass class CRRNetworks: """Network and pure functions for the CRR agent..""" policy_network: networks_lib.FeedForwardNetwork critic_network: networks_lib.FeedForwardNetwork log_prob: networks_lib.LogProbFn sample: networks_lib.SampleFn sample_eval: networks_lib.SampleFn def make_networks( spec: specs.EnvironmentSpec, policy_layer_sizes: Tuple[int, ...] = (256, 256), critic_layer_sizes: Tuple[int, ...] = (256, 256), activation: Callable[[jnp.ndarray], jnp.ndarray] = jax.nn.relu, ) -> CRRNetworks: """Creates networks used by the agent.""" num_actions = np.prod(spec.actions.shape, dtype=int) # Create dummy observations and actions to create network parameters. dummy_action = utils.add_batch_dim(utils.zeros_like(spec.actions)) dummy_obs = utils.add_batch_dim(utils.zeros_like(spec.observations)) def _policy_fn(obs: jnp.ndarray) -> jnp.ndarray: network = hk.Sequential([ hk.nets.MLP( list(policy_layer_sizes), w_init=hk.initializers.VarianceScaling(1.0, 'fan_in', 'uniform'), activation=activation, activate_final=True), networks_lib.NormalTanhDistribution(num_actions), ]) return network(obs) policy = hk.without_apply_rng(hk.transform(_policy_fn)) policy_network = networks_lib.FeedForwardNetwork( lambda key: policy.init(key, dummy_obs), policy.apply) def _critic_fn(obs, action): network = hk.Sequential([ hk.nets.MLP( list(critic_layer_sizes) + [1], w_init=hk.initializers.VarianceScaling(1.0, 'fan_in', 'uniform'), activation=activation), ]) data = jnp.concatenate([obs, action], axis=-1) return network(data) critic = hk.without_apply_rng(hk.transform(_critic_fn)) critic_network = networks_lib.FeedForwardNetwork( lambda key: critic.init(key, dummy_obs, dummy_action), critic.apply) return CRRNetworks( policy_network=policy_network, critic_network=critic_network, log_prob=lambda params, actions: params.log_prob(actions), sample=lambda params, key: params.sample(seed=key), sample_eval=lambda params, key: params.mode())
acme-master
acme/agents/jax/crr/networks.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Tests for the CRR agent.""" from acme import specs from acme.agents.jax import crr from acme.testing import fakes import jax import optax from absl.testing import absltest from absl.testing import parameterized class CRRTest(parameterized.TestCase): @parameterized.named_parameters( ('exp', crr.policy_loss_coeff_advantage_exp), ('indicator', crr.policy_loss_coeff_advantage_indicator), ('all', crr.policy_loss_coeff_constant)) def test_train(self, policy_loss_coeff_fn): seed = 0 num_iterations = 5 batch_size = 64 grad_updates_per_batch = 1 # Create a fake environment to test with. environment = fakes.ContinuousEnvironment( episode_length=10, bounded=True, action_dim=6) spec = specs.make_environment_spec(environment) # Construct the learner. networks = crr.make_networks( spec, policy_layer_sizes=(8, 8), critic_layer_sizes=(8, 8)) key = jax.random.PRNGKey(seed) dataset = fakes.transition_iterator(environment) learner = crr.CRRLearner( networks, key, discount=0.95, target_update_period=2, policy_loss_coeff_fn=policy_loss_coeff_fn, iterator=dataset(batch_size * grad_updates_per_batch), policy_optimizer=optax.adam(1e-4), critic_optimizer=optax.adam(1e-4), grad_updates_per_batch=grad_updates_per_batch) # Train the learner. for _ in range(num_iterations): learner.step() if __name__ == '__main__': absltest.main()
acme-master
acme/agents/jax/crr/agent_test.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Loss (weight) functions for CRR.""" from typing import Callable from acme import types from acme.agents.jax.crr.networks import CRRNetworks from acme.jax import networks as networks_lib import jax.numpy as jnp PolicyLossCoeff = Callable[[ CRRNetworks, networks_lib.Params, networks_lib.Params, types.Transition, networks_lib.PRNGKey, ], jnp.ndarray] def _compute_advantage(networks: CRRNetworks, policy_params: networks_lib.Params, critic_params: networks_lib.Params, transition: types.Transition, key: networks_lib.PRNGKey, num_action_samples: int = 4) -> jnp.ndarray: """Returns the advantage for the transition.""" # Sample count actions. replicated_observation = jnp.broadcast_to(transition.observation, (num_action_samples,) + transition.observation.shape) dist_params = networks.policy_network.apply(policy_params, replicated_observation) actions = networks.sample(dist_params, key) # Compute the state-action values for the sampled actions. q_actions = networks.critic_network.apply(critic_params, replicated_observation, actions) # Take the mean as the state-value estimate. It is also possible to take the # maximum, aka CRR(max); see table 1 in CRR paper. q_estimate = jnp.mean(q_actions, axis=0) # Compute the advantage. q = networks.critic_network.apply(critic_params, transition.observation, transition.action) return q - q_estimate def policy_loss_coeff_advantage_exp( networks: CRRNetworks, policy_params: networks_lib.Params, critic_params: networks_lib.Params, transition: types.Transition, key: networks_lib.PRNGKey, num_action_samples: int = 4, beta: float = 1.0, ratio_upper_bound: float = 20.0) -> jnp.ndarray: """Exponential advantage weigting; see equation (4) in CRR paper.""" advantage = _compute_advantage(networks, policy_params, critic_params, transition, key, num_action_samples) return jnp.minimum(jnp.exp(advantage / beta), ratio_upper_bound) def policy_loss_coeff_advantage_indicator( networks: CRRNetworks, policy_params: networks_lib.Params, critic_params: networks_lib.Params, transition: types.Transition, key: networks_lib.PRNGKey, num_action_samples: int = 4) -> jnp.ndarray: """Indicator advantage weighting; see equation (3) in CRR paper.""" advantage = _compute_advantage(networks, policy_params, critic_params, transition, key, num_action_samples) return jnp.heaviside(advantage, 0.) def policy_loss_coeff_constant(networks: CRRNetworks, policy_params: networks_lib.Params, critic_params: networks_lib.Params, transition: types.Transition, key: networks_lib.PRNGKey, value: float = 1.0) -> jnp.ndarray: """Constant weights.""" del networks del policy_params del critic_params del transition del key return value # pytype: disable=bad-return-type # jax-ndarray
acme-master
acme/agents/jax/crr/losses.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """CRR learner implementation.""" import time from typing import Dict, Iterator, List, NamedTuple, Optional, Tuple import acme from acme import types from acme.agents.jax.crr.losses import PolicyLossCoeff from acme.agents.jax.crr.networks import CRRNetworks from acme.jax import networks as networks_lib from acme.jax import utils from acme.utils import counting from acme.utils import loggers import jax import jax.numpy as jnp import optax class TrainingState(NamedTuple): """Contains training state for the learner.""" policy_params: networks_lib.Params target_policy_params: networks_lib.Params critic_params: networks_lib.Params target_critic_params: networks_lib.Params policy_opt_state: optax.OptState critic_opt_state: optax.OptState steps: int key: networks_lib.PRNGKey class CRRLearner(acme.Learner): """Critic Regularized Regression (CRR) learner. This is the learning component of a CRR agent as described in https://arxiv.org/abs/2006.15134. """ _state: TrainingState def __init__(self, networks: CRRNetworks, random_key: networks_lib.PRNGKey, discount: float, target_update_period: int, policy_loss_coeff_fn: PolicyLossCoeff, iterator: Iterator[types.Transition], policy_optimizer: optax.GradientTransformation, critic_optimizer: optax.GradientTransformation, counter: Optional[counting.Counter] = None, logger: Optional[loggers.Logger] = None, grad_updates_per_batch: int = 1, use_sarsa_target: bool = False): """Initializes the CRR learner. Args: networks: CRR networks. random_key: a key for random number generation. discount: discount to use for TD updates. target_update_period: period to update target's parameters. policy_loss_coeff_fn: set the loss function for the policy. iterator: an iterator over training data. policy_optimizer: the policy optimizer. critic_optimizer: the Q-function optimizer. counter: counter object used to keep track of steps. logger: logger object to be used by learner. grad_updates_per_batch: how many gradient updates given a sampled batch. use_sarsa_target: compute on-policy target using iterator's actions rather than sampled actions. Useful for 1-step offline RL (https://arxiv.org/pdf/2106.08909.pdf). When set to `True`, `target_policy_params` are unused. """ critic_network = networks.critic_network policy_network = networks.policy_network def policy_loss( policy_params: networks_lib.Params, critic_params: networks_lib.Params, transition: types.Transition, key: networks_lib.PRNGKey, ) -> jnp.ndarray: # Compute the loss coefficients. coeff = policy_loss_coeff_fn(networks, policy_params, critic_params, transition, key) coeff = jax.lax.stop_gradient(coeff) # Return the weighted loss. dist_params = policy_network.apply(policy_params, transition.observation) logp_action = networks.log_prob(dist_params, transition.action) # Make sure there is no broadcasting. logp_action *= coeff.flatten() assert len(logp_action.shape) == 1 return -jnp.mean(logp_action) def critic_loss( critic_params: networks_lib.Params, target_policy_params: networks_lib.Params, target_critic_params: networks_lib.Params, transition: types.Transition, key: networks_lib.PRNGKey, ): # Sample the next action. if use_sarsa_target: # TODO(b/222674779): use N-steps Trajectories to get the next actions. assert 'next_action' in transition.extras, ( 'next actions should be given as extras for one step RL.') next_action = transition.extras['next_action'] else: next_dist_params = policy_network.apply(target_policy_params, transition.next_observation) next_action = networks.sample(next_dist_params, key) # Calculate the value of the next state and action. next_q = critic_network.apply(target_critic_params, transition.next_observation, next_action) target_q = transition.reward + transition.discount * discount * next_q target_q = jax.lax.stop_gradient(target_q) q = critic_network.apply(critic_params, transition.observation, transition.action) q_error = q - target_q # Loss is MSE scaled by 0.5, so the gradient is equal to the TD error. # TODO(sertan): Replace with a distributional critic. CRR paper states # that this may perform better. return 0.5 * jnp.mean(jnp.square(q_error)) policy_loss_and_grad = jax.value_and_grad(policy_loss) critic_loss_and_grad = jax.value_and_grad(critic_loss) def sgd_step( state: TrainingState, transitions: types.Transition, ) -> Tuple[TrainingState, Dict[str, jnp.ndarray]]: key, key_policy, key_critic = jax.random.split(state.key, 3) # Compute losses and their gradients. policy_loss_value, policy_gradients = policy_loss_and_grad( state.policy_params, state.critic_params, transitions, key_policy) critic_loss_value, critic_gradients = critic_loss_and_grad( state.critic_params, state.target_policy_params, state.target_critic_params, transitions, key_critic) # Get optimizer updates and state. policy_updates, policy_opt_state = policy_optimizer.update( policy_gradients, state.policy_opt_state) critic_updates, critic_opt_state = critic_optimizer.update( critic_gradients, state.critic_opt_state) # Apply optimizer updates to parameters. policy_params = optax.apply_updates(state.policy_params, policy_updates) critic_params = optax.apply_updates(state.critic_params, critic_updates) steps = state.steps + 1 # Periodically update target networks. target_policy_params, target_critic_params = optax.periodic_update( # pytype: disable=wrong-arg-types # numpy-scalars (policy_params, critic_params), (state.target_policy_params, state.target_critic_params), steps, target_update_period) new_state = TrainingState( policy_params=policy_params, target_policy_params=target_policy_params, critic_params=critic_params, target_critic_params=target_critic_params, policy_opt_state=policy_opt_state, critic_opt_state=critic_opt_state, steps=steps, key=key, ) metrics = { 'policy_loss': policy_loss_value, 'critic_loss': critic_loss_value, } return new_state, metrics sgd_step = utils.process_multiple_batches(sgd_step, grad_updates_per_batch) self._sgd_step = jax.jit(sgd_step) # General learner book-keeping and loggers. self._counter = counter or counting.Counter() self._logger = logger or loggers.make_default_logger( 'learner', asynchronous=True, serialize_fn=utils.fetch_devicearray, steps_key=self._counter.get_steps_key()) # Create prefetching dataset iterator. self._iterator = iterator # Create the network parameters and copy into the target network parameters. key, key_policy, key_critic = jax.random.split(random_key, 3) initial_policy_params = policy_network.init(key_policy) initial_critic_params = critic_network.init(key_critic) initial_target_policy_params = initial_policy_params initial_target_critic_params = initial_critic_params # Initialize optimizers. initial_policy_opt_state = policy_optimizer.init(initial_policy_params) initial_critic_opt_state = critic_optimizer.init(initial_critic_params) # Create initial state. self._state = TrainingState( policy_params=initial_policy_params, target_policy_params=initial_target_policy_params, critic_params=initial_critic_params, target_critic_params=initial_target_critic_params, policy_opt_state=initial_policy_opt_state, critic_opt_state=initial_critic_opt_state, steps=0, key=key, ) # Do not record timestamps until after the first learning step is done. # This is to avoid including the time it takes for actors to come online and # fill the replay buffer. self._timestamp = None def step(self): transitions = next(self._iterator) self._state, metrics = self._sgd_step(self._state, transitions) # Compute elapsed time. timestamp = time.time() elapsed_time = timestamp - self._timestamp if self._timestamp else 0 self._timestamp = timestamp # Increment counts and record the current time counts = self._counter.increment(steps=1, walltime=elapsed_time) # Attempts to write the logs. self._logger.write({**metrics, **counts}) def get_variables(self, names: List[str]) -> List[networks_lib.Params]: # We only expose the variables for the learned policy and critic. The target # policy and critic are internal details. variables = { 'policy': self._state.target_policy_params, 'critic': self._state.target_critic_params, } return [variables[name] for name in names] def save(self) -> TrainingState: return self._state def restore(self, state: TrainingState): self._state = state
acme-master
acme/agents/jax/crr/learning.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """AIL config.""" import dataclasses from typing import Optional import optax @dataclasses.dataclass class AILConfig: """Configuration options for AIL. Attributes: direct_rl_batch_size: Batch size of a direct rl algorithm (measured in transitions). is_sequence_based: If True, a direct rl algorithm is using SequenceAdder data format. Otherwise the learner assumes that the direct rl algorithm is using NStepTransitionAdder. share_iterator: If True, AIL will use the same iterator for the discriminator network training as the direct rl algorithm. num_sgd_steps_per_step: Only used if 'share_iterator' is False. Denotes how many gradient updates perform per one learner step. discriminator_batch_size: Batch size for training the discriminator. policy_variable_name: The name of the policy variable to retrieve direct_rl policy parameters. discriminator_optimizer: Optimizer for the discriminator. If not specified it is set to Adam with learning rate of 1e-5. replay_table_name: The name of the reverb replay table to use. prefetch_size: How many batches to prefetch discount: Discount to use for TD updates min_replay_size: Minimal size of replay buffer max_replay_size: Maximal size of replay buffer policy_to_expert_data_ratio: If not None, the direct RL learner will receive expert transitions in the given proportions. policy_to_expert_data_ratio + 1 must divide the direct RL batch size. """ direct_rl_batch_size: int is_sequence_based: bool = False share_iterator: bool = True num_sgd_steps_per_step: int = 1 discriminator_batch_size: int = 256 policy_variable_name: Optional[str] = None discriminator_optimizer: Optional[optax.GradientTransformation] = None replay_table_name: str = 'ail_table' prefetch_size: int = 4 discount: float = 0.99 min_replay_size: int = 1000 max_replay_size: int = int(1e6) policy_to_expert_data_ratio: Optional[int] = None def __post_init__(self): assert self.direct_rl_batch_size % self.discriminator_batch_size == 0 def get_per_learner_step_batch_size(config: AILConfig) -> int: """Returns how many transitions should be sampled per direct learner step.""" # If the iterators are tied, the discriminator learning batch size has to # match the direct RL one. if config.share_iterator: assert (config.direct_rl_batch_size % config.discriminator_batch_size) == 0 return config.direct_rl_batch_size # Otherwise each iteration of the discriminator will sample a batch which will # be split in num_sgd_steps_per_step batches, each of size # discriminator_batch_size. return config.discriminator_batch_size * config.num_sgd_steps_per_step
acme-master
acme/agents/jax/ail/config.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Builder for DAC. https://arxiv.org/pdf/1809.02925.pdf """ import dataclasses from typing import Callable, Iterator from acme import types from acme.agents.jax import actor_core as actor_core_lib from acme.agents.jax import td3 from acme.agents.jax.ail import builder from acme.agents.jax.ail import config as ail_config from acme.agents.jax.ail import losses @dataclasses.dataclass class DACConfig: """Configuration options specific to DAC. Attributes: ail_config: AIL config. td3_config: TD3 config. entropy_coefficient: Entropy coefficient of the discriminator loss. gradient_penalty_coefficient: Coefficient for the gradient penalty term in the discriminator loss. """ ail_config: ail_config.AILConfig td3_config: td3.TD3Config entropy_coefficient: float = 1e-3 gradient_penalty_coefficient: float = 10. class DACBuilder(builder.AILBuilder[td3.TD3Networks, actor_core_lib.FeedForwardPolicy]): """DAC Builder.""" def __init__(self, config: DACConfig, make_demonstrations: Callable[[int], Iterator[types.Transition]]): td3_builder = td3.TD3Builder(config.td3_config) dac_loss = losses.add_gradient_penalty( losses.gail_loss(entropy_coefficient=config.entropy_coefficient), gradient_penalty_coefficient=config.gradient_penalty_coefficient, gradient_penalty_target=1.) super().__init__( td3_builder, config=config.ail_config, discriminator_loss=dac_loss, make_demonstrations=make_demonstrations)
acme-master
acme/agents/jax/ail/dac.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Tests for the AIL learner.""" import functools from acme import specs from acme import types from acme.agents.jax.ail import learning as ail_learning from acme.agents.jax.ail import losses from acme.agents.jax.ail import networks as ail_networks from acme.jax import networks as networks_lib from acme.jax import utils import haiku as hk import jax import numpy as np import optax from absl.testing import absltest def _make_discriminator(spec): def discriminator(*args, **kwargs) -> networks_lib.Logits: return ail_networks.DiscriminatorModule( environment_spec=spec, use_action=False, use_next_obs=False, network_core=ail_networks.DiscriminatorMLP([]))(*args, **kwargs) discriminator_transformed = hk.without_apply_rng( hk.transform_with_state(discriminator)) return ail_networks.make_discriminator( environment_spec=spec, discriminator_transformed=discriminator_transformed) class AilLearnerTest(absltest.TestCase): def test_step(self): simple_spec = specs.Array(shape=(), dtype=float) spec = specs.EnvironmentSpec(simple_spec, simple_spec, simple_spec, simple_spec) discriminator = _make_discriminator(spec) ail_network = ail_networks.AILNetworks( discriminator, imitation_reward_fn=lambda x: x, direct_rl_networks=None) loss = losses.gail_loss() optimizer = optax.adam(.01) step = jax.jit(functools.partial( ail_learning.ail_update_step, optimizer=optimizer, ail_network=ail_network, loss_fn=loss)) zero_transition = types.Transition( np.array([0.]), np.array([0.]), 0., 0., np.array([0.])) zero_transition = utils.add_batch_dim(zero_transition) one_transition = types.Transition( np.array([1.]), np.array([0.]), 0., 0., np.array([0.])) one_transition = utils.add_batch_dim(one_transition) key = jax.random.PRNGKey(0) discriminator_params, discriminator_state = discriminator.init(key) state = ail_learning.DiscriminatorTrainingState( optimizer_state=optimizer.init(discriminator_params), discriminator_params=discriminator_params, discriminator_state=discriminator_state, policy_params=None, key=key, steps=0, ) expected_loss = [1.062, 1.057, 1.052] for i in range(3): state, loss = step(state, (one_transition, zero_transition)) self.assertAlmostEqual(loss['total_loss'], expected_loss[i], places=3) if __name__ == '__main__': absltest.main()
acme-master
acme/agents/jax/ail/learning_test.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Implementations of a AIL agent.""" from acme.agents.jax.ail import losses from acme.agents.jax.ail import rewards from acme.agents.jax.ail.builder import AILBuilder from acme.agents.jax.ail.config import AILConfig from acme.agents.jax.ail.dac import DACBuilder from acme.agents.jax.ail.dac import DACConfig from acme.agents.jax.ail.gail import GAILBuilder from acme.agents.jax.ail.gail import GAILConfig from acme.agents.jax.ail.learning import AILLearner from acme.agents.jax.ail.networks import AILNetworks from acme.agents.jax.ail.networks import AIRLModule from acme.agents.jax.ail.networks import compute_ail_reward from acme.agents.jax.ail.networks import DiscriminatorMLP from acme.agents.jax.ail.networks import DiscriminatorModule from acme.agents.jax.ail.networks import make_discriminator
acme-master
acme/agents/jax/ail/__init__.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Adversarial Imitation Learning (AIL) Builder.""" import functools import itertools from typing import Callable, Generic, Iterator, List, Optional, Tuple from acme import adders from acme import core from acme import specs from acme import types from acme.adders import reverb as adders_reverb from acme.agents.jax import builders from acme.agents.jax.ail import config as ail_config from acme.agents.jax.ail import learning from acme.agents.jax.ail import losses from acme.agents.jax.ail import networks as ail_networks from acme.datasets import reverb as datasets from acme.jax import types as jax_types from acme.jax import utils from acme.jax.imitation_learning_types import DirectPolicyNetwork from acme.utils import counting from acme.utils import loggers from acme.utils import reverb_utils import jax import numpy as np import optax import reverb from reverb import rate_limiters import tree def _split_transitions( transitions: types.Transition, index: int) -> Tuple[types.Transition, types.Transition]: """Splits the given transition on the first axis at the given index. Args: transitions: Transitions to split. index: Spliting index. Returns: A pair of transitions, the first containing elements before the index (exclusive) and the second after the index (inclusive) """ return (tree.map_structure(lambda x: x[:index], transitions), tree.map_structure(lambda x: x[index:], transitions)) def _rebatch(iterator: Iterator[types.Transition], batch_size: int) -> Iterator[types.Transition]: """Rebatch the itererator with the given batch size. Args: iterator: Iterator to rebatch. batch_size: New batch size. Yields: Transitions with the new batch size. """ data = next(iterator) while True: while len(data.reward) < batch_size: # Ensure we can get enough demonstrations. next_data = next(iterator) data = tree.map_structure(lambda *args: np.concatenate(list(args)), data, next_data) output, data = _split_transitions(data, batch_size) yield output def _mix_arrays( replay: np.ndarray, demo: np.ndarray, index: int, seed: int) -> np.ndarray: """Mixes `replay` and `demo`. Args: replay: Replay data to mix. Only index element will be selected. demo: Demonstration data to mix. index: Amount of replay data we should include. seed: RNG seed. Returns: An array with replay elements up to 'index' and all the demos. """ # We're throwing away some replay data here. We have to if we want to make # sure the output info field is correct. output = np.concatenate((replay[:index], demo)) return np.random.default_rng(seed=seed).permutation(output) def _generate_samples_with_demonstrations( demonstration_iterator: Iterator[types.Transition], replay_iterator: Iterator[reverb.ReplaySample], policy_to_expert_data_ratio: int, batch_size) -> Iterator[reverb.ReplaySample]: """Generator which creates the sample having demonstrations in them. It takes the demonstrations and replay iterators and generates batches with same size as the replay iterator, such that each batches have the ratio of policy and expert data specified in policy_to_expert_data_ratio on average. There is no constraints on how the demonstrations and replay samples should be batched. Args: demonstration_iterator: Iterator of demonstrations. replay_iterator: Replay buffer sample iterator. policy_to_expert_data_ratio: Amount of policy transitions for 1 expert transitions. batch_size: Output batch size, which should match the replay batch size. Yields: Samples having a mix of demonstrations and policy data. The info will match the current replay sample info and the batch size will be the same as the replay_iterator data batch size. """ count = 0 if batch_size % (policy_to_expert_data_ratio + 1) != 0: raise ValueError( 'policy_to_expert_data_ratio + 1 must divide the batch size but ' f'{batch_size} % {policy_to_expert_data_ratio+1} !=0') demo_insertion_size = batch_size // (policy_to_expert_data_ratio + 1) policy_insertion_size = batch_size - demo_insertion_size demonstration_iterator = _rebatch(demonstration_iterator, demo_insertion_size) for sample, demos in zip(replay_iterator, demonstration_iterator): output_transitions = tree.map_structure( functools.partial(_mix_arrays, index=policy_insertion_size, seed=count), sample.data, demos) count += 1 yield reverb.ReplaySample(info=sample.info, data=output_transitions) class AILBuilder(builders.ActorLearnerBuilder[ail_networks.AILNetworks, DirectPolicyNetwork, learning.AILSample], Generic[ail_networks.DirectRLNetworks, DirectPolicyNetwork]): """AIL Builder.""" def __init__( self, rl_agent: builders.ActorLearnerBuilder[ail_networks.DirectRLNetworks, DirectPolicyNetwork, reverb.ReplaySample], config: ail_config.AILConfig, discriminator_loss: losses.Loss, make_demonstrations: Callable[[int], Iterator[types.Transition]]): """Implements a builder for AIL using rl_agent as forward RL algorithm. Args: rl_agent: The standard RL agent used by AIL to optimize the generator. config: a AIL config discriminator_loss: The loss function for the discriminator to minimize. make_demonstrations: A function that returns an iterator with demonstrations to be imitated. """ self._rl_agent = rl_agent self._config = config self._discriminator_loss = discriminator_loss self._make_demonstrations = make_demonstrations def make_learner(self, random_key: jax_types.PRNGKey, networks: ail_networks.AILNetworks, dataset: Iterator[learning.AILSample], logger_fn: loggers.LoggerFactory, environment_spec: specs.EnvironmentSpec, replay_client: Optional[reverb.Client] = None, counter: Optional[counting.Counter] = None) -> core.Learner: counter = counter or counting.Counter() direct_rl_counter = counting.Counter(counter, 'direct_rl') batch_size_per_learner_step = ail_config.get_per_learner_step_batch_size( self._config) direct_rl_learner_key, discriminator_key = jax.random.split(random_key) direct_rl_learner = functools.partial( self._rl_agent.make_learner, direct_rl_learner_key, networks.direct_rl_networks, logger_fn=logger_fn, environment_spec=environment_spec, replay_client=replay_client, counter=direct_rl_counter) discriminator_optimizer = ( self._config.discriminator_optimizer or optax.adam(1e-5)) return learning.AILLearner( counter, direct_rl_learner_factory=direct_rl_learner, loss_fn=self._discriminator_loss, iterator=dataset, discriminator_optimizer=discriminator_optimizer, ail_network=networks, discriminator_key=discriminator_key, is_sequence_based=self._config.is_sequence_based, num_sgd_steps_per_step=batch_size_per_learner_step // self._config.discriminator_batch_size, policy_variable_name=self._config.policy_variable_name, logger=logger_fn('learner', steps_key=counter.get_steps_key())) def make_replay_tables( self, environment_spec: specs.EnvironmentSpec, policy: DirectPolicyNetwork, ) -> List[reverb.Table]: replay_tables = self._rl_agent.make_replay_tables(environment_spec, policy) if self._config.share_iterator: return replay_tables replay_tables.append( reverb.Table( name=self._config.replay_table_name, sampler=reverb.selectors.Uniform(), remover=reverb.selectors.Fifo(), max_size=self._config.max_replay_size, rate_limiter=rate_limiters.MinSize(self._config.min_replay_size), signature=adders_reverb.NStepTransitionAdder.signature( environment_spec))) return replay_tables # This function does not expose all the iterators used by the learner when # share_iterator is False, making further wrapping impossible. # TODO(eorsini): Expose all iterators. # Currently GAIL uses 3 iterators, instead we can make it use a single # iterator and return this one here. The way to achieve this would be: # * Create the 3 iterators here. # * zip them and return them here. # * upzip them in the learner (this step will not be necessary once we move to # stateless learners) # This should work fine as the 3 iterators are always iterated in parallel # (i.e. at every step we call next once on each of them). def make_dataset_iterator( self, replay_client: reverb.Client) -> Iterator[learning.AILSample]: batch_size_per_learner_step = ail_config.get_per_learner_step_batch_size( self._config) iterator_demonstration = self._make_demonstrations( batch_size_per_learner_step) direct_iterator = self._rl_agent.make_dataset_iterator(replay_client) if self._config.share_iterator: # In order to reuse the iterator return values and not lose a 2x factor on # sample efficiency, we need to use itertools.tee(). discriminator_iterator, direct_iterator = itertools.tee(direct_iterator) else: discriminator_iterator = datasets.make_reverb_dataset( table=self._config.replay_table_name, server_address=replay_client.server_address, batch_size=ail_config.get_per_learner_step_batch_size(self._config), prefetch_size=self._config.prefetch_size).as_numpy_iterator() if self._config.policy_to_expert_data_ratio is not None: iterator_demonstration, iterator_demonstration2 = itertools.tee( iterator_demonstration) direct_iterator = _generate_samples_with_demonstrations( iterator_demonstration2, direct_iterator, self._config.policy_to_expert_data_ratio, self._config.direct_rl_batch_size) is_sequence_based = self._config.is_sequence_based # Don't flatten the discriminator batch if the iterator is not shared. process_discriminator_sample = functools.partial( reverb_utils.replay_sample_to_sars_transition, is_sequence=is_sequence_based and self._config.share_iterator, flatten_batch=is_sequence_based and self._config.share_iterator, strip_last_transition=is_sequence_based and self._config.share_iterator) discriminator_iterator = ( # Remove the extras to have the same nested structure as demonstrations. process_discriminator_sample(sample)._replace(extras=()) for sample in discriminator_iterator) return utils.device_put((learning.AILSample(*sample) for sample in zip( discriminator_iterator, direct_iterator, iterator_demonstration)), jax.devices()[0]) def make_adder( self, replay_client: reverb.Client, environment_spec: Optional[specs.EnvironmentSpec], policy: Optional[DirectPolicyNetwork]) -> Optional[adders.Adder]: direct_rl_adder = self._rl_agent.make_adder(replay_client, environment_spec, policy) if self._config.share_iterator: return direct_rl_adder ail_adder = adders_reverb.NStepTransitionAdder( priority_fns={self._config.replay_table_name: None}, client=replay_client, n_step=1, discount=self._config.discount) # Some direct rl algorithms (such as PPO), might be passing extra data # which we won't be able to process here properly, so we need to ignore them return adders.ForkingAdder( [adders.IgnoreExtrasAdder(ail_adder), direct_rl_adder]) def make_actor( self, random_key: jax_types.PRNGKey, policy: DirectPolicyNetwork, environment_spec: specs.EnvironmentSpec, variable_source: Optional[core.VariableSource] = None, adder: Optional[adders.Adder] = None, ) -> core.Actor: return self._rl_agent.make_actor(random_key, policy, environment_spec, variable_source, adder) def make_policy(self, networks: ail_networks.AILNetworks, environment_spec: specs.EnvironmentSpec, evaluation: bool = False) -> DirectPolicyNetwork: return self._rl_agent.make_policy(networks.direct_rl_networks, environment_spec, evaluation)
acme-master
acme/agents/jax/ail/builder.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Networks definitions for the BC agent. AIRL network architecture follows https://arxiv.org/pdf/1710.11248.pdf. """ import dataclasses import functools from typing import Any, Callable, Generic, Iterable, Optional from acme import specs from acme import types from acme.jax import networks as networks_lib from acme.jax import utils from acme.jax.imitation_learning_types import DirectRLNetworks import haiku as hk import jax from jax import numpy as jnp import numpy as np # Function from discriminator logit to imitation reward. ImitationRewardFn = Callable[[networks_lib.Logits], jnp.ndarray] State = networks_lib.Params @dataclasses.dataclass class AILNetworks(Generic[DirectRLNetworks]): """AIL networks data class. Attributes: discriminator_network: Networks which takes as input: (observations, actions, next_observations, direct_rl_params) to return the logit of the discriminator. If the discriminator does not need direct_rl_params you can pass (). imitation_reward_fn: Function from logit of the discriminator to imitation reward. direct_rl_networks: Networks of the direct RL algorithm. """ discriminator_network: networks_lib.FeedForwardNetwork imitation_reward_fn: ImitationRewardFn direct_rl_networks: DirectRLNetworks def compute_ail_reward(discriminator_params: networks_lib.Params, discriminator_state: State, policy_params: Optional[networks_lib.Params], transitions: types.Transition, networks: AILNetworks) -> jnp.ndarray: """Computes the AIL reward for a given transition. Args: discriminator_params: Parameters of the discriminator network. discriminator_state: State of the discriminator network. policy_params: Parameters of the direct RL policy. transitions: Transitions to compute the reward for. networks: AIL networks. Returns: The rewards as an ndarray. """ logits, _ = networks.discriminator_network.apply( discriminator_params, policy_params, discriminator_state, transitions, is_training=False, rng=None) return networks.imitation_reward_fn(logits) class SpectralNormalizedLinear(hk.Module): """SpectralNormalizedLinear module. This is a Linear layer with a upper-bounded Lipschitz. It is used in iResNet. Reference: Behrmann et al. Invertible Residual Networks. ICML 2019. https://arxiv.org/pdf/1811.00995.pdf """ def __init__( self, output_size: int, lipschitz_coeff: float, with_bias: bool = True, w_init: Optional[hk.initializers.Initializer] = None, b_init: Optional[hk.initializers.Initializer] = None, name: Optional[str] = None, ): """Constructs the SpectralNormalizedLinear module. Args: output_size: Output dimensionality. lipschitz_coeff: Spectral normalization coefficient. with_bias: Whether to add a bias to the output. w_init: Optional initializer for weights. By default, uses random values from truncated normal, with stddev ``1 / sqrt(fan_in)``. See https://arxiv.org/abs/1502.03167v3. b_init: Optional initializer for bias. By default, zero. name: Name of the module. """ super().__init__(name=name) self.input_size = None self.output_size = output_size self.with_bias = with_bias self.w_init = w_init self.b_init = b_init or jnp.zeros self.lipschitz_coeff = lipschitz_coeff self.num_iterations = 100 self.eps = 1e-6 def get_normalized_weights(self, weights: jnp.ndarray, renormalize: bool = False) -> jnp.ndarray: def _l2_normalize(x, axis=None, eps=1e-12): return x * jax.lax.rsqrt((x * x).sum(axis=axis, keepdims=True) + eps) output_size = self.output_size dtype = weights.dtype assert output_size == weights.shape[-1] sigma = hk.get_state('sigma', (), init=jnp.ones) if renormalize: # Power iterations to compute spectral norm V*W*U^T. u = hk.get_state( 'u', (1, output_size), dtype, init=hk.initializers.RandomNormal()) for _ in range(self.num_iterations): v = _l2_normalize(jnp.matmul(u, weights.transpose()), eps=self.eps) u = _l2_normalize(jnp.matmul(v, weights), eps=self.eps) u = jax.lax.stop_gradient(u) v = jax.lax.stop_gradient(v) sigma = jnp.matmul(jnp.matmul(v, weights), jnp.transpose(u))[0, 0] hk.set_state('u', u) hk.set_state('v', v) hk.set_state('sigma', sigma) factor = jnp.maximum(1, sigma / self.lipschitz_coeff) return weights / factor def __call__(self, inputs: jnp.ndarray) -> jnp.ndarray: """Computes a linear transform of the input.""" if not inputs.shape: raise ValueError('Input must not be scalar.') input_size = self.input_size = inputs.shape[-1] output_size = self.output_size dtype = inputs.dtype w_init = self.w_init if w_init is None: stddev = 1. / np.sqrt(self.input_size) w_init = hk.initializers.TruncatedNormal(stddev=stddev) w = hk.get_parameter('w', [input_size, output_size], dtype, init=w_init) w = self.get_normalized_weights(w, renormalize=True) out = jnp.dot(inputs, w) if self.with_bias: b = hk.get_parameter('b', [self.output_size], dtype, init=self.b_init) b = jnp.broadcast_to(b, out.shape) out = out + b return out class DiscriminatorMLP(hk.Module): """A multi-layer perceptron module.""" def __init__( self, hidden_layer_sizes: Iterable[int], w_init: Optional[hk.initializers.Initializer] = None, b_init: Optional[hk.initializers.Initializer] = None, with_bias: bool = True, activation: Callable[[jnp.ndarray], jnp.ndarray] = jax.nn.relu, input_dropout_rate: float = 0., hidden_dropout_rate: float = 0., spectral_normalization_lipschitz_coeff: Optional[float] = None, name: Optional[str] = None ): """Constructs an MLP. Args: hidden_layer_sizes: Hiddent layer sizes. w_init: Initializer for :class:`~haiku.Linear` weights. b_init: Initializer for :class:`~haiku.Linear` bias. Must be ``None`` if ``with_bias=False``. with_bias: Whether or not to apply a bias in each layer. activation: Activation function to apply between :class:`~haiku.Linear` layers. Defaults to ReLU. input_dropout_rate: Dropout on the input. hidden_dropout_rate: Dropout on the hidden layer outputs. spectral_normalization_lipschitz_coeff: If not None, the network will have spectral normalization with the given constant. name: Optional name for this module. Raises: ValueError: If ``with_bias`` is ``False`` and ``b_init`` is not ``None``. """ if not with_bias and b_init is not None: raise ValueError('When with_bias=False b_init must not be set.') super().__init__(name=name) self._activation = activation self._input_dropout_rate = input_dropout_rate self._hidden_dropout_rate = hidden_dropout_rate layer_sizes = list(hidden_layer_sizes) + [1] if spectral_normalization_lipschitz_coeff is not None: layer_lipschitz_coeff = np.power(spectral_normalization_lipschitz_coeff, 1. / len(layer_sizes)) layer_module = functools.partial( SpectralNormalizedLinear, lipschitz_coeff=layer_lipschitz_coeff, w_init=w_init, b_init=b_init, with_bias=with_bias) else: layer_module = functools.partial( hk.Linear, w_init=w_init, b_init=b_init, with_bias=with_bias) layers = [] for index, output_size in enumerate(layer_sizes): layers.append( layer_module(output_size=output_size, name=f'linear_{index}')) self._layers = tuple(layers) def __call__( self, inputs: jnp.ndarray, is_training: bool, rng: Optional[networks_lib.PRNGKey], ) -> networks_lib.Logits: rng = hk.PRNGSequence(rng) if rng is not None else None out = inputs for i, layer in enumerate(self._layers): if is_training: dropout_rate = ( self._input_dropout_rate if i == 0 else self._hidden_dropout_rate) out = hk.dropout(next(rng), dropout_rate, out) out = layer(out) if i < len(self._layers) - 1: out = self._activation(out) return out class DiscriminatorModule(hk.Module): """Discriminator module that concatenates its inputs.""" def __init__(self, environment_spec: specs.EnvironmentSpec, use_action: bool, use_next_obs: bool, network_core: Callable[..., Any], observation_embedding: Callable[[networks_lib.Observation], jnp.ndarray] = lambda x: x, name='discriminator'): super().__init__(name=name) self._use_action = use_action self._environment_spec = environment_spec self._use_next_obs = use_next_obs self._network_core = network_core self._observation_embedding = observation_embedding def __call__(self, observations: networks_lib.Observation, actions: networks_lib.Action, next_observations: networks_lib.Observation, is_training: bool, rng: networks_lib.PRNGKey) -> networks_lib.Logits: observations = self._observation_embedding(observations) if self._use_next_obs: next_observations = self._observation_embedding(next_observations) data = jnp.concatenate([observations, next_observations], axis=-1) else: data = observations if self._use_action: action_spec = self._environment_spec.actions if isinstance(action_spec, specs.DiscreteArray): actions = jax.nn.one_hot(actions, action_spec.num_values) data = jnp.concatenate([data, actions], axis=-1) output = self._network_core(data, is_training, rng) output = jnp.squeeze(output, axis=-1) return output class AIRLModule(hk.Module): """AIRL Module.""" def __init__(self, environment_spec: specs.EnvironmentSpec, use_action: bool, use_next_obs: bool, discount: float, g_core: Callable[..., Any], h_core: Callable[..., Any], observation_embedding: Callable[[networks_lib.Observation], jnp.ndarray] = lambda x: x, name='airl'): super().__init__(name=name) self._environment_spec = environment_spec self._use_action = use_action self._use_next_obs = use_next_obs self._discount = discount self._g_core = g_core self._h_core = h_core self._observation_embedding = observation_embedding def __call__(self, observations: networks_lib.Observation, actions: networks_lib.Action, next_observations: networks_lib.Observation, is_training: bool, rng: networks_lib.PRNGKey) -> networks_lib.Logits: g_output = DiscriminatorModule( environment_spec=self._environment_spec, use_action=self._use_action, use_next_obs=self._use_next_obs, network_core=self._g_core, observation_embedding=self._observation_embedding, name='airl_g')(observations, actions, next_observations, is_training, rng) h_module = DiscriminatorModule( environment_spec=self._environment_spec, use_action=False, use_next_obs=False, network_core=self._h_core, observation_embedding=self._observation_embedding, name='airl_h') return (g_output + self._discount * h_module(next_observations, (), (), is_training, rng) - h_module(observations, (), (), is_training, rng)) # TODO(eorsini): Manipulate FeedForwardNetworks instead of transforms to # increase compatibility with Flax. def make_discriminator( environment_spec: specs.EnvironmentSpec, discriminator_transformed: hk.TransformedWithState, logpi_fn: Optional[Callable[ [networks_lib.Params, networks_lib.Observation, networks_lib.Action], jnp.ndarray]] = None ) -> networks_lib.FeedForwardNetwork: """Creates the discriminator network. Args: environment_spec: Environment spec discriminator_transformed: Haiku transformed of the discriminator. logpi_fn: If the policy logpi function is provided, its output will be removed from the discriminator logit. Returns: The network. """ def apply_fn(params: hk.Params, policy_params: networks_lib.Params, state: hk.State, transitions: types.Transition, is_training: bool, rng: networks_lib.PRNGKey) -> networks_lib.Logits: output, state = discriminator_transformed.apply( params, state, transitions.observation, transitions.action, transitions.next_observation, is_training, rng) if logpi_fn is not None: logpi = logpi_fn(policy_params, transitions.observation, transitions.action) # Quick Maths: # D = exp(output)/(exp(output) + pi(a|s)) # logit(D) = log(D/(1-D)) = log(exp(output)/pi(a|s)) # logit(D) = output - logpi return output - logpi, state # pytype: disable=bad-return-type # jax-ndarray return output, state # pytype: disable=bad-return-type # jax-ndarray dummy_obs = utils.zeros_like(environment_spec.observations) dummy_obs = utils.add_batch_dim(dummy_obs) dummy_actions = utils.zeros_like(environment_spec.actions) dummy_actions = utils.add_batch_dim(dummy_actions) return networks_lib.FeedForwardNetwork( # pylint: disable=g-long-lambda init=lambda rng: discriminator_transformed.init( rng, dummy_obs, dummy_actions, dummy_obs, False, rng), apply=apply_fn)
acme-master
acme/agents/jax/ail/networks.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """AIL logits to AIL reward.""" from typing import Optional from acme.agents.jax.ail import networks as ail_networks from acme.jax import networks as networks_lib import jax import jax.numpy as jnp def fairl_reward( max_reward_magnitude: Optional[float] = None ) -> ail_networks.ImitationRewardFn: """The FAIRL reward function (https://arxiv.org/pdf/1911.02256.pdf). Args: max_reward_magnitude: Clipping value for the reward. Returns: The function from logit to imitation reward. """ def imitation_reward(logits: networks_lib.Logits) -> float: rewards = jnp.exp(jnp.clip(logits, a_max=20.)) * -logits if max_reward_magnitude is not None: # pylint: disable=invalid-unary-operand-type rewards = jnp.clip( rewards, a_min=-max_reward_magnitude, a_max=max_reward_magnitude) return rewards # pytype: disable=bad-return-type # jax-types return imitation_reward # pytype: disable=bad-return-type # jax-ndarray def gail_reward( reward_balance: float = .5, max_reward_magnitude: Optional[float] = None ) -> ail_networks.ImitationRewardFn: """GAIL reward function (https://arxiv.org/pdf/1606.03476.pdf). Args: reward_balance: 1 means log(D) reward, 0 means -log(1-D) and other values mean an average of the two. max_reward_magnitude: Clipping value for the reward. Returns: The function from logit to imitation reward. """ def imitation_reward(logits: networks_lib.Logits) -> float: # Quick Maths: # logits = ln(D) - ln(1-D) # -softplus(-logits) = ln(D) # softplus(logits) = -ln(1-D) rewards = ( reward_balance * -jax.nn.softplus(-logits) + (1 - reward_balance) * jax.nn.softplus(logits)) if max_reward_magnitude is not None: # pylint: disable=invalid-unary-operand-type rewards = jnp.clip( rewards, a_min=-max_reward_magnitude, a_max=max_reward_magnitude) return rewards return imitation_reward # pytype: disable=bad-return-type # jax-ndarray
acme-master
acme/agents/jax/ail/rewards.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """AIL discriminator losses.""" import functools from typing import Callable, Dict, Optional, Tuple from acme import types from acme.jax import networks as networks_lib import jax import jax.numpy as jnp import tensorflow_probability as tfp import tree tfp = tfp.experimental.substrates.jax tfd = tfp.distributions # The loss is a function taking the discriminator, its state, the demo # transition and the replay buffer transitions. # It returns the loss as a float and a debug dictionary with the new state. State = networks_lib.Params DiscriminatorOutput = Tuple[networks_lib.Logits, State] DiscriminatorFn = Callable[[State, types.Transition], DiscriminatorOutput] Metrics = Dict[str, float] LossOutput = Tuple[float, Tuple[Metrics, State]] Loss = Callable[[ DiscriminatorFn, State, types.Transition, types.Transition, networks_lib .PRNGKey ], LossOutput] def _binary_cross_entropy_loss(logit: jnp.ndarray, label: jnp.ndarray) -> jnp.ndarray: return label * jax.nn.softplus(-logit) + (1 - label) * jax.nn.softplus(logit) @jax.vmap def _weighted_average(x: jnp.ndarray, y: jnp.ndarray, lambdas: jnp.ndarray) -> jnp.ndarray: return lambdas * x + (1. - lambdas) * y def _label_data( rb_transitions: types.Transition, demonstration_transitions: types.Transition, mixup_alpha: Optional[float], key: networks_lib.PRNGKey) -> Tuple[types.Transition, jnp.ndarray]: """Create a tuple data, labels by concatenating the rb and dem transitions.""" data = tree.map_structure(lambda x, y: jnp.concatenate([x, y]), rb_transitions, demonstration_transitions) labels = jnp.concatenate([ jnp.zeros(rb_transitions.reward.shape), jnp.ones(demonstration_transitions.reward.shape) ]) if mixup_alpha is not None: lambda_key, mixup_key = jax.random.split(key) lambdas = tfd.Beta(mixup_alpha, mixup_alpha).sample( len(labels), seed=lambda_key) shuffled_data = tree.map_structure( lambda x: jax.random.permutation(key=mixup_key, x=x), data) shuffled_labels = jax.random.permutation(key=mixup_key, x=labels) data = tree.map_structure(lambda x, y: _weighted_average(x, y, lambdas), data, shuffled_data) labels = _weighted_average(labels, shuffled_labels, lambdas) return data, labels def _logit_bernoulli_entropy(logits: networks_lib.Logits) -> jnp.ndarray: return (1. - jax.nn.sigmoid(logits)) * logits - jax.nn.log_sigmoid(logits) def gail_loss(entropy_coefficient: float = 0., mixup_alpha: Optional[float] = None) -> Loss: """Computes the standard GAIL loss.""" def loss_fn( discriminator_fn: DiscriminatorFn, discriminator_state: State, demo_transitions: types.Transition, rb_transitions: types.Transition, rng_key: networks_lib.PRNGKey) -> LossOutput: data, labels = _label_data( rb_transitions=rb_transitions, demonstration_transitions=demo_transitions, mixup_alpha=mixup_alpha, key=rng_key) logits, discriminator_state = discriminator_fn(discriminator_state, data) classification_loss = jnp.mean(_binary_cross_entropy_loss(logits, labels)) entropy = jnp.mean(_logit_bernoulli_entropy(logits)) entropy_loss = -entropy_coefficient * entropy total_loss = classification_loss + entropy_loss metrics = { 'total_loss': total_loss, 'entropy_loss': entropy_loss, 'classification_loss': classification_loss } return total_loss, (metrics, discriminator_state) return loss_fn def pugail_loss(positive_class_prior: float, entropy_coefficient: float, pugail_beta: Optional[float] = None) -> Loss: """Computes the PUGAIL loss (https://arxiv.org/pdf/1911.00459.pdf).""" def loss_fn( discriminator_fn: DiscriminatorFn, discriminator_state: State, demo_transitions: types.Transition, rb_transitions: types.Transition, rng_key: networks_lib.PRNGKey) -> LossOutput: del rng_key demo_logits, discriminator_state = discriminator_fn(discriminator_state, demo_transitions) rb_logits, discriminator_state = discriminator_fn(discriminator_state, rb_transitions) # Quick Maths: # output = logit(D) = ln(D) - ln(1-D) # -softplus(-output) = ln(D) # softplus(output) = -ln(1-D) # prior * -ln(D(expert)) positive_loss = positive_class_prior * -jax.nn.log_sigmoid(demo_logits) # -ln(1 - D(policy)) - prior * -ln(1 - D(expert)) negative_loss = jax.nn.softplus( rb_logits) - positive_class_prior * jax.nn.softplus(demo_logits) if pugail_beta is not None: negative_loss = jnp.clip(negative_loss, a_min=-1. * pugail_beta) classification_loss = jnp.mean(positive_loss + negative_loss) entropy = jnp.mean( _logit_bernoulli_entropy(jnp.concatenate([demo_logits, rb_logits]))) entropy_loss = -entropy_coefficient * entropy total_loss = classification_loss + entropy_loss metrics = { 'total_loss': total_loss, 'positive_loss': jnp.mean(positive_loss), 'negative_loss': jnp.mean(negative_loss), 'demo_logits': jnp.mean(demo_logits), 'rb_logits': jnp.mean(rb_logits), 'entropy_loss': entropy_loss, 'classification_loss': classification_loss } return total_loss, (metrics, discriminator_state) return loss_fn def _make_gradient_penalty_data(rb_transitions: types.Transition, demonstration_transitions: types.Transition, key: networks_lib.PRNGKey) -> types.Transition: lambdas = tfd.Uniform().sample(len(rb_transitions.reward), seed=key) return tree.map_structure(lambda x, y: _weighted_average(x, y, lambdas), rb_transitions, demonstration_transitions) @functools.partial(jax.vmap, in_axes=(0, None, None)) def _compute_gradient_penalty(gradient_penalty_data: types.Transition, discriminator_fn: Callable[[types.Transition], float], gradient_penalty_target: float) -> float: """Computes a penalty based on the gradient norm on the data.""" # The input should not be batched. assert not gradient_penalty_data.reward.shape discriminator_gradient_fn = jax.grad(discriminator_fn) gradients = discriminator_gradient_fn(gradient_penalty_data) gradients = tree.map_structure(lambda x: x.flatten(), gradients) gradients = jnp.concatenate([gradients.observation, gradients.action, gradients.next_observation]) gradient_norms = jnp.linalg.norm(gradients + 1e-8) k = gradient_penalty_target * jnp.ones_like(gradient_norms) return jnp.mean(jnp.square(gradient_norms - k)) def add_gradient_penalty(base_loss: Loss, gradient_penalty_coefficient: float, gradient_penalty_target: float) -> Loss: """Adds a gradient penalty to the base_loss.""" if not gradient_penalty_coefficient: return base_loss def loss_fn(discriminator_fn: DiscriminatorFn, discriminator_state: State, demo_transitions: types.Transition, rb_transitions: types.Transition, rng_key: networks_lib.PRNGKey) -> LossOutput: super_key, gradient_penalty_key = jax.random.split(rng_key) partial_loss, (losses, discriminator_state) = base_loss( discriminator_fn, discriminator_state, demo_transitions, rb_transitions, super_key) gradient_penalty_data = _make_gradient_penalty_data( rb_transitions=rb_transitions, demonstration_transitions=demo_transitions, key=gradient_penalty_key) def apply_discriminator_fn(transitions: types.Transition) -> float: logits, _ = discriminator_fn(discriminator_state, transitions) return logits # pytype: disable=bad-return-type # jax-ndarray gradient_penalty = gradient_penalty_coefficient * jnp.mean( _compute_gradient_penalty(gradient_penalty_data, apply_discriminator_fn, gradient_penalty_target)) losses['gradient_penalty'] = gradient_penalty total_loss = partial_loss + gradient_penalty losses['total_loss'] = total_loss return total_loss, (losses, discriminator_state) return loss_fn
acme-master
acme/agents/jax/ail/losses.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """AIL learner implementation.""" import functools import itertools import time from typing import Any, Callable, Iterator, List, NamedTuple, Optional, Tuple import acme from acme import types from acme.agents.jax.ail import losses from acme.agents.jax.ail import networks as ail_networks from acme.jax import networks as networks_lib from acme.jax import utils from acme.utils import counting from acme.utils import loggers from acme.utils import reverb_utils import jax import optax import reverb class DiscriminatorTrainingState(NamedTuple): """Contains training state for the discriminator.""" # State of the optimizer used to optimize the discriminator parameters. optimizer_state: optax.OptState # Parameters of the discriminator. discriminator_params: networks_lib.Params # State of the discriminator discriminator_state: losses.State # For AIRL variants, we need the policy params to compute the loss. policy_params: Optional[networks_lib.Params] # Key for random number generation. key: networks_lib.PRNGKey # Training step of the discriminator. steps: int class TrainingState(NamedTuple): """Contains training state of the AIL learner.""" rewarder_state: DiscriminatorTrainingState learner_state: Any def ail_update_step( state: DiscriminatorTrainingState, data: Tuple[types.Transition, types.Transition], optimizer: optax.GradientTransformation, ail_network: ail_networks.AILNetworks, loss_fn: losses.Loss) -> Tuple[DiscriminatorTrainingState, losses.Metrics]: """Run an update steps on the given transitions. Args: state: The learner state. data: Demo and rb transitions. optimizer: Discriminator optimizer. ail_network: AIL networks. loss_fn: Discriminator loss to minimize. Returns: A new state and metrics. """ demo_transitions, rb_transitions = data key, discriminator_key, loss_key = jax.random.split(state.key, 3) def compute_loss( discriminator_params: networks_lib.Params) -> losses.LossOutput: discriminator_fn = functools.partial( ail_network.discriminator_network.apply, discriminator_params, state.policy_params, is_training=True, rng=discriminator_key) return loss_fn(discriminator_fn, state.discriminator_state, demo_transitions, rb_transitions, loss_key) loss_grad = jax.grad(compute_loss, has_aux=True) grads, (loss, new_discriminator_state) = loss_grad(state.discriminator_params) update, optimizer_state = optimizer.update( grads, state.optimizer_state, params=state.discriminator_params) discriminator_params = optax.apply_updates(state.discriminator_params, update) new_state = DiscriminatorTrainingState( optimizer_state=optimizer_state, discriminator_params=discriminator_params, discriminator_state=new_discriminator_state, policy_params=state.policy_params, # Not modified. key=key, steps=state.steps + 1, ) return new_state, loss class AILSample(NamedTuple): discriminator_sample: types.Transition direct_sample: reverb.ReplaySample demonstration_sample: types.Transition class AILLearner(acme.Learner): """AIL learner.""" def __init__( self, counter: counting.Counter, direct_rl_learner_factory: Callable[[Iterator[reverb.ReplaySample]], acme.Learner], loss_fn: losses.Loss, iterator: Iterator[AILSample], discriminator_optimizer: optax.GradientTransformation, ail_network: ail_networks.AILNetworks, discriminator_key: networks_lib.PRNGKey, is_sequence_based: bool, num_sgd_steps_per_step: int = 1, policy_variable_name: Optional[str] = None, logger: Optional[loggers.Logger] = None): """AIL Learner. Args: counter: Counter. direct_rl_learner_factory: Function that creates the direct RL learner when passed a replay sample iterator. loss_fn: Discriminator loss. iterator: Iterator that returns AILSamples. discriminator_optimizer: Discriminator optax optimizer. ail_network: AIL networks. discriminator_key: RNG key. is_sequence_based: If True, a direct rl algorithm is using SequenceAdder data format. Otherwise the learner assumes that the direct rl algorithm is using NStepTransitionAdder. num_sgd_steps_per_step: Number of discriminator gradient updates per step. policy_variable_name: The name of the policy variable to retrieve direct_rl policy parameters. logger: Logger. """ self._is_sequence_based = is_sequence_based state_key, networks_key = jax.random.split(discriminator_key) # Generator expression that works the same as an iterator. # https://pymbook.readthedocs.io/en/latest/igd.html#generator-expressions iterator, direct_rl_iterator = itertools.tee(iterator) direct_rl_iterator = ( self._process_sample(sample.direct_sample) for sample in direct_rl_iterator) self._direct_rl_learner = direct_rl_learner_factory(direct_rl_iterator) self._iterator = iterator if policy_variable_name is not None: def get_policy_params(): return self._direct_rl_learner.get_variables([policy_variable_name])[0] self._get_policy_params = get_policy_params else: self._get_policy_params = lambda: None # General learner book-keeping and loggers. self._counter = counter or counting.Counter() self._logger = logger or loggers.make_default_logger( 'learner', asynchronous=True, serialize_fn=utils.fetch_devicearray, steps_key=self._counter.get_steps_key()) # Use the JIT compiler. self._update_step = functools.partial( ail_update_step, optimizer=discriminator_optimizer, ail_network=ail_network, loss_fn=loss_fn) self._update_step = utils.process_multiple_batches(self._update_step, num_sgd_steps_per_step) self._update_step = jax.jit(self._update_step) discriminator_params, discriminator_state = ( ail_network.discriminator_network.init(networks_key)) self._state = DiscriminatorTrainingState( optimizer_state=discriminator_optimizer.init(discriminator_params), discriminator_params=discriminator_params, discriminator_state=discriminator_state, policy_params=self._get_policy_params(), key=state_key, steps=0, ) # Do not record timestamps until after the first learning step is done. # This is to avoid including the time it takes for actors to come online and # fill the replay buffer. self._timestamp = None self._get_reward = jax.jit( functools.partial( ail_networks.compute_ail_reward, networks=ail_network)) def _process_sample(self, sample: reverb.ReplaySample) -> reverb.ReplaySample: """Updates the reward of the replay sample. Args: sample: Replay sample to update the reward to. Returns: The replay sample with an updated reward. """ transitions = reverb_utils.replay_sample_to_sars_transition( sample, is_sequence=self._is_sequence_based) rewards = self._get_reward(self._state.discriminator_params, self._state.discriminator_state, self._state.policy_params, transitions) return sample._replace(data=sample.data._replace(reward=rewards)) def step(self): sample = next(self._iterator) rb_transitions = sample.discriminator_sample demo_transitions = sample.demonstration_sample if demo_transitions.reward.shape != rb_transitions.reward.shape: raise ValueError( 'Different shapes for demo transitions and rb_transitions: ' f'{demo_transitions.reward.shape} != {rb_transitions.reward.shape}') # Update the parameters of the policy before doing a gradient step. state = self._state._replace(policy_params=self._get_policy_params()) self._state, metrics = self._update_step(state, (demo_transitions, rb_transitions)) # The order is important for AIRL. # In AIRL, the discriminator update depends on the logpi of the direct rl # policy. # When updating the discriminator, we want the logpi for which the # transitions were made with and not an updated one. # Get data from replay (dropping extras if any). Note there is no # extra data here because we do not insert any into Reverb. self._direct_rl_learner.step() # Compute elapsed time. timestamp = time.time() elapsed_time = timestamp - self._timestamp if self._timestamp else 0 self._timestamp = timestamp # Increment counts and record the current time. counts = self._counter.increment(steps=1, walltime=elapsed_time) # Attempts to write the logs. self._logger.write({**metrics, **counts}) def get_variables(self, names: List[str]) -> List[Any]: rewarder_dict = {'discriminator': self._state.discriminator_params} learner_names = [name for name in names if name not in rewarder_dict] learner_dict = {} if learner_names: learner_dict = dict( zip(learner_names, self._direct_rl_learner.get_variables(learner_names))) variables = [ rewarder_dict.get(name, learner_dict.get(name, None)) for name in names ] return variables def save(self) -> TrainingState: return TrainingState( rewarder_state=self._state, learner_state=self._direct_rl_learner.save()) def restore(self, state: TrainingState): self._state = state.rewarder_state self._direct_rl_learner.restore(state.learner_state)
acme-master
acme/agents/jax/ail/learning.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Builder for GAIL. https://arxiv.org/pdf/1606.03476.pdf """ import dataclasses from typing import Callable, Iterator from acme import types from acme.agents.jax import actor_core as actor_core_lib from acme.agents.jax import ppo from acme.agents.jax.ail import builder from acme.agents.jax.ail import config as ail_config from acme.agents.jax.ail import losses @dataclasses.dataclass class GAILConfig: """Configuration options specific to GAIL.""" ail_config: ail_config.AILConfig ppo_config: ppo.PPOConfig class GAILBuilder(builder.AILBuilder[ppo.PPONetworks, actor_core_lib.FeedForwardPolicyWithExtra] ): """GAIL Builder.""" def __init__(self, config: GAILConfig, make_demonstrations: Callable[[int], Iterator[types.Transition]]): ppo_builder = ppo.PPOBuilder(config.ppo_config) super().__init__( ppo_builder, config=config.ail_config, discriminator_loss=losses.gail_loss(), make_demonstrations=make_demonstrations)
acme-master
acme/agents/jax/ail/gail.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Tests for the AIL discriminator losses.""" from acme import types from acme.agents.jax.ail import losses from acme.jax import networks as networks_lib import jax import jax.numpy as jnp import tree from absl.testing import absltest class AilLossTest(absltest.TestCase): def test_gradient_penalty(self): def dummy_discriminator( transition: types.Transition) -> networks_lib.Logits: return transition.observation + jnp.square(transition.action) zero_transition = types.Transition(0., 0., 0., 0., 0.) zero_transition = tree.map_structure(lambda x: jnp.expand_dims(x, axis=0), zero_transition) self.assertEqual( losses._compute_gradient_penalty(zero_transition, dummy_discriminator, 0.), 1**2 + 0**2) one_transition = types.Transition(1., 1., 0., 0., 0.) one_transition = tree.map_structure(lambda x: jnp.expand_dims(x, axis=0), one_transition) self.assertEqual( losses._compute_gradient_penalty(one_transition, dummy_discriminator, 0.), 1**2 + 2**2) def test_pugail(self): def dummy_discriminator( state: losses.State, transition: types.Transition) -> losses.DiscriminatorOutput: return transition.observation, state zero_transition = types.Transition(.1, 0., 0., 0., 0.) zero_transition = tree.map_structure(lambda x: jnp.expand_dims(x, axis=0), zero_transition) one_transition = types.Transition(1., 0., 0., 0., 0.) one_transition = tree.map_structure(lambda x: jnp.expand_dims(x, axis=0), one_transition) prior = .7 loss_fn = losses.pugail_loss( positive_class_prior=prior, entropy_coefficient=0.) loss, _ = loss_fn(dummy_discriminator, {}, one_transition, zero_transition, ()) d_one = jax.nn.sigmoid(dummy_discriminator({}, one_transition)[0]) d_zero = jax.nn.sigmoid(dummy_discriminator({}, zero_transition)[0]) expected_loss = -prior * jnp.log( d_one) + -jnp.log(1. - d_zero) - prior * -jnp.log(1 - d_one) self.assertAlmostEqual(loss, expected_loss, places=6) if __name__ == '__main__': absltest.main()
acme-master
acme/agents/jax/ail/losses_test.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Tests for the builder generator.""" from acme import types from acme.agents.jax.ail import builder import numpy as np import reverb from absl.testing import absltest _REWARD = np.zeros((3,)) class BuilderTest(absltest.TestCase): def test_weighted_generator(self): data0 = types.Transition(np.array([[1], [2], [3]]), (), _REWARD, (), ()) it0 = iter([data0]) data1 = types.Transition(np.array([[4], [5], [6]]), (), _REWARD, (), ()) data2 = types.Transition(np.array([[7], [8], [9]]), (), _REWARD, (), ()) it1 = iter([ reverb.ReplaySample( info=reverb.SampleInfo( *[() for _ in reverb.SampleInfo.tf_dtypes()]), data=data1), reverb.ReplaySample( info=reverb.SampleInfo( *[() for _ in reverb.SampleInfo.tf_dtypes()]), data=data2) ]) weighted_it = builder._generate_samples_with_demonstrations( it0, it1, policy_to_expert_data_ratio=2, batch_size=3) np.testing.assert_array_equal( next(weighted_it).data.observation, np.array([[1], [4], [5]])) np.testing.assert_array_equal( next(weighted_it).data.observation, np.array([[7], [8], [2]])) self.assertRaises(StopIteration, lambda: next(weighted_it)) if __name__ == '__main__': absltest.main()
acme-master
acme/agents/jax/ail/builder_test.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """TD3 config.""" import dataclasses from typing import Optional, Union from acme.adders import reverb as adders_reverb import optax @dataclasses.dataclass class TD3Config: """Configuration options for TD3.""" # Loss options batch_size: int = 256 policy_learning_rate: Union[optax.Schedule, float] = 3e-4 critic_learning_rate: Union[optax.Schedule, float] = 3e-4 # Policy gradient clipping is not part of the original TD3 implementation, # used e.g. in DAC https://arxiv.org/pdf/1809.02925.pdf policy_gradient_clipping: Optional[float] = None discount: float = 0.99 n_step: int = 1 # TD3 specific options (https://arxiv.org/pdf/1802.09477.pdf) sigma: float = 0.1 delay: int = 2 target_sigma: float = 0.2 noise_clip: float = 0.5 tau: float = 0.005 # Replay options min_replay_size: int = 1000 max_replay_size: int = 1000000 replay_table_name: str = adders_reverb.DEFAULT_PRIORITY_TABLE prefetch_size: int = 4 samples_per_insert: float = 256 # Rate to be used for the SampleToInsertRatio rate limiter tolerance. # See a formula in make_replay_tables for more details. samples_per_insert_tolerance_rate: float = 0.1 # How many gradient updates to perform per step. num_sgd_steps_per_step: int = 1 # Offline RL options # if bc_alpha: if given, will add a bc regularization term to the policy loss, # (https://arxiv.org/pdf/2106.06860.pdf), useful for offline training. bc_alpha: Optional[float] = None
acme-master
acme/agents/jax/td3/config.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """TD3 agent.""" from acme.agents.jax.td3.builder import TD3Builder from acme.agents.jax.td3.config import TD3Config from acme.agents.jax.td3.learning import TD3Learner from acme.agents.jax.td3.networks import get_default_behavior_policy from acme.agents.jax.td3.networks import make_networks from acme.agents.jax.td3.networks import TD3Networks
acme-master
acme/agents/jax/td3/__init__.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """TD3 Builder.""" from typing import Iterator, List, Optional from acme import adders from acme import core from acme import specs from acme.adders import reverb as adders_reverb from acme.agents.jax import actor_core as actor_core_lib from acme.agents.jax import actors from acme.agents.jax import builders from acme.agents.jax.td3 import config as td3_config from acme.agents.jax.td3 import learning from acme.agents.jax.td3 import networks as td3_networks from acme.datasets import reverb as datasets from acme.jax import networks as networks_lib from acme.jax import utils from acme.jax import variable_utils from acme.utils import counting from acme.utils import loggers import jax import optax import reverb from reverb import rate_limiters class TD3Builder(builders.ActorLearnerBuilder[td3_networks.TD3Networks, actor_core_lib.FeedForwardPolicy, reverb.ReplaySample]): """TD3 Builder.""" def __init__( self, config: td3_config.TD3Config, ): """Creates a TD3 learner, a behavior policy and an eval actor. Args: config: a config with TD3 hps """ self._config = config def make_learner( self, random_key: networks_lib.PRNGKey, networks: td3_networks.TD3Networks, dataset: Iterator[reverb.ReplaySample], logger_fn: loggers.LoggerFactory, environment_spec: specs.EnvironmentSpec, replay_client: Optional[reverb.Client] = None, counter: Optional[counting.Counter] = None, ) -> core.Learner: del environment_spec, replay_client critic_optimizer = optax.adam(self._config.critic_learning_rate) twin_critic_optimizer = optax.adam(self._config.critic_learning_rate) policy_optimizer = optax.adam(self._config.policy_learning_rate) if self._config.policy_gradient_clipping is not None: policy_optimizer = optax.chain( optax.clip_by_global_norm(self._config.policy_gradient_clipping), policy_optimizer) return learning.TD3Learner( networks=networks, random_key=random_key, discount=self._config.discount, target_sigma=self._config.target_sigma, noise_clip=self._config.noise_clip, policy_optimizer=policy_optimizer, critic_optimizer=critic_optimizer, twin_critic_optimizer=twin_critic_optimizer, num_sgd_steps_per_step=self._config.num_sgd_steps_per_step, bc_alpha=self._config.bc_alpha, iterator=dataset, logger=logger_fn('learner'), counter=counter) def make_actor( self, random_key: networks_lib.PRNGKey, policy: actor_core_lib.FeedForwardPolicy, environment_spec: specs.EnvironmentSpec, variable_source: Optional[core.VariableSource] = None, adder: Optional[adders.Adder] = None, ) -> core.Actor: del environment_spec assert variable_source is not None actor_core = actor_core_lib.batched_feed_forward_to_actor_core(policy) # Inference happens on CPU, so it's better to move variables there too. variable_client = variable_utils.VariableClient(variable_source, 'policy', device='cpu') return actors.GenericActor( actor_core, random_key, variable_client, adder, backend='cpu') def make_replay_tables( self, environment_spec: specs.EnvironmentSpec, policy: actor_core_lib.FeedForwardPolicy, ) -> List[reverb.Table]: """Creates reverb tables for the algorithm.""" del policy samples_per_insert_tolerance = ( self._config.samples_per_insert_tolerance_rate * self._config.samples_per_insert) error_buffer = self._config.min_replay_size * samples_per_insert_tolerance limiter = rate_limiters.SampleToInsertRatio( min_size_to_sample=self._config.min_replay_size, samples_per_insert=self._config.samples_per_insert, error_buffer=error_buffer) return [reverb.Table( name=self._config.replay_table_name, sampler=reverb.selectors.Uniform(), remover=reverb.selectors.Fifo(), max_size=self._config.max_replay_size, rate_limiter=limiter, signature=adders_reverb.NStepTransitionAdder.signature( environment_spec))] def make_dataset_iterator( self, replay_client: reverb.Client) -> Iterator[reverb.ReplaySample]: """Creates a dataset iterator to use for learning.""" dataset = datasets.make_reverb_dataset( table=self._config.replay_table_name, server_address=replay_client.server_address, batch_size=( self._config.batch_size * self._config.num_sgd_steps_per_step), prefetch_size=self._config.prefetch_size, transition_adder=True) return utils.device_put(dataset.as_numpy_iterator(), jax.devices()[0]) def make_adder( self, replay_client: reverb.Client, environment_spec: Optional[specs.EnvironmentSpec], policy: Optional[actor_core_lib.FeedForwardPolicy] ) -> Optional[adders.Adder]: """Creates an adder which handles observations.""" del environment_spec, policy return adders_reverb.NStepTransitionAdder( priority_fns={self._config.replay_table_name: None}, client=replay_client, n_step=self._config.n_step, discount=self._config.discount) def make_policy(self, networks: td3_networks.TD3Networks, environment_spec: specs.EnvironmentSpec, evaluation: bool = False) -> actor_core_lib.FeedForwardPolicy: """Creates a policy.""" sigma = 0 if evaluation else self._config.sigma return td3_networks.get_default_behavior_policy( networks=networks, action_specs=environment_spec.actions, sigma=sigma)
acme-master
acme/agents/jax/td3/builder.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """TD3 networks definition.""" import dataclasses from typing import Callable, Sequence from acme import specs from acme import types from acme.agents.jax import actor_core as actor_core_lib from acme.jax import networks as networks_lib from acme.jax import utils import haiku as hk import jax import jax.numpy as jnp import numpy as np @dataclasses.dataclass class TD3Networks: """Network and pure functions for the TD3 agent.""" policy_network: networks_lib.FeedForwardNetwork critic_network: networks_lib.FeedForwardNetwork twin_critic_network: networks_lib.FeedForwardNetwork add_policy_noise: Callable[[types.NestedArray, networks_lib.PRNGKey, float, float], types.NestedArray] def get_default_behavior_policy( networks: TD3Networks, action_specs: specs.BoundedArray, sigma: float) -> actor_core_lib.FeedForwardPolicy: """Selects action according to the policy.""" def behavior_policy(params: networks_lib.Params, key: networks_lib.PRNGKey, observation: types.NestedArray): action = networks.policy_network.apply(params, observation) noise = jax.random.normal(key, shape=action.shape) * sigma noisy_action = jnp.clip(action + noise, action_specs.minimum, action_specs.maximum) return noisy_action return behavior_policy def make_networks( spec: specs.EnvironmentSpec, hidden_layer_sizes: Sequence[int] = (256, 256)) -> TD3Networks: """Creates networks used by the agent. The networks used are based on LayerNormMLP, which is different than the MLP with relu activation described in TD3 (which empirically performs worse). Args: spec: Environment specs hidden_layer_sizes: list of sizes of hidden layers in actor/critic networks Returns: network: TD3Networks """ action_specs = spec.actions num_dimensions = np.prod(action_specs.shape, dtype=int) def add_policy_noise(action: types.NestedArray, key: networks_lib.PRNGKey, target_sigma: float, noise_clip: float) -> types.NestedArray: """Adds action noise to bootstrapped Q-value estimate in critic loss.""" noise = jax.random.normal(key=key, shape=action_specs.shape) * target_sigma noise = jnp.clip(noise, -noise_clip, noise_clip) return jnp.clip(action + noise, action_specs.minimum, action_specs.maximum) def _actor_fn(obs: types.NestedArray) -> types.NestedArray: network = hk.Sequential([ networks_lib.LayerNormMLP(hidden_layer_sizes, activate_final=True), networks_lib.NearZeroInitializedLinear(num_dimensions), networks_lib.TanhToSpec(spec.actions), ]) return network(obs) def _critic_fn(obs: types.NestedArray, action: types.NestedArray) -> types.NestedArray: network1 = hk.Sequential([ networks_lib.LayerNormMLP(list(hidden_layer_sizes) + [1]), ]) input_ = jnp.concatenate([obs, action], axis=-1) value = network1(input_) return jnp.squeeze(value) policy = hk.without_apply_rng(hk.transform(_actor_fn)) critic = hk.without_apply_rng(hk.transform(_critic_fn)) # Create dummy observations and actions to create network parameters. dummy_action = utils.zeros_like(spec.actions) dummy_obs = utils.zeros_like(spec.observations) dummy_action = utils.add_batch_dim(dummy_action) dummy_obs = utils.add_batch_dim(dummy_obs) network = TD3Networks( policy_network=networks_lib.FeedForwardNetwork( lambda key: policy.init(key, dummy_obs), policy.apply), critic_network=networks_lib.FeedForwardNetwork( lambda key: critic.init(key, dummy_obs, dummy_action), critic.apply), twin_critic_network=networks_lib.FeedForwardNetwork( lambda key: critic.init(key, dummy_obs, dummy_action), critic.apply), add_policy_noise=add_policy_noise) return network
acme-master
acme/agents/jax/td3/networks.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """TD3 learner implementation.""" import time from typing import Dict, Iterator, List, NamedTuple, Optional, Tuple import acme from acme import types from acme.agents.jax.td3 import networks as td3_networks from acme.jax import networks as networks_lib from acme.jax import utils from acme.utils import counting from acme.utils import loggers import jax import jax.numpy as jnp import optax import reverb import rlax class TrainingState(NamedTuple): """Contains training state for the learner.""" policy_params: networks_lib.Params target_policy_params: networks_lib.Params critic_params: networks_lib.Params target_critic_params: networks_lib.Params twin_critic_params: networks_lib.Params target_twin_critic_params: networks_lib.Params policy_opt_state: optax.OptState critic_opt_state: optax.OptState twin_critic_opt_state: optax.OptState steps: int random_key: networks_lib.PRNGKey class TD3Learner(acme.Learner): """TD3 learner.""" _state: TrainingState def __init__(self, networks: td3_networks.TD3Networks, random_key: networks_lib.PRNGKey, discount: float, iterator: Iterator[reverb.ReplaySample], policy_optimizer: optax.GradientTransformation, critic_optimizer: optax.GradientTransformation, twin_critic_optimizer: optax.GradientTransformation, delay: int = 2, target_sigma: float = 0.2, noise_clip: float = 0.5, tau: float = 0.005, use_sarsa_target: bool = False, bc_alpha: Optional[float] = None, counter: Optional[counting.Counter] = None, logger: Optional[loggers.Logger] = None, num_sgd_steps_per_step: int = 1): """Initializes the TD3 learner. Args: networks: TD3 networks. random_key: a key for random number generation. discount: discount to use for TD updates iterator: an iterator over training data. policy_optimizer: the policy optimizer. critic_optimizer: the Q-function optimizer. twin_critic_optimizer: the twin Q-function optimizer. delay: ratio of policy updates for critic updates (see TD3), delay=2 means 2 updates of the critic for 1 policy update. target_sigma: std of zero mean Gaussian added to the action of the next_state, for critic evaluation (reducing overestimation bias). noise_clip: hard constraint on target noise. tau: target parameters smoothing coefficient. use_sarsa_target: compute on-policy target using iterator's actions rather than sampled actions. Useful for 1-step offline RL (https://arxiv.org/pdf/2106.08909.pdf). When set to `True`, `target_policy_params` are unused. This is only working when the learner is used as an offline algorithm. I.e. TD3Builder does not support adding the SARSA target to the replay buffer. bc_alpha: bc_alpha: Implements TD3+BC. See comments in TD3Config.bc_alpha for details. counter: counter object used to keep track of steps. logger: logger object to be used by learner. num_sgd_steps_per_step: number of sgd steps to perform per learner 'step'. """ def policy_loss( policy_params: networks_lib.Params, critic_params: networks_lib.Params, transition: types.NestedArray, ) -> jnp.ndarray: # Computes the discrete policy gradient loss. action = networks.policy_network.apply( policy_params, transition.observation) grad_critic = jax.vmap( jax.grad(networks.critic_network.apply, argnums=2), in_axes=(None, 0, 0)) dq_da = grad_critic(critic_params, transition.observation, action) batch_dpg_learning = jax.vmap(rlax.dpg_loss, in_axes=(0, 0)) loss = jnp.mean(batch_dpg_learning(action, dq_da)) if bc_alpha is not None: # BC regularization for offline RL q_sa = networks.critic_network.apply(critic_params, transition.observation, action) bc_factor = jax.lax.stop_gradient(bc_alpha / jnp.mean(jnp.abs(q_sa))) loss += jnp.mean(jnp.square(action - transition.action)) / bc_factor return loss def critic_loss( critic_params: networks_lib.Params, state: TrainingState, transition: types.Transition, random_key: jnp.ndarray, ): # Computes the critic loss. q_tm1 = networks.critic_network.apply( critic_params, transition.observation, transition.action) if use_sarsa_target: # TODO(b/222674779): use N-steps Trajectories to get the next actions. assert 'next_action' in transition.extras, ( 'next actions should be given as extras for one step RL.') action = transition.extras['next_action'] else: action = networks.policy_network.apply(state.target_policy_params, transition.next_observation) action = networks.add_policy_noise(action, random_key, target_sigma, noise_clip) q_t = networks.critic_network.apply( state.target_critic_params, transition.next_observation, action) twin_q_t = networks.twin_critic_network.apply( state.target_twin_critic_params, transition.next_observation, action) q_t = jnp.minimum(q_t, twin_q_t) target_q_tm1 = transition.reward + discount * transition.discount * q_t td_error = jax.lax.stop_gradient(target_q_tm1) - q_tm1 return jnp.mean(jnp.square(td_error)) def update_step( state: TrainingState, transitions: types.Transition, ) -> Tuple[TrainingState, Dict[str, jnp.ndarray]]: random_key, key_critic, key_twin = jax.random.split(state.random_key, 3) # Updates on the critic: compute the gradients, and update using # Polyak averaging. critic_loss_and_grad = jax.value_and_grad(critic_loss) critic_loss_value, critic_gradients = critic_loss_and_grad( state.critic_params, state, transitions, key_critic) critic_updates, critic_opt_state = critic_optimizer.update( critic_gradients, state.critic_opt_state) critic_params = optax.apply_updates(state.critic_params, critic_updates) # In the original authors' implementation the critic target update is # delayed similarly to the policy update which we found empirically to # perform slightly worse. target_critic_params = optax.incremental_update( new_tensors=critic_params, old_tensors=state.target_critic_params, step_size=tau) # Updates on the twin critic: compute the gradients, and update using # Polyak averaging. twin_critic_loss_value, twin_critic_gradients = critic_loss_and_grad( state.twin_critic_params, state, transitions, key_twin) twin_critic_updates, twin_critic_opt_state = twin_critic_optimizer.update( twin_critic_gradients, state.twin_critic_opt_state) twin_critic_params = optax.apply_updates(state.twin_critic_params, twin_critic_updates) # In the original authors' implementation the twin critic target update is # delayed similarly to the policy update which we found empirically to # perform slightly worse. target_twin_critic_params = optax.incremental_update( new_tensors=twin_critic_params, old_tensors=state.target_twin_critic_params, step_size=tau) # Updates on the policy: compute the gradients, and update using # Polyak averaging (if delay enabled, the update might not be applied). policy_loss_and_grad = jax.value_and_grad(policy_loss) policy_loss_value, policy_gradients = policy_loss_and_grad( state.policy_params, state.critic_params, transitions) def update_policy_step(): policy_updates, policy_opt_state = policy_optimizer.update( policy_gradients, state.policy_opt_state) policy_params = optax.apply_updates(state.policy_params, policy_updates) target_policy_params = optax.incremental_update( new_tensors=policy_params, old_tensors=state.target_policy_params, step_size=tau) return policy_params, target_policy_params, policy_opt_state # The update on the policy is applied every `delay` steps. current_policy_state = (state.policy_params, state.target_policy_params, state.policy_opt_state) policy_params, target_policy_params, policy_opt_state = jax.lax.cond( state.steps % delay == 0, lambda _: update_policy_step(), lambda _: current_policy_state, operand=None) steps = state.steps + 1 new_state = TrainingState( policy_params=policy_params, critic_params=critic_params, twin_critic_params=twin_critic_params, target_policy_params=target_policy_params, target_critic_params=target_critic_params, target_twin_critic_params=target_twin_critic_params, policy_opt_state=policy_opt_state, critic_opt_state=critic_opt_state, twin_critic_opt_state=twin_critic_opt_state, steps=steps, random_key=random_key, ) metrics = { 'policy_loss': policy_loss_value, 'critic_loss': critic_loss_value, 'twin_critic_loss': twin_critic_loss_value, } return new_state, metrics # General learner book-keeping and loggers. self._counter = counter or counting.Counter() self._logger = logger or loggers.make_default_logger( 'learner', asynchronous=True, serialize_fn=utils.fetch_devicearray, steps_key=self._counter.get_steps_key()) # Create prefetching dataset iterator. self._iterator = iterator # Faster sgd step update_step = utils.process_multiple_batches(update_step, num_sgd_steps_per_step) # Use the JIT compiler. self._update_step = jax.jit(update_step) (key_init_policy, key_init_twin, key_init_target, key_state ) = jax.random.split(random_key, 4) # Create the network parameters and copy into the target network parameters. initial_policy_params = networks.policy_network.init(key_init_policy) initial_critic_params = networks.critic_network.init(key_init_twin) initial_twin_critic_params = networks.twin_critic_network.init( key_init_target) initial_target_policy_params = initial_policy_params initial_target_critic_params = initial_critic_params initial_target_twin_critic_params = initial_twin_critic_params # Initialize optimizers. initial_policy_opt_state = policy_optimizer.init(initial_policy_params) initial_critic_opt_state = critic_optimizer.init(initial_critic_params) initial_twin_critic_opt_state = twin_critic_optimizer.init( initial_twin_critic_params) # Create initial state. self._state = TrainingState( policy_params=initial_policy_params, target_policy_params=initial_target_policy_params, critic_params=initial_critic_params, twin_critic_params=initial_twin_critic_params, target_critic_params=initial_target_critic_params, target_twin_critic_params=initial_target_twin_critic_params, policy_opt_state=initial_policy_opt_state, critic_opt_state=initial_critic_opt_state, twin_critic_opt_state=initial_twin_critic_opt_state, steps=0, random_key=key_state ) # Do not record timestamps until after the first learning step is done. # This is to avoid including the time it takes for actors to come online and # fill the replay buffer. self._timestamp = None def step(self): # Get data from replay (dropping extras if any). Note there is no # extra data here because we do not insert any into Reverb. sample = next(self._iterator) transitions = types.Transition(*sample.data) self._state, metrics = self._update_step(self._state, transitions) # Compute elapsed time. timestamp = time.time() elapsed_time = timestamp - self._timestamp if self._timestamp else 0 self._timestamp = timestamp # Increment counts and record the current time counts = self._counter.increment(steps=1, walltime=elapsed_time) # Attempts to write the logs. self._logger.write({**metrics, **counts}) def get_variables(self, names: List[str]) -> List[networks_lib.Params]: variables = { 'policy': self._state.policy_params, 'critic': self._state.critic_params, 'twin_critic': self._state.twin_critic_params, } return [variables[name] for name in names] def save(self) -> TrainingState: return self._state def restore(self, state: TrainingState): self._state = state
acme-master
acme/agents/jax/td3/learning.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """BVE config.""" import dataclasses from typing import Callable, Sequence, Union import numpy as np @dataclasses.dataclass class BVEConfig: """Configuration options for BVE agent. Attributes: epsilon: for use by epsilon-greedy policies. If multiple, the epsilons are alternated randomly per-episode. seed: Random seed. learning_rate: Learning rate for Adam optimizer. Could be a number or a function defining a schedule. adam_eps: Epsilon for Adam optimizer. discount: Discount rate applied to value per timestep. target_update_period: Update target network every period. max_gradient_norm: For gradient clipping. max_abs_reward: Maximum absolute reward. huber_loss_parameter: The delta parameter of the huber loss. batch_size: Number of transitions per batch. prefetch_size: Prefetch size for reverb replay performance. num_sgd_steps_per_step: How many gradient updates to perform per learner step. """ epsilon: Union[float, Sequence[float]] = 0.05 # TODO(b/191706065): update all clients and remove this field. seed: int = 1 # Learning rule learning_rate: Union[float, Callable[[int], float]] = 3e-4 adam_eps: float = 1e-8 # Eps for Adam optimizer. discount: float = 0.99 # Discount rate applied to value per timestep. target_update_period: int = 2500 # Update target network every period. max_gradient_norm: float = np.inf # For gradient clipping. max_abs_reward: float = 1. # Maximum absolute value to clip the rewards. huber_loss_parameter: float = 1. # Huber loss delta parameter. batch_size: int = 256 # Minibatch size. prefetch_size = 500 # The amount of data to prefetch into the memory. num_sgd_steps_per_step: int = 1
acme-master
acme/agents/jax/bve/config.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Implementation of a behavior value estimation (BVE).""" from acme.agents.jax.bve.builder import BVEBuilder from acme.agents.jax.bve.config import BVEConfig from acme.agents.jax.bve.losses import BVELoss from acme.agents.jax.bve.networks import BVENetworks
acme-master
acme/agents/jax/bve/__init__.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """BVE Builder.""" from typing import Iterator, Optional from acme import core from acme import specs from acme.agents.jax import actor_core as actor_core_lib from acme.agents.jax import actors from acme.agents.jax import builders from acme.agents.jax.bve import losses from acme.agents.jax.bve import networks as bve_networks from acme.agents.jax.dqn import learning_lib from acme.jax import networks as networks_lib from acme.jax import types as jax_types from acme.jax import utils from acme.jax import variable_utils from acme.utils import counting from acme.utils import loggers import haiku as hk import optax class BVEBuilder(builders.OfflineBuilder[bve_networks.BVENetworks, actor_core_lib.ActorCore, utils.PrefetchingSplit]): """BVE Builder.""" def __init__(self, config): """Build a BVE agent. Args: config: The config of the BVE agent. """ self._config = config def make_learner(self, random_key: jax_types.PRNGKey, networks: bve_networks.BVENetworks, dataset: Iterator[utils.PrefetchingSplit], logger_fn: loggers.LoggerFactory, environment_spec: specs.EnvironmentSpec, counter: Optional[counting.Counter] = None) -> core.Learner: del environment_spec loss_fn = losses.BVELoss( discount=self._config.discount, max_abs_reward=self._config.max_abs_reward, huber_loss_parameter=self._config.huber_loss_parameter, ) return learning_lib.SGDLearner( network=networks.policy_network, random_key=random_key, optimizer=optax.adam( self._config.learning_rate, eps=self._config.adam_eps), target_update_period=self._config.target_update_period, data_iterator=dataset, loss_fn=loss_fn, counter=counter, num_sgd_steps_per_step=self._config.num_sgd_steps_per_step, logger=logger_fn('learner')) def make_actor( self, random_key: jax_types.PRNGKey, policy: actor_core_lib.ActorCore, environment_spec: specs.EnvironmentSpec, variable_source: Optional[core.VariableSource] = None) -> core.Actor: """Create the actor for the BVE to perform online evals. Args: random_key: prng key. policy: The DQN policy. environment_spec: The environment spec. variable_source: The source of where the variables are coming from. Returns: Return the actor for the evaluations. """ del environment_spec variable_client = variable_utils.VariableClient( variable_source, 'policy', device='cpu') return actors.GenericActor(policy, random_key, variable_client) def make_policy( self, networks: bve_networks.BVENetworks, environment_spec: specs.EnvironmentSpec, evaluation: Optional[bool] = False) -> actor_core_lib.ActorCore: """Creates a policy.""" del environment_spec, evaluation def behavior_policy( params: hk.Params, key: jax_types.PRNGKey, observation: networks_lib.Observation) -> networks_lib.Action: network_output = networks.policy_network.apply( params, observation, is_training=False) return networks.sample_fn(network_output, key) return actor_core_lib.batched_feed_forward_to_actor_core(behavior_policy)
acme-master
acme/agents/jax/bve/builder.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Network definitions for BVE.""" import dataclasses from typing import Optional from acme.jax import networks as networks_lib @dataclasses.dataclass class BVENetworks: """The network and pure functions for the BVE agent. Attributes: policy_network: The policy network. sample_fn: A pure function. Samples an action based on the network output. log_prob: A pure function. Computes log-probability for an action. """ policy_network: networks_lib.TypedFeedForwardNetwork sample_fn: networks_lib.SampleFn log_prob: Optional[networks_lib.LogProbFn] = None
acme-master
acme/agents/jax/bve/networks.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Behavior Value Estimation loss.""" import dataclasses from typing import Tuple from acme import types from acme.agents.jax import dqn from acme.jax import networks as networks_lib import jax import jax.numpy as jnp import reverb import rlax @dataclasses.dataclass class BVELoss(dqn.LossFn): """This loss implements TD-loss to estimate behavior value. This loss function uses the next action to learn with the SARSA tuples. It is intended to be used with dqn.SGDLearner. The method was proposed in "Regularized Behavior Value Estimation" by Gulcehre et al to overcome the extrapolation error in offline RL setting: https://arxiv.org/abs/2103.09575 """ discount: float = 0.99 max_abs_reward: float = 1. huber_loss_parameter: float = 1. def __call__( self, network: networks_lib.TypedFeedForwardNetwork, params: networks_lib.Params, target_params: networks_lib.Params, batch: reverb.ReplaySample, key: networks_lib.PRNGKey, ) -> Tuple[jax.Array, dqn.LossExtra]: """Calculate a loss on a single batch of data.""" transitions: types.Transition = batch.data # Forward pass. key1, key2 = jax.random.split(key) q_tm1 = network.apply( params, transitions.observation, is_training=True, key=key1) q_t_value = network.apply( target_params, transitions.next_observation, is_training=True, key=key2) # Cast and clip rewards. d_t = (transitions.discount * self.discount).astype(jnp.float32) r_t = jnp.clip(transitions.reward, -self.max_abs_reward, self.max_abs_reward).astype(jnp.float32) # Compute double Q-learning n-step TD-error. batch_error = jax.vmap(rlax.sarsa) next_action = transitions.extras['next_action'] td_error = batch_error(q_tm1, transitions.action, r_t, d_t, q_t_value, next_action) batch_loss = rlax.huber_loss(td_error, self.huber_loss_parameter) # Average: loss = jnp.mean(batch_loss) # [] metrics = {'td_error': td_error, 'batch_loss': batch_loss} return loss, dqn.LossExtra( metrics=metrics, reverb_priorities=jnp.abs(td_error).astype(jnp.float64))
acme-master
acme/agents/jax/bve/losses.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License.
acme-master
acme/agents/tf/__init__.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Generic actor implementation, using TensorFlow and Sonnet.""" from typing import Optional, Tuple from acme import adders from acme import core from acme import types # Internal imports. from acme.tf import utils as tf2_utils from acme.tf import variable_utils as tf2_variable_utils import dm_env import sonnet as snt import tensorflow as tf import tensorflow_probability as tfp tfd = tfp.distributions class FeedForwardActor(core.Actor): """A feed-forward actor. An actor based on a feed-forward policy which takes non-batched observations and outputs non-batched actions. It also allows adding experiences to replay and updating the weights from the policy on the learner. """ def __init__( self, policy_network: snt.Module, adder: Optional[adders.Adder] = None, variable_client: Optional[tf2_variable_utils.VariableClient] = None, ): """Initializes the actor. Args: policy_network: the policy to run. adder: the adder object to which allows to add experiences to a dataset/replay buffer. variable_client: object which allows to copy weights from the learner copy of the policy to the actor copy (in case they are separate). """ # Store these for later use. self._adder = adder self._variable_client = variable_client self._policy_network = policy_network @tf.function def _policy(self, observation: types.NestedTensor) -> types.NestedTensor: # Add a dummy batch dimension and as a side effect convert numpy to TF. batched_observation = tf2_utils.add_batch_dim(observation) # Compute the policy, conditioned on the observation. policy = self._policy_network(batched_observation) # Sample from the policy if it is stochastic. action = policy.sample() if isinstance(policy, tfd.Distribution) else policy return action def select_action(self, observation: types.NestedArray) -> types.NestedArray: # Pass the observation through the policy network. action = self._policy(observation) # Return a numpy array with squeezed out batch dimension. return tf2_utils.to_numpy_squeeze(action) def observe_first(self, timestep: dm_env.TimeStep): if self._adder: self._adder.add_first(timestep) def observe(self, action: types.NestedArray, next_timestep: dm_env.TimeStep): if self._adder: self._adder.add(action, next_timestep) def update(self, wait: bool = False): if self._variable_client: self._variable_client.update(wait) class RecurrentActor(core.Actor): """A recurrent actor. An actor based on a recurrent policy which takes non-batched observations and outputs non-batched actions, and keeps track of the recurrent state inside. It also allows adding experiences to replay and updating the weights from the policy on the learner. """ def __init__( self, policy_network: snt.RNNCore, adder: Optional[adders.Adder] = None, variable_client: Optional[tf2_variable_utils.VariableClient] = None, store_recurrent_state: bool = True, ): """Initializes the actor. Args: policy_network: the (recurrent) policy to run. adder: the adder object to which allows to add experiences to a dataset/replay buffer. variable_client: object which allows to copy weights from the learner copy of the policy to the actor copy (in case they are separate). store_recurrent_state: Whether to pass the recurrent state to the adder. """ # Store these for later use. self._adder = adder self._variable_client = variable_client self._network = policy_network self._state = None self._prev_state = None self._store_recurrent_state = store_recurrent_state @tf.function def _policy( self, observation: types.NestedTensor, state: types.NestedTensor, ) -> Tuple[types.NestedTensor, types.NestedTensor]: # Add a dummy batch dimension and as a side effect convert numpy to TF. batched_observation = tf2_utils.add_batch_dim(observation) # Compute the policy, conditioned on the observation. policy, new_state = self._network(batched_observation, state) # Sample from the policy if it is stochastic. action = policy.sample() if isinstance(policy, tfd.Distribution) else policy return action, new_state def select_action(self, observation: types.NestedArray) -> types.NestedArray: # Initialize the RNN state if necessary. if self._state is None: self._state = self._network.initial_state(1) # Step the recurrent policy forward given the current observation and state. policy_output, new_state = self._policy(observation, self._state) # Bookkeeping of recurrent states for the observe method. self._prev_state = self._state self._state = new_state # Return a numpy array with squeezed out batch dimension. return tf2_utils.to_numpy_squeeze(policy_output) def observe_first(self, timestep: dm_env.TimeStep): if self._adder: self._adder.add_first(timestep) # Set the state to None so that we re-initialize at the next policy call. self._state = None def observe(self, action: types.NestedArray, next_timestep: dm_env.TimeStep): if not self._adder: return if not self._store_recurrent_state: self._adder.add(action, next_timestep) return numpy_state = tf2_utils.to_numpy_squeeze(self._prev_state) self._adder.add(action, next_timestep, extras=(numpy_state,)) def update(self, wait: bool = False): if self._variable_client: self._variable_client.update(wait) # Internal class 1. # Internal class 2.
acme-master
acme/agents/tf/actors.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Tests for actors_tf2.""" from acme import environment_loop from acme import specs from acme.agents.tf import actors from acme.testing import fakes import dm_env import numpy as np import sonnet as snt import tensorflow as tf from absl.testing import absltest def _make_fake_env() -> dm_env.Environment: env_spec = specs.EnvironmentSpec( observations=specs.Array(shape=(10, 5), dtype=np.float32), actions=specs.DiscreteArray(num_values=3), rewards=specs.Array(shape=(), dtype=np.float32), discounts=specs.BoundedArray( shape=(), dtype=np.float32, minimum=0., maximum=1.), ) return fakes.Environment(env_spec, episode_length=10) class ActorTest(absltest.TestCase): def test_feedforward(self): environment = _make_fake_env() env_spec = specs.make_environment_spec(environment) network = snt.Sequential([ snt.Flatten(), snt.Linear(env_spec.actions.num_values), lambda x: tf.argmax(x, axis=-1, output_type=env_spec.actions.dtype), ]) actor = actors.FeedForwardActor(network) loop = environment_loop.EnvironmentLoop(environment, actor) loop.run(20) def test_recurrent(self): environment = _make_fake_env() env_spec = specs.make_environment_spec(environment) network = snt.DeepRNN([ snt.Flatten(), snt.Linear(env_spec.actions.num_values), lambda x: tf.argmax(x, axis=-1, output_type=env_spec.actions.dtype), ]) actor = actors.RecurrentActor(network) loop = environment_loop.EnvironmentLoop(environment, actor) loop.run(20) if __name__ == '__main__': absltest.main()
acme-master
acme/agents/tf/actors_test.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Defines the (MoG) distributional MPO distributed agent class.""" from typing import Callable, Dict, Optional import acme from acme import datasets from acme import specs from acme.adders import reverb as adders from acme.agents.tf import actors from acme.agents.tf.mog_mpo import learning from acme.tf import networks from acme.tf import savers as tf2_savers from acme.tf import variable_utils as tf2_variable_utils from acme.utils import counting from acme.utils import loggers from acme.utils import lp_utils import dm_env import launchpad as lp import reverb import sonnet as snt class DistributedMoGMPO: """Program definition for distributional (MoG) MPO.""" def __init__( self, environment_factory: Callable[[bool], dm_env.Environment], network_factory: Callable[[specs.EnvironmentSpec], Dict[str, snt.Module]], num_actors: int = 1, num_caches: int = 0, environment_spec: Optional[specs.EnvironmentSpec] = None, batch_size: int = 256, prefetch_size: int = 4, min_replay_size: int = 1_000, max_replay_size: int = 1_000_000, samples_per_insert: Optional[float] = 32.0, n_step: int = 5, num_samples: int = 20, policy_evaluation_config: Optional[ learning.PolicyEvaluationConfig] = None, additional_discount: float = 0.99, target_policy_update_period: int = 100, target_critic_update_period: int = 100, policy_loss_factory: Optional[Callable[[], snt.Module]] = None, max_actor_steps: Optional[int] = None, log_every: float = 10.0, ): if environment_spec is None: environment_spec = specs.make_environment_spec(environment_factory(False)) self._environment_factory = environment_factory self._network_factory = network_factory self._policy_loss_factory = policy_loss_factory self._environment_spec = environment_spec self._num_actors = num_actors self._num_caches = num_caches self._batch_size = batch_size self._prefetch_size = prefetch_size self._min_replay_size = min_replay_size self._max_replay_size = max_replay_size self._samples_per_insert = samples_per_insert self._n_step = n_step self._additional_discount = additional_discount self._num_samples = num_samples self._policy_evaluation_config = policy_evaluation_config self._target_policy_update_period = target_policy_update_period self._target_critic_update_period = target_critic_update_period self._max_actor_steps = max_actor_steps self._log_every = log_every def replay(self): """The replay storage.""" if self._samples_per_insert is not None: # Create enough of an error buffer to give a 10% tolerance in rate. samples_per_insert_tolerance = 0.1 * self._samples_per_insert error_buffer = self._min_replay_size * samples_per_insert_tolerance limiter = reverb.rate_limiters.SampleToInsertRatio( min_size_to_sample=self._min_replay_size, samples_per_insert=self._samples_per_insert, error_buffer=error_buffer) else: limiter = reverb.rate_limiters.MinSize(self._min_replay_size) replay_table = reverb.Table( name=adders.DEFAULT_PRIORITY_TABLE, sampler=reverb.selectors.Uniform(), remover=reverb.selectors.Fifo(), max_size=self._max_replay_size, rate_limiter=limiter, signature=adders.NStepTransitionAdder.signature( self._environment_spec)) return [replay_table] def counter(self): return tf2_savers.CheckpointingRunner( counting.Counter(), time_delta_minutes=1, subdirectory='counter') def coordinator(self, counter: counting.Counter, max_actor_steps: int): return lp_utils.StepsLimiter(counter, max_actor_steps) def learner( self, replay: reverb.Client, counter: counting.Counter, ): """The Learning part of the agent.""" # Create online and target networks. online_networks = self._network_factory(self._environment_spec) target_networks = self._network_factory(self._environment_spec) # The dataset object to learn from. dataset = datasets.make_reverb_dataset( server_address=replay.server_address, batch_size=self._batch_size, prefetch_size=self._prefetch_size, ) counter = counting.Counter(counter, 'learner') logger = loggers.make_default_logger('learner', time_delta=self._log_every) # Create policy loss module if a factory is passed. if self._policy_loss_factory: policy_loss_module = self._policy_loss_factory() else: policy_loss_module = None # Return the learning agent. return learning.MoGMPOLearner( policy_network=online_networks['policy'], critic_network=online_networks['critic'], observation_network=online_networks['observation'], target_policy_network=target_networks['policy'], target_critic_network=target_networks['critic'], target_observation_network=target_networks['observation'], discount=self._additional_discount, num_samples=self._num_samples, policy_evaluation_config=self._policy_evaluation_config, target_policy_update_period=self._target_policy_update_period, target_critic_update_period=self._target_critic_update_period, policy_loss_module=policy_loss_module, dataset=dataset, counter=counter, logger=logger) def actor( self, replay: reverb.Client, variable_source: acme.VariableSource, counter: counting.Counter, actor_id: int, ) -> acme.EnvironmentLoop: """The actor process.""" # Create environment and target networks to act with. environment = self._environment_factory(False) agent_networks = self._network_factory(self._environment_spec) # Create a stochastic behavior policy. behavior_network = snt.Sequential([ agent_networks['observation'], agent_networks['policy'], networks.StochasticSamplingHead(), ]) # Ensure network variables are created. policy_variables = {'policy': behavior_network.variables} # Create the variable client responsible for keeping the actor up-to-date. variable_client = tf2_variable_utils.VariableClient( variable_source, policy_variables, update_period=1000) # Make sure not to use a random policy after checkpoint restoration by # assigning variables before running the environment loop. variable_client.update_and_wait() # Component to add things into replay. adder = adders.NStepTransitionAdder( client=replay, n_step=self._n_step, discount=self._additional_discount) # Create the agent. actor = actors.FeedForwardActor( policy_network=behavior_network, adder=adder, variable_client=variable_client) # Create logger and counter; actors will not spam bigtable. save_data = actor_id == 0 counter = counting.Counter(counter, 'actor') logger = loggers.make_default_logger( 'actor', save_data=save_data, time_delta=self._log_every) # Create the run loop and return it. return acme.EnvironmentLoop(environment, actor, counter, logger) def evaluator( self, variable_source: acme.VariableSource, counter: counting.Counter, ): """The evaluation process.""" # Create environment and target networks to act with. environment = self._environment_factory(True) agent_networks = self._network_factory(self._environment_spec) # Create a stochastic behavior policy. evaluator_network = snt.Sequential([ agent_networks['observation'], agent_networks['policy'], networks.StochasticMeanHead(), ]) # Create the variable client responsible for keeping the actor up-to-date. variable_client = tf2_variable_utils.VariableClient( variable_source, variables={'policy': evaluator_network.variables}, update_period=1000) # Make sure not to evaluate a random actor by assigning variables before # running the environment loop. variable_client.update_and_wait() # Create the agent. evaluator = actors.FeedForwardActor( policy_network=evaluator_network, variable_client=variable_client) # Create logger and counter. counter = counting.Counter(counter, 'evaluator') logger = loggers.make_default_logger( 'evaluator', time_delta=self._log_every) # Create the run loop and return it. return acme.EnvironmentLoop(environment, evaluator, counter, logger) def build(self, name='dmpo'): """Build the distributed agent topology.""" program = lp.Program(name=name) with program.group('replay'): replay = program.add_node(lp.ReverbNode(self.replay)) with program.group('counter'): counter = program.add_node(lp.CourierNode(self.counter)) if self._max_actor_steps: _ = program.add_node( lp.CourierNode(self.coordinator, counter, self._max_actor_steps)) with program.group('learner'): learner = program.add_node(lp.CourierNode(self.learner, replay, counter)) with program.group('evaluator'): program.add_node(lp.CourierNode(self.evaluator, learner, counter)) if not self._num_caches: # Use our learner as a single variable source. sources = [learner] else: with program.group('cacher'): # Create a set of learner caches. sources = [] for _ in range(self._num_caches): cacher = program.add_node( lp.CacherNode( learner, refresh_interval_ms=2000, stale_after_ms=4000)) sources.append(cacher) with program.group('actor'): # Add actors which pull round-robin from our variable sources. for actor_id in range(self._num_actors): source = sources[actor_id % len(sources)] program.add_node( lp.CourierNode(self.actor, replay, source, counter, actor_id)) return program
acme-master
acme/agents/tf/mog_mpo/agent_distributed.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Implementations of a (MoG) distributional MPO agent.""" from acme.agents.tf.mog_mpo.agent_distributed import DistributedMoGMPO from acme.agents.tf.mog_mpo.learning import MoGMPOLearner from acme.agents.tf.mog_mpo.learning import PolicyEvaluationConfig from acme.agents.tf.mog_mpo.networks import make_default_networks
acme-master
acme/agents/tf/mog_mpo/__init__.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Shared helpers for different experiment flavours.""" from typing import Mapping, Sequence from acme import specs from acme.tf import networks from acme.tf import utils as tf2_utils import numpy as np import sonnet as snt def make_default_networks( environment_spec: specs.EnvironmentSpec, *, policy_layer_sizes: Sequence[int] = (256, 256, 256), critic_layer_sizes: Sequence[int] = (512, 512, 256), policy_init_scale: float = 0.7, critic_init_scale: float = 1e-3, critic_num_components: int = 5, ) -> Mapping[str, snt.Module]: """Creates networks used by the agent.""" # Unpack the environment spec to get appropriate shapes, dtypes, etc. act_spec = environment_spec.actions obs_spec = environment_spec.observations num_dimensions = np.prod(act_spec.shape, dtype=int) # Create the observation network and make sure it's a Sonnet module. observation_network = tf2_utils.batch_concat observation_network = tf2_utils.to_sonnet_module(observation_network) # Create the policy network. policy_network = snt.Sequential([ networks.LayerNormMLP(policy_layer_sizes, activate_final=True), networks.MultivariateNormalDiagHead( num_dimensions, init_scale=policy_init_scale, use_tfd_independent=True) ]) # The multiplexer concatenates the (maybe transformed) observations/actions. critic_network = snt.Sequential([ networks.CriticMultiplexer(action_network=networks.ClipToSpec(act_spec)), networks.LayerNormMLP(critic_layer_sizes, activate_final=True), networks.GaussianMixtureHead( num_dimensions=1, num_components=critic_num_components, init_scale=critic_init_scale) ]) # Create network variables. # Get embedding spec by creating observation network variables. emb_spec = tf2_utils.create_variables(observation_network, [obs_spec]) tf2_utils.create_variables(policy_network, [emb_spec]) tf2_utils.create_variables(critic_network, [emb_spec, act_spec]) return { 'policy': policy_network, 'critic': critic_network, 'observation': observation_network, }
acme-master
acme/agents/tf/mog_mpo/networks.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Distributional MPO with MoG critic learner implementation.""" import dataclasses import time from typing import List, Optional import acme from acme import types from acme.tf import losses from acme.tf import savers as tf2_savers from acme.tf import utils as tf2_utils from acme.utils import counting from acme.utils import loggers import numpy as np import reverb import sonnet as snt import tensorflow as tf import tensorflow_probability as tfp tfd = tfp.distributions @dataclasses.dataclass class PolicyEvaluationConfig: evaluate_stochastic_policy: bool = True num_value_samples: int = 128 class MoGMPOLearner(acme.Learner): """Distributional (MoG) MPO learner.""" def __init__( self, policy_network: snt.Module, critic_network: snt.Module, target_policy_network: snt.Module, target_critic_network: snt.Module, discount: float, num_samples: int, target_policy_update_period: int, target_critic_update_period: int, dataset: tf.data.Dataset, observation_network: snt.Module, target_observation_network: snt.Module, policy_evaluation_config: Optional[PolicyEvaluationConfig] = None, policy_loss_module: Optional[snt.Module] = None, policy_optimizer: Optional[snt.Optimizer] = None, critic_optimizer: Optional[snt.Optimizer] = None, dual_optimizer: Optional[snt.Optimizer] = None, clipping: bool = True, counter: Optional[counting.Counter] = None, logger: Optional[loggers.Logger] = None, checkpoint: bool = True, ): # Store online and target networks. self._policy_network = policy_network self._critic_network = critic_network self._observation_network = observation_network self._target_policy_network = target_policy_network self._target_critic_network = target_critic_network self._target_observation_network = target_observation_network # General learner book-keeping and loggers. self._counter = counter or counting.Counter() self._logger = logger or loggers.make_default_logger('learner') # Other learner parameters. self._discount = discount self._num_samples = num_samples if policy_evaluation_config is None: policy_evaluation_config = PolicyEvaluationConfig() self._policy_evaluation_config = policy_evaluation_config self._clipping = clipping # Necessary to track when to update target networks. self._num_steps = tf.Variable(0, dtype=tf.int32) self._target_policy_update_period = target_policy_update_period self._target_critic_update_period = target_critic_update_period # Batch dataset and create iterator. self._iterator = iter(dataset) # pytype: disable=wrong-arg-types self._policy_loss_module = policy_loss_module or losses.MPO( epsilon=1e-1, epsilon_mean=3e-3, epsilon_stddev=1e-6, epsilon_penalty=1e-3, init_log_temperature=10., init_log_alpha_mean=10., init_log_alpha_stddev=1000.) # Create the optimizers. self._critic_optimizer = critic_optimizer or snt.optimizers.Adam(1e-4) self._policy_optimizer = policy_optimizer or snt.optimizers.Adam(1e-4) self._dual_optimizer = dual_optimizer or snt.optimizers.Adam(1e-2) # Expose the variables. policy_network_to_expose = snt.Sequential( [self._target_observation_network, self._target_policy_network]) self._variables = { 'critic': self._target_critic_network.variables, 'policy': policy_network_to_expose.variables, } # Create a checkpointer and snapshotter object. self._checkpointer = None self._snapshotter = None if checkpoint: self._checkpointer = tf2_savers.Checkpointer( subdirectory='mog_mpo_learner', objects_to_save={ 'counter': self._counter, 'policy': self._policy_network, 'critic': self._critic_network, 'observation': self._observation_network, 'target_policy': self._target_policy_network, 'target_critic': self._target_critic_network, 'target_observation': self._target_observation_network, 'policy_optimizer': self._policy_optimizer, 'critic_optimizer': self._critic_optimizer, 'dual_optimizer': self._dual_optimizer, 'policy_loss_module': self._policy_loss_module, 'num_steps': self._num_steps, }) self._snapshotter = tf2_savers.Snapshotter( objects_to_save={ 'policy': snt.Sequential([ self._target_observation_network, self._target_policy_network ]), }) # Do not record timestamps until after the first learning step is done. # This is to avoid including the time it takes for actors to come online and # fill the replay buffer. self._timestamp = None @tf.function def _step(self, inputs: reverb.ReplaySample) -> types.NestedTensor: # Get data from replay (dropping extras if any). Note there is no # extra data here because we do not insert any into Reverb. o_tm1, a_tm1, r_t, d_t, o_t = (inputs.data.observation, inputs.data.action, inputs.data.reward, inputs.data.discount, inputs.data.next_observation) # Cast the additional discount to match the environment discount dtype. discount = tf.cast(self._discount, dtype=d_t.dtype) with tf.GradientTape(persistent=True) as tape: # Maybe transform the observation before feeding into policy and critic. # Transforming the observations this way at the start of the learning # step effectively means that the policy and critic share observation # network weights. o_tm1 = self._observation_network(o_tm1) # This stop_gradient prevents gradients to propagate into the target # observation network. In addition, since the online policy network is # evaluated at o_t, this also means the policy loss does not influence # the observation network training. o_t = tf.stop_gradient(self._target_observation_network(o_t)) # Get online and target action distributions from policy networks. online_action_distribution = self._policy_network(o_t) target_action_distribution = self._target_policy_network(o_t) # Sample actions to evaluate policy; of size [N, B, ...]. sampled_actions = target_action_distribution.sample(self._num_samples) # Tile embedded observations to feed into the target critic network. # Note: this is more efficient than tiling before the embedding layer. tiled_o_t = tf2_utils.tile_tensor(o_t, self._num_samples) # [N, B, ...] # Compute target-estimated distributional value of sampled actions at o_t. sampled_q_t_distributions = self._target_critic_network( # Merge batch dimensions; to shape [N*B, ...]. snt.merge_leading_dims(tiled_o_t, num_dims=2), snt.merge_leading_dims(sampled_actions, num_dims=2)) # Compute online critic value distribution of a_tm1 in state o_tm1. q_tm1_distribution = self._critic_network(o_tm1, a_tm1) # [B, ...] # Get the return distributions used in the policy evaluation bootstrap. if self._policy_evaluation_config.evaluate_stochastic_policy: z_distributions = sampled_q_t_distributions num_joint_samples = self._num_samples else: z_distributions = self._target_critic_network( o_t, target_action_distribution.mean()) num_joint_samples = 1 num_value_samples = self._policy_evaluation_config.num_value_samples num_joint_samples *= num_value_samples z_samples = z_distributions.sample(num_value_samples) z_samples = tf.reshape(z_samples, (num_joint_samples, -1, 1)) # Expand dims of reward and discount tensors. reward = r_t[..., tf.newaxis] # [B, 1] full_discount = discount * d_t[..., tf.newaxis] target_q = reward + full_discount * z_samples # [N, B, 1] target_q = tf.stop_gradient(target_q) # Compute sample-based cross-entropy. log_probs_q = q_tm1_distribution.log_prob(target_q) # [N, B, 1] critic_loss = -tf.reduce_mean(log_probs_q, axis=0) # [B, 1] critic_loss = tf.reduce_mean(critic_loss) # Compute Q-values of sampled actions and reshape to [N, B]. sampled_q_values = sampled_q_t_distributions.mean() sampled_q_values = tf.reshape(sampled_q_values, (self._num_samples, -1)) # Compute MPO policy loss. policy_loss, policy_stats = self._policy_loss_module( online_action_distribution=online_action_distribution, target_action_distribution=target_action_distribution, actions=sampled_actions, q_values=sampled_q_values) policy_loss = tf.reduce_mean(policy_loss) # For clarity, explicitly define which variables are trained by which loss. critic_trainable_variables = ( # In this agent, the critic loss trains the observation network. self._observation_network.trainable_variables + self._critic_network.trainable_variables) policy_trainable_variables = self._policy_network.trainable_variables # The following are the MPO dual variables, stored in the loss module. dual_trainable_variables = self._policy_loss_module.trainable_variables # Compute gradients. critic_gradients = tape.gradient(critic_loss, critic_trainable_variables) policy_gradients, dual_gradients = tape.gradient( policy_loss, (policy_trainable_variables, dual_trainable_variables)) # Delete the tape manually because of the persistent=True flag. del tape # Maybe clip gradients. if self._clipping: policy_gradients = tuple(tf.clip_by_global_norm(policy_gradients, 40.)[0]) critic_gradients = tuple(tf.clip_by_global_norm(critic_gradients, 40.)[0]) # Apply gradients. self._critic_optimizer.apply(critic_gradients, critic_trainable_variables) self._policy_optimizer.apply(policy_gradients, policy_trainable_variables) self._dual_optimizer.apply(dual_gradients, dual_trainable_variables) # Losses to track. fetches = { 'critic_loss': critic_loss, 'policy_loss': policy_loss, } # Log MPO stats. fetches.update(policy_stats) return fetches def step(self): self._maybe_update_target_networks() self._num_steps.assign_add(1) # Run the learning step. fetches = self._step(next(self._iterator)) # Compute elapsed time. timestamp = time.time() elapsed_time = timestamp - self._timestamp if self._timestamp else 0 self._timestamp = timestamp # Update our counts and record it. counts = self._counter.increment(steps=1, walltime=elapsed_time) fetches.update(counts) # Checkpoint and attempt to write the logs. if self._checkpointer is not None: self._checkpointer.save() if self._snapshotter is not None: self._snapshotter.save() self._logger.write(fetches) def get_variables(self, names: List[str]) -> List[List[np.ndarray]]: return [tf2_utils.to_numpy(self._variables[name]) for name in names] def _maybe_update_target_networks(self): # Update target network. online_policy_variables = self._policy_network.variables target_policy_variables = self._target_policy_network.variables online_critic_variables = (*self._observation_network.variables, *self._critic_network.variables) target_critic_variables = (*self._target_observation_network.variables, *self._target_critic_network.variables) # Make online policy -> target policy network update ops. if tf.math.mod(self._num_steps, self._target_policy_update_period) == 0: for src, dest in zip(online_policy_variables, target_policy_variables): dest.assign(src) # Make online critic -> target critic network update ops. if tf.math.mod(self._num_steps, self._target_critic_update_period) == 0: for src, dest in zip(online_critic_variables, target_critic_variables): dest.assign(src)
acme-master
acme/agents/tf/mog_mpo/learning.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Integration test for the distributed agent.""" from typing import Sequence, Tuple import acme from acme import specs from acme import wrappers from acme.agents.tf import mompo from acme.tf import networks from acme.tf import utils as tf2_utils from acme.utils import lp_utils from dm_control import suite import launchpad as lp import numpy as np import sonnet as snt import tensorflow as tf from absl.testing import absltest from absl.testing import parameterized def make_networks( action_spec: specs.BoundedArray, num_critic_heads: int, policy_layer_sizes: Sequence[int] = (50,), critic_layer_sizes: Sequence[int] = (50,), num_layers_shared: int = 1, distributional_critic: bool = True, vmin: float = -150., vmax: float = 150., num_atoms: int = 51, ): """Creates networks used by the agent.""" num_dimensions = np.prod(action_spec.shape, dtype=int) policy_network = snt.Sequential([ networks.LayerNormMLP(policy_layer_sizes, activate_final=True), networks.MultivariateNormalDiagHead( num_dimensions, tanh_mean=False, init_scale=0.69) ]) if not distributional_critic: critic_layer_sizes = list(critic_layer_sizes) + [1] if not num_layers_shared: # No layers are shared critic_network_base = None else: critic_network_base = networks.LayerNormMLP( critic_layer_sizes[:num_layers_shared], activate_final=True) critic_network_heads = [ snt.nets.MLP(critic_layer_sizes, activation=tf.nn.elu, activate_final=False) for _ in range(num_critic_heads)] if distributional_critic: critic_network_heads = [ snt.Sequential([ c, networks.DiscreteValuedHead(vmin, vmax, num_atoms) ]) for c in critic_network_heads] # The multiplexer concatenates the (maybe transformed) observations/actions. critic_network = snt.Sequential([ networks.CriticMultiplexer( critic_network=critic_network_base, action_network=networks.ClipToSpec(action_spec)), networks.Multihead(network_heads=critic_network_heads), ]) return { 'policy': policy_network, 'critic': critic_network, 'observation': tf2_utils.batch_concat, } def make_environment(evaluation: bool = False): del evaluation # Unused. environment = suite.load('cartpole', 'balance') wrapped = wrappers.SinglePrecisionWrapper(environment) return wrapped def compute_action_norm(target_pi_samples: tf.Tensor, target_q_target_pi_samples: tf.Tensor) -> tf.Tensor: """Compute Q-values for the action norm objective from action samples.""" del target_q_target_pi_samples action_norm = tf.norm(target_pi_samples, ord=2, axis=-1) return tf.stop_gradient(-1 * action_norm) def task_reward_fn(observation: tf.Tensor, action: tf.Tensor, reward: tf.Tensor) -> tf.Tensor: del observation, action return tf.stop_gradient(reward) def make_objectives() -> Tuple[ Sequence[mompo.RewardObjective], Sequence[mompo.QValueObjective]]: """Define the multiple objectives for the policy to learn.""" task_reward = mompo.RewardObjective( name='task', reward_fn=task_reward_fn) action_norm = mompo.QValueObjective( name='action_norm_q', qvalue_fn=compute_action_norm) return [task_reward], [action_norm] class DistributedAgentTest(parameterized.TestCase): """Simple integration/smoke test for the distributed agent.""" @parameterized.named_parameters( ('distributional_critic', True), ('vanilla_critic', False)) def test_agent(self, distributional_critic): # Create objectives. reward_objectives, qvalue_objectives = make_objectives() network_factory = lp_utils.partial_kwargs( make_networks, distributional_critic=distributional_critic) agent = mompo.DistributedMultiObjectiveMPO( reward_objectives, qvalue_objectives, environment_factory=make_environment, network_factory=network_factory, num_actors=2, batch_size=32, min_replay_size=32, max_replay_size=1000, ) program = agent.build() (learner_node,) = program.groups['learner'] learner_node.disable_run() lp.launch(program, launch_type='test_mt') learner: acme.Learner = learner_node.create_handle().dereference() for _ in range(5): learner.step() if __name__ == '__main__': absltest.main()
acme-master
acme/agents/tf/mompo/agent_distributed_test.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Defines the multi-objective MPO distributed agent class.""" from typing import Callable, Dict, Optional, Sequence import acme from acme import datasets from acme import specs from acme.adders import reverb as adders from acme.agents.tf import actors from acme.agents.tf.mompo import learning from acme.tf import losses from acme.tf import networks from acme.tf import savers as tf2_savers from acme.tf import utils as tf2_utils from acme.tf import variable_utils as tf2_variable_utils from acme.utils import counting from acme.utils import loggers from acme.utils import lp_utils import dm_env import launchpad as lp import reverb import sonnet as snt import tensorflow as tf MultiObjectiveNetworkFactorySpec = Callable[ [specs.BoundedArray, int], Dict[str, snt.Module]] MultiObjectivePolicyLossFactorySpec = Callable[[], losses.MultiObjectiveMPO] class DistributedMultiObjectiveMPO: """Program definition for multi-objective MPO. This agent distinguishes itself from the distributed MPO agent in two ways: - Allowing for one or more objectives (see `acme/agents/tf/mompo/learning.py` for details on what form this sequence of objectives should take) - Optionally using a distributional critic (state-action value approximator) as in DMPO. In other words, the critic network can output either scalar Q-values or a DiscreteValuedDistribution. """ def __init__( self, reward_objectives: Sequence[learning.RewardObjective], qvalue_objectives: Sequence[learning.QValueObjective], environment_factory: Callable[[bool], dm_env.Environment], network_factory: MultiObjectiveNetworkFactorySpec, num_actors: int = 1, num_caches: int = 0, environment_spec: Optional[specs.EnvironmentSpec] = None, batch_size: int = 512, prefetch_size: int = 4, min_replay_size: int = 1000, max_replay_size: int = 1000000, samples_per_insert: Optional[float] = None, n_step: int = 5, max_in_flight_items: int = 5, num_samples: int = 20, additional_discount: float = 0.99, target_policy_update_period: int = 200, target_critic_update_period: int = 200, policy_loss_factory: Optional[MultiObjectivePolicyLossFactorySpec] = None, max_actor_steps: Optional[int] = None, log_every: float = 10.0, ): if environment_spec is None: environment_spec = specs.make_environment_spec(environment_factory(False)) self._environment_factory = environment_factory self._network_factory = network_factory self._policy_loss_factory = policy_loss_factory self._environment_spec = environment_spec self._num_actors = num_actors self._num_caches = num_caches self._batch_size = batch_size self._prefetch_size = prefetch_size self._min_replay_size = min_replay_size self._max_replay_size = max_replay_size self._samples_per_insert = samples_per_insert self._n_step = n_step self._max_in_flight_items = max_in_flight_items self._additional_discount = additional_discount self._num_samples = num_samples self._target_policy_update_period = target_policy_update_period self._target_critic_update_period = target_critic_update_period self._max_actor_steps = max_actor_steps self._log_every = log_every self._reward_objectives = reward_objectives self._qvalue_objectives = qvalue_objectives self._num_critic_heads = len(self._reward_objectives) if not self._reward_objectives: raise ValueError('Must specify at least one reward objective.') def replay(self): """The replay storage.""" if self._samples_per_insert is not None: # Create enough of an error buffer to give a 10% tolerance in rate. samples_per_insert_tolerance = 0.1 * self._samples_per_insert error_buffer = self._min_replay_size * samples_per_insert_tolerance limiter = reverb.rate_limiters.SampleToInsertRatio( min_size_to_sample=self._min_replay_size, samples_per_insert=self._samples_per_insert, error_buffer=error_buffer) else: limiter = reverb.rate_limiters.MinSize(self._min_replay_size) replay_table = reverb.Table( name=adders.DEFAULT_PRIORITY_TABLE, sampler=reverb.selectors.Uniform(), remover=reverb.selectors.Fifo(), max_size=self._max_replay_size, rate_limiter=limiter, signature=adders.NStepTransitionAdder.signature( self._environment_spec)) return [replay_table] def counter(self): return tf2_savers.CheckpointingRunner(counting.Counter(), time_delta_minutes=1, subdirectory='counter') def coordinator(self, counter: counting.Counter, max_actor_steps: int): return lp_utils.StepsLimiter(counter, max_actor_steps) def learner( self, replay: reverb.Client, counter: counting.Counter, ): """The Learning part of the agent.""" act_spec = self._environment_spec.actions obs_spec = self._environment_spec.observations # Create online and target networks. online_networks = self._network_factory(act_spec, self._num_critic_heads) target_networks = self._network_factory(act_spec, self._num_critic_heads) # Make sure observation network is a Sonnet Module. observation_network = online_networks.get('observation', tf.identity) target_observation_network = target_networks.get('observation', tf.identity) observation_network = tf2_utils.to_sonnet_module(observation_network) target_observation_network = tf2_utils.to_sonnet_module( target_observation_network) # Get embedding spec and create observation network variables. emb_spec = tf2_utils.create_variables(observation_network, [obs_spec]) # Create variables. tf2_utils.create_variables(online_networks['policy'], [emb_spec]) tf2_utils.create_variables(online_networks['critic'], [emb_spec, act_spec]) tf2_utils.create_variables(target_networks['policy'], [emb_spec]) tf2_utils.create_variables(target_networks['critic'], [emb_spec, act_spec]) tf2_utils.create_variables(target_observation_network, [obs_spec]) # The dataset object to learn from. dataset = datasets.make_reverb_dataset(server_address=replay.server_address) dataset = dataset.batch(self._batch_size, drop_remainder=True) dataset = dataset.prefetch(self._prefetch_size) counter = counting.Counter(counter, 'learner') logger = loggers.make_default_logger( 'learner', time_delta=self._log_every, steps_key='learner_steps') # Create policy loss module if a factory is passed. if self._policy_loss_factory: policy_loss_module = self._policy_loss_factory() else: policy_loss_module = None # Return the learning agent. return learning.MultiObjectiveMPOLearner( reward_objectives=self._reward_objectives, qvalue_objectives=self._qvalue_objectives, policy_network=online_networks['policy'], critic_network=online_networks['critic'], observation_network=observation_network, target_policy_network=target_networks['policy'], target_critic_network=target_networks['critic'], target_observation_network=target_observation_network, discount=self._additional_discount, num_samples=self._num_samples, target_policy_update_period=self._target_policy_update_period, target_critic_update_period=self._target_critic_update_period, policy_loss_module=policy_loss_module, dataset=dataset, counter=counter, logger=logger) def actor( self, replay: reverb.Client, variable_source: acme.VariableSource, counter: counting.Counter, ) -> acme.EnvironmentLoop: """The actor process.""" action_spec = self._environment_spec.actions observation_spec = self._environment_spec.observations # Create environment and target networks to act with. environment = self._environment_factory(False) agent_networks = self._network_factory(action_spec, self._num_critic_heads) # Make sure observation network is defined. observation_network = agent_networks.get('observation', tf.identity) # Create a stochastic behavior policy. behavior_network = snt.Sequential([ observation_network, agent_networks['policy'], networks.StochasticSamplingHead(), ]) # Ensure network variables are created. tf2_utils.create_variables(behavior_network, [observation_spec]) policy_variables = {'policy': behavior_network.variables} # Create the variable client responsible for keeping the actor up-to-date. variable_client = tf2_variable_utils.VariableClient( variable_source, policy_variables, update_period=1000) # Make sure not to use a random policy after checkpoint restoration by # assigning variables before running the environment loop. variable_client.update_and_wait() # Component to add things into replay. adder = adders.NStepTransitionAdder( client=replay, n_step=self._n_step, max_in_flight_items=self._max_in_flight_items, discount=self._additional_discount) # Create the agent. actor = actors.FeedForwardActor( policy_network=behavior_network, adder=adder, variable_client=variable_client) # Create logger and counter; actors will not spam bigtable. counter = counting.Counter(counter, 'actor') logger = loggers.make_default_logger( 'actor', save_data=False, time_delta=self._log_every, steps_key='actor_steps') # Create the run loop and return it. return acme.EnvironmentLoop( environment, actor, counter, logger) def evaluator( self, variable_source: acme.VariableSource, counter: counting.Counter, ): """The evaluation process.""" action_spec = self._environment_spec.actions observation_spec = self._environment_spec.observations # Create environment and target networks to act with. environment = self._environment_factory(True) agent_networks = self._network_factory(action_spec, self._num_critic_heads) # Make sure observation network is defined. observation_network = agent_networks.get('observation', tf.identity) # Create a deterministic behavior policy. evaluator_modules = [ observation_network, agent_networks['policy'], networks.StochasticMeanHead(), ] if isinstance(action_spec, specs.BoundedArray): evaluator_modules += [networks.ClipToSpec(action_spec)] evaluator_network = snt.Sequential(evaluator_modules) # Ensure network variables are created. tf2_utils.create_variables(evaluator_network, [observation_spec]) policy_variables = {'policy': evaluator_network.variables} # Create the variable client responsible for keeping the actor up-to-date. variable_client = tf2_variable_utils.VariableClient( variable_source, policy_variables, update_period=1000) # Make sure not to evaluate a random actor by assigning variables before # running the environment loop. variable_client.update_and_wait() # Create the agent. evaluator = actors.FeedForwardActor( policy_network=evaluator_network, variable_client=variable_client) # Create logger and counter. counter = counting.Counter(counter, 'evaluator') logger = loggers.make_default_logger( 'evaluator', time_delta=self._log_every, steps_key='evaluator_steps') # Create the run loop and return it. return acme.EnvironmentLoop( environment, evaluator, counter, logger) def build(self, name='mompo'): """Build the distributed agent topology.""" program = lp.Program(name=name) with program.group('replay'): replay = program.add_node(lp.ReverbNode(self.replay)) with program.group('counter'): counter = program.add_node(lp.CourierNode(self.counter)) if self._max_actor_steps: _ = program.add_node( lp.CourierNode(self.coordinator, counter, self._max_actor_steps)) with program.group('learner'): learner = program.add_node( lp.CourierNode(self.learner, replay, counter)) with program.group('evaluator'): program.add_node( lp.CourierNode(self.evaluator, learner, counter)) if not self._num_caches: # Use our learner as a single variable source. sources = [learner] else: with program.group('cacher'): # Create a set of learner caches. sources = [] for _ in range(self._num_caches): cacher = program.add_node( lp.CacherNode( learner, refresh_interval_ms=2000, stale_after_ms=4000)) sources.append(cacher) with program.group('actor'): # Add actors which pull round-robin from our variable sources. for actor_id in range(self._num_actors): source = sources[actor_id % len(sources)] program.add_node(lp.CourierNode(self.actor, replay, source, counter)) return program
acme-master
acme/agents/tf/mompo/agent_distributed.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Implementations of a distributional MPO agent.""" from acme.agents.tf.mompo.agent import MultiObjectiveMPO from acme.agents.tf.mompo.agent_distributed import DistributedMultiObjectiveMPO from acme.agents.tf.mompo.learning import MultiObjectiveMPOLearner from acme.agents.tf.mompo.learning import QValueObjective from acme.agents.tf.mompo.learning import RewardObjective
acme-master
acme/agents/tf/mompo/__init__.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Tests for the multi-objective MPO agent.""" from typing import Dict, Sequence, Tuple import acme from acme import specs from acme.agents.tf import mompo from acme.testing import fakes from acme.tf import networks import numpy as np import sonnet as snt import tensorflow as tf from absl.testing import absltest from absl.testing import parameterized def make_networks( action_spec: specs.Array, num_critic_heads: int, policy_layer_sizes: Sequence[int] = (300, 200), critic_layer_sizes: Sequence[int] = (400, 300), num_layers_shared: int = 1, distributional_critic: bool = True, vmin: float = -150., vmax: float = 150., num_atoms: int = 51, ) -> Dict[str, snt.Module]: """Creates networks used by the agent.""" num_dimensions = np.prod(action_spec.shape, dtype=int) policy_network = snt.Sequential([ networks.LayerNormMLP(policy_layer_sizes, activate_final=True), networks.MultivariateNormalDiagHead( num_dimensions, tanh_mean=False, init_scale=0.69) ]) if not distributional_critic: critic_layer_sizes = list(critic_layer_sizes) + [1] if not num_layers_shared: # No layers are shared critic_network_base = None else: critic_network_base = networks.LayerNormMLP( critic_layer_sizes[:num_layers_shared], activate_final=True) critic_network_heads = [ snt.nets.MLP(critic_layer_sizes, activation=tf.nn.elu, activate_final=False) for _ in range(num_critic_heads)] if distributional_critic: critic_network_heads = [ snt.Sequential([ c, networks.DiscreteValuedHead(vmin, vmax, num_atoms) ]) for c in critic_network_heads] # The multiplexer concatenates the (maybe transformed) observations/actions. critic_network = snt.Sequential([ networks.CriticMultiplexer( critic_network=critic_network_base), networks.Multihead(network_heads=critic_network_heads), ]) return { 'policy': policy_network, 'critic': critic_network, } def compute_action_norm(target_pi_samples: tf.Tensor, target_q_target_pi_samples: tf.Tensor) -> tf.Tensor: """Compute Q-values for the action norm objective from action samples.""" del target_q_target_pi_samples action_norm = tf.norm(target_pi_samples, ord=2, axis=-1) return tf.stop_gradient(-1 * action_norm) def task_reward_fn(observation: tf.Tensor, action: tf.Tensor, reward: tf.Tensor) -> tf.Tensor: del observation, action return tf.stop_gradient(reward) def make_objectives() -> Tuple[ Sequence[mompo.RewardObjective], Sequence[mompo.QValueObjective]]: """Define the multiple objectives for the policy to learn.""" task_reward = mompo.RewardObjective( name='task', reward_fn=task_reward_fn) action_norm = mompo.QValueObjective( name='action_norm_q', qvalue_fn=compute_action_norm) return [task_reward], [action_norm] class MOMPOTest(parameterized.TestCase): @parameterized.named_parameters( ('distributional_critic', True), ('vanilla_critic', False)) def test_mompo(self, distributional_critic): # Create a fake environment to test with. environment = fakes.ContinuousEnvironment(episode_length=10) spec = specs.make_environment_spec(environment) # Create objectives. reward_objectives, qvalue_objectives = make_objectives() num_critic_heads = len(reward_objectives) # Create networks. agent_networks = make_networks( spec.actions, num_critic_heads=num_critic_heads, distributional_critic=distributional_critic) # Construct the agent. agent = mompo.MultiObjectiveMPO( reward_objectives, qvalue_objectives, spec, policy_network=agent_networks['policy'], critic_network=agent_networks['critic'], batch_size=10, samples_per_insert=2, min_replay_size=10) # Try running the environment loop. We have no assertions here because all # we care about is that the agent runs without raising any errors. loop = acme.EnvironmentLoop(environment, agent) loop.run(num_episodes=2) if __name__ == '__main__': absltest.main()
acme-master
acme/agents/tf/mompo/agent_test.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Multi-objective MPO agent implementation.""" import copy from typing import Optional, Sequence from acme import datasets from acme import specs from acme import types from acme.adders import reverb as adders from acme.agents import agent from acme.agents.tf import actors from acme.agents.tf.mompo import learning from acme.tf import losses from acme.tf import networks from acme.tf import utils as tf2_utils from acme.utils import counting from acme.utils import loggers import reverb import sonnet as snt import tensorflow as tf class MultiObjectiveMPO(agent.Agent): """Multi-objective MPO Agent. This implements a single-process multi-objective MPO agent. This is an actor-critic algorithm that generates data via a behavior policy, inserts N-step transitions into a replay buffer, and periodically updates the policy (and as a result the behavior) by sampling uniformly from this buffer. This agent distinguishes itself from the MPO agent in two ways: - Allowing for one or more objectives (see `acme/agents/tf/mompo/learning.py` for details on what form this sequence of objectives should take) - Optionally using a distributional critic (state-action value approximator) as in DMPO. In other words, the critic network can output either scalar Q-values or a DiscreteValuedDistribution. """ def __init__(self, reward_objectives: Sequence[learning.RewardObjective], qvalue_objectives: Sequence[learning.QValueObjective], environment_spec: specs.EnvironmentSpec, policy_network: snt.Module, critic_network: snt.Module, observation_network: types.TensorTransformation = tf.identity, discount: float = 0.99, batch_size: int = 512, prefetch_size: int = 4, target_policy_update_period: int = 200, target_critic_update_period: int = 200, min_replay_size: int = 1000, max_replay_size: int = 1000000, samples_per_insert: float = 16., policy_loss_module: Optional[losses.MultiObjectiveMPO] = None, policy_optimizer: Optional[snt.Optimizer] = None, critic_optimizer: Optional[snt.Optimizer] = None, n_step: int = 5, num_samples: int = 20, clipping: bool = True, logger: Optional[loggers.Logger] = None, counter: Optional[counting.Counter] = None, checkpoint: bool = True, replay_table_name: str = adders.DEFAULT_PRIORITY_TABLE): """Initialize the agent. Args: reward_objectives: list of the objectives that the policy should optimize; each objective is defined by its reward function qvalue_objectives: list of the objectives that the policy should optimize; each objective is defined by its Q-value function environment_spec: description of the actions, observations, etc. policy_network: the online (optimized) policy. critic_network: the online critic. observation_network: optional network to transform the observations before they are fed into any network. discount: discount to use for TD updates. batch_size: batch size for updates. prefetch_size: size to prefetch from replay. target_policy_update_period: number of updates to perform before updating the target policy network. target_critic_update_period: number of updates to perform before updating the target critic network. min_replay_size: minimum replay size before updating. max_replay_size: maximum replay size. samples_per_insert: number of samples to take from replay for every insert that is made. policy_loss_module: configured MO-MPO loss function for the policy optimization; defaults to sensible values on the control suite. See `acme/tf/losses/mompo.py` for more details. policy_optimizer: optimizer to be used on the policy. critic_optimizer: optimizer to be used on the critic. n_step: number of steps to squash into a single transition. num_samples: number of actions to sample when doing a Monte Carlo integration with respect to the policy. clipping: whether to clip gradients by global norm. logger: logging object used to write to logs. counter: counter object used to keep track of steps. checkpoint: boolean indicating whether to checkpoint the learner. replay_table_name: string indicating what name to give the replay table. """ # Check that at least one objective's reward function is specified. if not reward_objectives: raise ValueError('Must specify at least one reward objective.') # Create a replay server to add data to. replay_table = reverb.Table( name=adders.DEFAULT_PRIORITY_TABLE, sampler=reverb.selectors.Uniform(), remover=reverb.selectors.Fifo(), max_size=max_replay_size, rate_limiter=reverb.rate_limiters.MinSize(min_size_to_sample=1), signature=adders.NStepTransitionAdder.signature(environment_spec)) self._server = reverb.Server([replay_table], port=None) # The adder is used to insert observations into replay. address = f'localhost:{self._server.port}' adder = adders.NStepTransitionAdder( client=reverb.Client(address), n_step=n_step, discount=discount) # The dataset object to learn from. dataset = datasets.make_reverb_dataset( table=replay_table_name, server_address=address, batch_size=batch_size, prefetch_size=prefetch_size) # Make sure observation network is a Sonnet Module. observation_network = tf2_utils.to_sonnet_module(observation_network) # Create target networks before creating online/target network variables. target_policy_network = copy.deepcopy(policy_network) target_critic_network = copy.deepcopy(critic_network) target_observation_network = copy.deepcopy(observation_network) # Get observation and action specs. act_spec = environment_spec.actions obs_spec = environment_spec.observations emb_spec = tf2_utils.create_variables(observation_network, [obs_spec]) # Create the behavior policy. behavior_network = snt.Sequential([ observation_network, policy_network, networks.StochasticSamplingHead(), ]) # Create variables. tf2_utils.create_variables(policy_network, [emb_spec]) tf2_utils.create_variables(critic_network, [emb_spec, act_spec]) tf2_utils.create_variables(target_policy_network, [emb_spec]) tf2_utils.create_variables(target_critic_network, [emb_spec, act_spec]) tf2_utils.create_variables(target_observation_network, [obs_spec]) # Create the actor which defines how we take actions. actor = actors.FeedForwardActor( policy_network=behavior_network, adder=adder) # Create optimizers. policy_optimizer = policy_optimizer or snt.optimizers.Adam(1e-4) critic_optimizer = critic_optimizer or snt.optimizers.Adam(1e-4) # The learner updates the parameters (and initializes them). learner = learning.MultiObjectiveMPOLearner( reward_objectives=reward_objectives, qvalue_objectives=qvalue_objectives, policy_network=policy_network, critic_network=critic_network, observation_network=observation_network, target_policy_network=target_policy_network, target_critic_network=target_critic_network, target_observation_network=target_observation_network, policy_loss_module=policy_loss_module, policy_optimizer=policy_optimizer, critic_optimizer=critic_optimizer, clipping=clipping, discount=discount, num_samples=num_samples, target_policy_update_period=target_policy_update_period, target_critic_update_period=target_critic_update_period, dataset=dataset, logger=logger, counter=counter, checkpoint=checkpoint) super().__init__( actor=actor, learner=learner, min_observations=max(batch_size, min_replay_size), observations_per_step=float(batch_size) / samples_per_insert)
acme-master
acme/agents/tf/mompo/agent.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Multi-objective MPO learner implementation.""" import dataclasses import time from typing import Callable, List, Optional, Sequence import acme from acme import types from acme.tf import losses from acme.tf import networks from acme.tf import savers as tf2_savers from acme.tf import utils as tf2_utils from acme.utils import counting from acme.utils import loggers import numpy as np import sonnet as snt import tensorflow as tf import trfl QValueFunctionSpec = Callable[[tf.Tensor, tf.Tensor], tf.Tensor] RewardFunctionSpec = Callable[[tf.Tensor, tf.Tensor, tf.Tensor], tf.Tensor] _DEFAULT_EPSILON = 1e-1 _DEFAULT_EPSILON_MEAN = 1e-3 _DEFAULT_EPSILON_STDDEV = 1e-6 _DEFAULT_INIT_LOG_TEMPERATURE = 1. _DEFAULT_INIT_LOG_ALPHA_MEAN = 1. _DEFAULT_INIT_LOG_ALPHA_STDDEV = 10. @dataclasses.dataclass class QValueObjective: """Defines an objective by specifying its 'Q-values' directly.""" name: str # This computes "Q-values" directly from the sampled actions and other Q's. qvalue_fn: QValueFunctionSpec @dataclasses.dataclass class RewardObjective: """Defines an objective by specifying its reward function.""" name: str # This computes the reward from observations, actions, and environment task # reward. In the learner, a head will automatically be added to the critic # network, to learn Q-values for this objective. reward_fn: RewardFunctionSpec class MultiObjectiveMPOLearner(acme.Learner): """Distributional MPO learner. This is the learning component of a multi-objective MPO (MO-MPO) agent. Two sequences of objectives must be specified. Otherwise, the inputs are identical to those of the MPO / DMPO learners. Each objective must be defined as either a RewardObjective or an QValueObjective. These objectives are provided by the reward_objectives and qvalue_objectives parameters, respectively. For each RewardObjective, a critic will be trained to estimate Q-values for that objective. Whereas for each QValueObjective, the Q-values are computed directly by its qvalue_fn. A RewardObjective's reward_fn takes the observation, action, and environment reward as input, and returns the reward for that objective. For example, if the environment reward is a scalar, then an objective corresponding to the = task would simply return the environment reward. A QValueObjective's qvalue_fn takes the actions and reward-based objectives' Q-values as input, and outputs the "Q-values" for that objective. For instance, in the MO-MPO paper ([Abdolmaleki, Huang et al., 2020]), the action norm objective in the Humanoid run task is defined by setting the qvalue_fn to be the l2-norm of the actions. Note: If there is only one objective and that is the task reward, then this algorithm becomes exactly the same as (D)MPO. (Abdolmaleki, Huang et al., 2020): https://arxiv.org/pdf/2005.07513.pdf """ def __init__( self, reward_objectives: Sequence[RewardObjective], qvalue_objectives: Sequence[QValueObjective], policy_network: snt.Module, critic_network: snt.Module, target_policy_network: snt.Module, target_critic_network: snt.Module, discount: float, num_samples: int, target_policy_update_period: int, target_critic_update_period: int, dataset: tf.data.Dataset, observation_network: types.TensorTransformation = tf.identity, target_observation_network: types.TensorTransformation = tf.identity, policy_loss_module: Optional[losses.MultiObjectiveMPO] = None, policy_optimizer: Optional[snt.Optimizer] = None, critic_optimizer: Optional[snt.Optimizer] = None, dual_optimizer: Optional[snt.Optimizer] = None, clipping: bool = True, counter: Optional[counting.Counter] = None, logger: Optional[loggers.Logger] = None, checkpoint: bool = True, ): # Store online and target networks. self._policy_network = policy_network self._critic_network = critic_network self._target_policy_network = target_policy_network self._target_critic_network = target_critic_network # Make sure observation networks are snt.Module's so they have variables. self._observation_network = tf2_utils.to_sonnet_module(observation_network) self._target_observation_network = tf2_utils.to_sonnet_module( target_observation_network) # General learner book-keeping and loggers. self._counter = counter or counting.Counter() self._logger = logger or loggers.make_default_logger('learner') # Other learner parameters. self._discount = discount self._num_samples = num_samples self._clipping = clipping # Necessary to track when to update target networks. self._num_steps = tf.Variable(0, dtype=tf.int32) self._target_policy_update_period = target_policy_update_period self._target_critic_update_period = target_critic_update_period # Batch dataset and create iterator. # TODO(b/155086959): Fix type stubs and remove. self._iterator = iter(dataset) # pytype: disable=wrong-arg-types # Store objectives self._reward_objectives = reward_objectives self._qvalue_objectives = qvalue_objectives if self._qvalue_objectives is None: self._qvalue_objectives = [] self._num_critic_heads = len(self._reward_objectives) # C self._objective_names = ( [x.name for x in self._reward_objectives] + [x.name for x in self._qvalue_objectives]) self._policy_loss_module = policy_loss_module or losses.MultiObjectiveMPO( epsilons=[losses.KLConstraint(name, _DEFAULT_EPSILON) for name in self._objective_names], epsilon_mean=_DEFAULT_EPSILON_MEAN, epsilon_stddev=_DEFAULT_EPSILON_STDDEV, init_log_temperature=_DEFAULT_INIT_LOG_TEMPERATURE, init_log_alpha_mean=_DEFAULT_INIT_LOG_ALPHA_MEAN, init_log_alpha_stddev=_DEFAULT_INIT_LOG_ALPHA_STDDEV) # Check that ordering of objectives matches the policy_loss_module's if self._objective_names != list(self._policy_loss_module.objective_names): raise ValueError("Agent's ordering of objectives doesn't match " "the policy loss module's ordering of epsilons.") # Create the optimizers. self._critic_optimizer = critic_optimizer or snt.optimizers.Adam(1e-4) self._policy_optimizer = policy_optimizer or snt.optimizers.Adam(1e-4) self._dual_optimizer = dual_optimizer or snt.optimizers.Adam(1e-2) # Expose the variables. policy_network_to_expose = snt.Sequential( [self._target_observation_network, self._target_policy_network]) self._variables = { 'critic': self._target_critic_network.variables, 'policy': policy_network_to_expose.variables, } # Create a checkpointer and snapshotter object. self._checkpointer = None self._snapshotter = None if checkpoint: self._checkpointer = tf2_savers.Checkpointer( subdirectory='mompo_learner', objects_to_save={ 'counter': self._counter, 'policy': self._policy_network, 'critic': self._critic_network, 'observation': self._observation_network, 'target_policy': self._target_policy_network, 'target_critic': self._target_critic_network, 'target_observation': self._target_observation_network, 'policy_optimizer': self._policy_optimizer, 'critic_optimizer': self._critic_optimizer, 'dual_optimizer': self._dual_optimizer, 'policy_loss_module': self._policy_loss_module, 'num_steps': self._num_steps, }) self._snapshotter = tf2_savers.Snapshotter( objects_to_save={ 'policy': snt.Sequential([ self._target_observation_network, self._target_policy_network ]), }) # Do not record timestamps until after the first learning step is done. # This is to avoid including the time it takes for actors to come online and # fill the replay buffer. self._timestamp: float = None @tf.function def _step(self) -> types.NestedTensor: # Update target network. online_policy_variables = self._policy_network.variables target_policy_variables = self._target_policy_network.variables online_critic_variables = ( *self._observation_network.variables, *self._critic_network.variables, ) target_critic_variables = ( *self._target_observation_network.variables, *self._target_critic_network.variables, ) # Make online policy -> target policy network update ops. if tf.math.mod(self._num_steps, self._target_policy_update_period) == 0: for src, dest in zip(online_policy_variables, target_policy_variables): dest.assign(src) # Make online critic -> target critic network update ops. if tf.math.mod(self._num_steps, self._target_critic_update_period) == 0: for src, dest in zip(online_critic_variables, target_critic_variables): dest.assign(src) self._num_steps.assign_add(1) # Get data from replay (dropping extras if any). Note there is no # extra data here because we do not insert any into Reverb. inputs = next(self._iterator) transitions: types.Transition = inputs.data with tf.GradientTape(persistent=True) as tape: # Maybe transform the observation before feeding into policy and critic. # Transforming the observations this way at the start of the learning # step effectively means that the policy and critic share observation # network weights. o_tm1 = self._observation_network(transitions.observation) # This stop_gradient prevents gradients to propagate into the target # observation network. In addition, since the online policy network is # evaluated at o_t, this also means the policy loss does not influence # the observation network training. o_t = tf.stop_gradient( self._target_observation_network(transitions.next_observation)) # Get online and target action distributions from policy networks. online_action_distribution = self._policy_network(o_t) target_action_distribution = self._target_policy_network(o_t) # Sample actions to evaluate policy; of size [N, B, ...]. sampled_actions = target_action_distribution.sample(self._num_samples) # Tile embedded observations to feed into the target critic network. # Note: this is more efficient than tiling before the embedding layer. tiled_o_t = tf2_utils.tile_tensor(o_t, self._num_samples) # [N, B, ...] # Compute target-estimated distributional value of sampled actions at o_t. sampled_q_t_all = self._target_critic_network( # Merge batch dimensions; to shape [N*B, ...]. snt.merge_leading_dims(tiled_o_t, num_dims=2), snt.merge_leading_dims(sampled_actions, num_dims=2)) # Compute online critic value distribution of a_tm1 in state o_tm1. q_tm1_all = self._critic_network(o_tm1, transitions.action) # Compute rewards for objectives with defined reward_fn reward_stats = {} r_t_all = [] for objective in self._reward_objectives: r = objective.reward_fn(o_tm1, transitions.action, transitions.reward) reward_stats['{}_reward'.format(objective.name)] = tf.reduce_mean(r) r_t_all.append(r) r_t_all = tf.stack(r_t_all, axis=-1) r_t_all.get_shape().assert_has_rank(2) # [B, C] if isinstance(sampled_q_t_all, list): # Distributional critics critic_loss, sampled_q_t = _compute_distributional_critic_loss( sampled_q_t_all, q_tm1_all, r_t_all, transitions.discount, self._discount, self._num_samples) else: critic_loss, sampled_q_t = _compute_critic_loss( sampled_q_t_all, q_tm1_all, r_t_all, transitions.discount, self._discount, self._num_samples, self._num_critic_heads) # Add sampled Q-values for objectives with defined qvalue_fn sampled_q_t_k = [sampled_q_t] for objective in self._qvalue_objectives: sampled_q_t_k.append(tf.expand_dims(tf.stop_gradient( objective.qvalue_fn(sampled_actions, sampled_q_t)), axis=-1)) sampled_q_t_k = tf.concat(sampled_q_t_k, axis=-1) # [N, B, K] # Compute MPO policy loss. policy_loss, policy_stats = self._policy_loss_module( online_action_distribution=online_action_distribution, target_action_distribution=target_action_distribution, actions=sampled_actions, q_values=sampled_q_t_k) # For clarity, explicitly define which variables are trained by which loss. critic_trainable_variables = ( # In this agent, the critic loss trains the observation network. self._observation_network.trainable_variables + self._critic_network.trainable_variables) policy_trainable_variables = self._policy_network.trainable_variables # The following are the MPO dual variables, stored in the loss module. dual_trainable_variables = self._policy_loss_module.trainable_variables # Compute gradients. critic_gradients = tape.gradient(critic_loss, critic_trainable_variables) policy_gradients, dual_gradients = tape.gradient( policy_loss, (policy_trainable_variables, dual_trainable_variables)) # Delete the tape manually because of the persistent=True flag. del tape # Maybe clip gradients. if self._clipping: policy_gradients = tuple(tf.clip_by_global_norm(policy_gradients, 40.)[0]) critic_gradients = tuple(tf.clip_by_global_norm(critic_gradients, 40.)[0]) # Apply gradients. self._critic_optimizer.apply(critic_gradients, critic_trainable_variables) self._policy_optimizer.apply(policy_gradients, policy_trainable_variables) self._dual_optimizer.apply(dual_gradients, dual_trainable_variables) # Losses to track. fetches = { 'critic_loss': critic_loss, 'policy_loss': policy_loss, } fetches.update(policy_stats) # Log MPO stats. fetches.update(reward_stats) # Log reward stats. return fetches def step(self): # Run the learning step. fetches = self._step() # Compute elapsed time. timestamp = time.time() elapsed_time = timestamp - self._timestamp if self._timestamp else 0 self._timestamp = timestamp # Update our counts and record it. counts = self._counter.increment(steps=1, walltime=elapsed_time) fetches.update(counts) # Checkpoint and attempt to write the logs. if self._checkpointer is not None: self._checkpointer.save() if self._snapshotter is not None: self._snapshotter.save() self._logger.write(fetches) def get_variables(self, names: List[str]) -> List[List[np.ndarray]]: return [tf2_utils.to_numpy(self._variables[name]) for name in names] def _compute_distributional_critic_loss( sampled_q_t_all: List[tf.Tensor], q_tm1_all: List[tf.Tensor], r_t_all: tf.Tensor, d_t: tf.Tensor, discount: float, num_samples: int): """Compute loss and sampled Q-values for distributional critics.""" # Compute average logits by first reshaping them and normalizing them # across atoms. batch_size = r_t_all.get_shape()[0] # Cast the additional discount to match the environment discount dtype. discount = tf.cast(discount, dtype=d_t.dtype) critic_losses = [] sampled_q_ts = [] for idx, (sampled_q_t_distributions, q_tm1_distribution) in enumerate( zip(sampled_q_t_all, q_tm1_all)): # Compute loss for distributional critic for objective c sampled_logits = tf.reshape( sampled_q_t_distributions.logits, [num_samples, batch_size, -1]) # [N, B, A] sampled_logprobs = tf.math.log_softmax(sampled_logits, axis=-1) averaged_logits = tf.reduce_logsumexp(sampled_logprobs, axis=0) # Construct the expected distributional value for bootstrapping. q_t_distribution = networks.DiscreteValuedDistribution( values=sampled_q_t_distributions.values, logits=averaged_logits) # Compute critic distributional loss. critic_loss = losses.categorical( q_tm1_distribution, r_t_all[:, idx], discount * d_t, q_t_distribution) critic_losses.append(tf.reduce_mean(critic_loss)) # Compute Q-values of sampled actions and reshape to [N, B]. sampled_q_ts.append(tf.reshape( sampled_q_t_distributions.mean(), (num_samples, -1))) critic_loss = tf.reduce_mean(critic_losses) sampled_q_t = tf.stack(sampled_q_ts, axis=-1) # [N, B, C] return critic_loss, sampled_q_t def _compute_critic_loss( sampled_q_t_all: tf.Tensor, q_tm1_all: tf.Tensor, r_t_all: tf.Tensor, d_t: tf.Tensor, discount: float, num_samples: int, num_critic_heads: int): """Compute loss and sampled Q-values for (non-distributional) critics.""" # Reshape Q-value samples back to original batch dimensions and average # them to compute the TD-learning bootstrap target. batch_size = r_t_all.get_shape()[0] sampled_q_t = tf.reshape( sampled_q_t_all, (num_samples, batch_size, num_critic_heads)) # [N,B,C] q_t = tf.reduce_mean(sampled_q_t, axis=0) # [B, C] # Flatten q_t and q_tm1; necessary for trfl.td_learning q_t = tf.reshape(q_t, [-1]) # [B*C] q_tm1 = tf.reshape(q_tm1_all, [-1]) # [B*C] # Flatten r_t_all; necessary for trfl.td_learning r_t_all = tf.reshape(r_t_all, [-1]) # [B*C] # Broadcast and then flatten d_t, to match shape of q_t and q_tm1 d_t = tf.tile(d_t, [num_critic_heads]) # [B*C] # Cast the additional discount to match the environment discount dtype. discount = tf.cast(discount, dtype=d_t.dtype) # Critic loss. critic_loss = trfl.td_learning(q_tm1, r_t_all, discount * d_t, q_t).loss critic_loss = tf.reduce_mean(critic_loss) return critic_loss, sampled_q_t
acme-master
acme/agents/tf/mompo/learning.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Integration test for the distributed agent.""" from typing import Sequence import acme from acme import specs from acme.agents.tf import dmpo from acme.testing import fakes from acme.tf import networks from acme.tf import utils as tf2_utils import launchpad as lp import numpy as np import sonnet as snt from absl.testing import absltest def make_networks( action_spec: specs.BoundedArray, policy_layer_sizes: Sequence[int] = (50,), critic_layer_sizes: Sequence[int] = (50,), vmin: float = -150., vmax: float = 150., num_atoms: int = 51, ): """Creates networks used by the agent.""" num_dimensions = np.prod(action_spec.shape, dtype=int) policy_network = snt.Sequential([ networks.LayerNormMLP(policy_layer_sizes, activate_final=True), networks.MultivariateNormalDiagHead( num_dimensions, tanh_mean=True, init_scale=0.3, fixed_scale=True, use_tfd_independent=False) ]) # The multiplexer concatenates the (maybe transformed) observations/actions. critic_network = networks.CriticMultiplexer( critic_network=networks.LayerNormMLP( critic_layer_sizes, activate_final=True), action_network=networks.ClipToSpec(action_spec)) critic_network = snt.Sequential( [critic_network, networks.DiscreteValuedHead(vmin, vmax, num_atoms)]) return { 'policy': policy_network, 'critic': critic_network, 'observation': tf2_utils.batch_concat, } class DistributedAgentTest(absltest.TestCase): """Simple integration/smoke test for the distributed agent.""" def test_agent(self): agent = dmpo.DistributedDistributionalMPO( environment_factory=lambda x: fakes.ContinuousEnvironment(bounded=True), network_factory=make_networks, num_actors=2, batch_size=32, min_replay_size=32, max_replay_size=1000, ) program = agent.build() (learner_node,) = program.groups['learner'] learner_node.disable_run() lp.launch(program, launch_type='test_mt') learner: acme.Learner = learner_node.create_handle().dereference() for _ in range(5): learner.step() if __name__ == '__main__': absltest.main()
acme-master
acme/agents/tf/dmpo/agent_distributed_test.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Defines the distributional MPO distributed agent class.""" from typing import Callable, Dict, Optional, Sequence import acme from acme import datasets from acme import specs from acme import types from acme.adders import reverb as adders from acme.agents.tf import actors from acme.agents.tf.dmpo import learning from acme.datasets import image_augmentation from acme.tf import networks from acme.tf import savers as tf2_savers from acme.tf import utils as tf2_utils from acme.tf import variable_utils as tf2_variable_utils from acme.utils import counting from acme.utils import loggers from acme.utils import lp_utils from acme.utils import observers as observers_lib import dm_env import launchpad as lp import reverb import sonnet as snt import tensorflow as tf class DistributedDistributionalMPO: """Program definition for distributional MPO.""" def __init__( self, environment_factory: Callable[[bool], dm_env.Environment], network_factory: Callable[[specs.BoundedArray], Dict[str, snt.Module]], num_actors: int = 1, num_caches: int = 0, environment_spec: Optional[specs.EnvironmentSpec] = None, batch_size: int = 256, prefetch_size: int = 4, observation_augmentation: Optional[types.TensorTransformation] = None, min_replay_size: int = 1000, max_replay_size: int = 1000000, samples_per_insert: Optional[float] = 32.0, n_step: int = 5, num_samples: int = 20, additional_discount: float = 0.99, target_policy_update_period: int = 100, target_critic_update_period: int = 100, variable_update_period: int = 1000, policy_loss_factory: Optional[Callable[[], snt.Module]] = None, max_actor_steps: Optional[int] = None, log_every: float = 10.0, make_observers: Optional[Callable[ [], Sequence[observers_lib.EnvLoopObserver]]] = None): if environment_spec is None: environment_spec = specs.make_environment_spec(environment_factory(False)) self._environment_factory = environment_factory self._network_factory = network_factory self._policy_loss_factory = policy_loss_factory self._environment_spec = environment_spec self._num_actors = num_actors self._num_caches = num_caches self._batch_size = batch_size self._prefetch_size = prefetch_size self._observation_augmentation = observation_augmentation self._min_replay_size = min_replay_size self._max_replay_size = max_replay_size self._samples_per_insert = samples_per_insert self._n_step = n_step self._additional_discount = additional_discount self._num_samples = num_samples self._target_policy_update_period = target_policy_update_period self._target_critic_update_period = target_critic_update_period self._variable_update_period = variable_update_period self._max_actor_steps = max_actor_steps self._log_every = log_every self._make_observers = make_observers def replay(self): """The replay storage.""" if self._samples_per_insert is not None: # Create enough of an error buffer to give a 10% tolerance in rate. samples_per_insert_tolerance = 0.1 * self._samples_per_insert error_buffer = self._min_replay_size * samples_per_insert_tolerance limiter = reverb.rate_limiters.SampleToInsertRatio( min_size_to_sample=self._min_replay_size, samples_per_insert=self._samples_per_insert, error_buffer=error_buffer) else: limiter = reverb.rate_limiters.MinSize(self._min_replay_size) replay_table = reverb.Table( name=adders.DEFAULT_PRIORITY_TABLE, sampler=reverb.selectors.Uniform(), remover=reverb.selectors.Fifo(), max_size=self._max_replay_size, rate_limiter=limiter, signature=adders.NStepTransitionAdder.signature( self._environment_spec)) return [replay_table] def counter(self): return tf2_savers.CheckpointingRunner(counting.Counter(), time_delta_minutes=1, subdirectory='counter') def coordinator(self, counter: counting.Counter, max_actor_steps: int): return lp_utils.StepsLimiter(counter, max_actor_steps) def learner( self, replay: reverb.Client, counter: counting.Counter, ): """The Learning part of the agent.""" act_spec = self._environment_spec.actions obs_spec = self._environment_spec.observations # Create online and target networks. online_networks = self._network_factory(act_spec) target_networks = self._network_factory(act_spec) # Make sure observation network is a Sonnet Module. observation_network = online_networks.get('observation', tf.identity) target_observation_network = target_networks.get('observation', tf.identity) observation_network = tf2_utils.to_sonnet_module(observation_network) target_observation_network = tf2_utils.to_sonnet_module( target_observation_network) # Get embedding spec and create observation network variables. emb_spec = tf2_utils.create_variables(observation_network, [obs_spec]) # Create variables. tf2_utils.create_variables(online_networks['policy'], [emb_spec]) tf2_utils.create_variables(online_networks['critic'], [emb_spec, act_spec]) tf2_utils.create_variables(target_networks['policy'], [emb_spec]) tf2_utils.create_variables(target_networks['critic'], [emb_spec, act_spec]) tf2_utils.create_variables(target_observation_network, [obs_spec]) # The dataset object to learn from. dataset = datasets.make_reverb_dataset(server_address=replay.server_address) dataset = dataset.batch(self._batch_size, drop_remainder=True) if self._observation_augmentation: transform = image_augmentation.make_transform( observation_transform=self._observation_augmentation) dataset = dataset.map( transform, num_parallel_calls=16, deterministic=False) dataset = dataset.prefetch(self._prefetch_size) counter = counting.Counter(counter, 'learner') logger = loggers.make_default_logger( 'learner', time_delta=self._log_every, steps_key='learner_steps') # Create policy loss module if a factory is passed. if self._policy_loss_factory: policy_loss_module = self._policy_loss_factory() else: policy_loss_module = None # Return the learning agent. return learning.DistributionalMPOLearner( policy_network=online_networks['policy'], critic_network=online_networks['critic'], observation_network=observation_network, target_policy_network=target_networks['policy'], target_critic_network=target_networks['critic'], target_observation_network=target_observation_network, discount=self._additional_discount, num_samples=self._num_samples, target_policy_update_period=self._target_policy_update_period, target_critic_update_period=self._target_critic_update_period, policy_loss_module=policy_loss_module, dataset=dataset, counter=counter, logger=logger) def actor( self, replay: reverb.Client, variable_source: acme.VariableSource, counter: counting.Counter, actor_id: int, ) -> acme.EnvironmentLoop: """The actor process.""" action_spec = self._environment_spec.actions observation_spec = self._environment_spec.observations # Create environment and target networks to act with. environment = self._environment_factory(False) agent_networks = self._network_factory(action_spec) # Make sure observation network is defined. observation_network = agent_networks.get('observation', tf.identity) # Create a stochastic behavior policy. behavior_network = snt.Sequential([ observation_network, agent_networks['policy'], networks.StochasticSamplingHead(), ]) # Ensure network variables are created. tf2_utils.create_variables(behavior_network, [observation_spec]) policy_variables = {'policy': behavior_network.variables} # Create the variable client responsible for keeping the actor up-to-date. variable_client = tf2_variable_utils.VariableClient( variable_source, policy_variables, update_period=self._variable_update_period) # Make sure not to use a random policy after checkpoint restoration by # assigning variables before running the environment loop. variable_client.update_and_wait() # Component to add things into replay. adder = adders.NStepTransitionAdder( client=replay, n_step=self._n_step, discount=self._additional_discount) # Create the agent. actor = actors.FeedForwardActor( policy_network=behavior_network, adder=adder, variable_client=variable_client) # Create logger and counter; only the first actor stores logs to bigtable. save_data = actor_id == 0 counter = counting.Counter(counter, 'actor') logger = loggers.make_default_logger( 'actor', save_data=save_data, time_delta=self._log_every, steps_key='actor_steps') observers = self._make_observers() if self._make_observers else () # Create the run loop and return it. return acme.EnvironmentLoop( environment, actor, counter, logger, observers=observers) def evaluator( self, variable_source: acme.VariableSource, counter: counting.Counter, ): """The evaluation process.""" action_spec = self._environment_spec.actions observation_spec = self._environment_spec.observations # Create environment and target networks to act with. environment = self._environment_factory(True) agent_networks = self._network_factory(action_spec) # Make sure observation network is defined. observation_network = agent_networks.get('observation', tf.identity) # Create a stochastic behavior policy. evaluator_network = snt.Sequential([ observation_network, agent_networks['policy'], networks.StochasticMeanHead(), ]) # Ensure network variables are created. tf2_utils.create_variables(evaluator_network, [observation_spec]) policy_variables = {'policy': evaluator_network.variables} # Create the variable client responsible for keeping the actor up-to-date. variable_client = tf2_variable_utils.VariableClient( variable_source, policy_variables, update_period=self._variable_update_period) # Make sure not to evaluate a random actor by assigning variables before # running the environment loop. variable_client.update_and_wait() # Create the agent. evaluator = actors.FeedForwardActor( policy_network=evaluator_network, variable_client=variable_client) # Create logger and counter. counter = counting.Counter(counter, 'evaluator') logger = loggers.make_default_logger( 'evaluator', time_delta=self._log_every, steps_key='evaluator_steps') observers = self._make_observers() if self._make_observers else () # Create the run loop and return it. return acme.EnvironmentLoop( environment, evaluator, counter, logger, observers=observers) def build(self, name='dmpo'): """Build the distributed agent topology.""" program = lp.Program(name=name) with program.group('replay'): replay = program.add_node(lp.ReverbNode(self.replay)) with program.group('counter'): counter = program.add_node(lp.CourierNode(self.counter)) if self._max_actor_steps: _ = program.add_node( lp.CourierNode(self.coordinator, counter, self._max_actor_steps)) with program.group('learner'): learner = program.add_node( lp.CourierNode(self.learner, replay, counter)) with program.group('evaluator'): program.add_node( lp.CourierNode(self.evaluator, learner, counter)) if not self._num_caches: # Use our learner as a single variable source. sources = [learner] else: with program.group('cacher'): # Create a set of learner caches. sources = [] for _ in range(self._num_caches): cacher = program.add_node( lp.CacherNode( learner, refresh_interval_ms=2000, stale_after_ms=4000)) sources.append(cacher) with program.group('actor'): # Add actors which pull round-robin from our variable sources. for actor_id in range(self._num_actors): source = sources[actor_id % len(sources)] program.add_node( lp.CourierNode(self.actor, replay, source, counter, actor_id)) return program
acme-master
acme/agents/tf/dmpo/agent_distributed.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Implementations of a distributional MPO agent.""" from acme.agents.tf.dmpo.agent import DistributionalMPO from acme.agents.tf.dmpo.agent_distributed import DistributedDistributionalMPO from acme.agents.tf.dmpo.learning import DistributionalMPOLearner
acme-master
acme/agents/tf/dmpo/__init__.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Tests for the distributional MPO agent.""" from typing import Dict, Sequence import acme from acme import specs from acme.agents.tf import dmpo from acme.testing import fakes from acme.tf import networks import numpy as np import sonnet as snt from absl.testing import absltest def make_networks( action_spec: specs.Array, policy_layer_sizes: Sequence[int] = (300, 200), critic_layer_sizes: Sequence[int] = (400, 300), ) -> Dict[str, snt.Module]: """Creates networks used by the agent.""" num_dimensions = np.prod(action_spec.shape, dtype=int) critic_layer_sizes = list(critic_layer_sizes) policy_network = snt.Sequential([ networks.LayerNormMLP(policy_layer_sizes), networks.MultivariateNormalDiagHead(num_dimensions), ]) # The multiplexer concatenates the (maybe transformed) observations/actions. critic_network = snt.Sequential([ networks.CriticMultiplexer( critic_network=networks.LayerNormMLP(critic_layer_sizes)), networks.DiscreteValuedHead(0., 1., 10), ]) return { 'policy': policy_network, 'critic': critic_network, } class DMPOTest(absltest.TestCase): def test_dmpo(self): # Create a fake environment to test with. environment = fakes.ContinuousEnvironment(episode_length=10) spec = specs.make_environment_spec(environment) # Create networks. agent_networks = make_networks(spec.actions) # Construct the agent. agent = dmpo.DistributionalMPO( spec, policy_network=agent_networks['policy'], critic_network=agent_networks['critic'], batch_size=10, samples_per_insert=2, min_replay_size=10) # Try running the environment loop. We have no assertions here because all # we care about is that the agent runs without raising any errors. loop = acme.EnvironmentLoop(environment, agent) loop.run(num_episodes=2) if __name__ == '__main__': absltest.main()
acme-master
acme/agents/tf/dmpo/agent_test.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Distributional MPO agent implementation.""" import copy from typing import Optional from acme import datasets from acme import specs from acme import types from acme.adders import reverb as adders from acme.agents import agent from acme.agents.tf import actors from acme.agents.tf.dmpo import learning from acme.tf import networks from acme.tf import utils as tf2_utils from acme.utils import counting from acme.utils import loggers import reverb import sonnet as snt import tensorflow as tf class DistributionalMPO(agent.Agent): """Distributional MPO Agent. This implements a single-process distributional MPO agent. This is an actor-critic algorithm that generates data via a behavior policy, inserts N-step transitions into a replay buffer, and periodically updates the policy (and as a result the behavior) by sampling uniformly from this buffer. This agent distinguishes itself from the MPO agent by using a distributional critic (state-action value approximator). """ def __init__(self, environment_spec: specs.EnvironmentSpec, policy_network: snt.Module, critic_network: snt.Module, observation_network: types.TensorTransformation = tf.identity, discount: float = 0.99, batch_size: int = 256, prefetch_size: int = 4, target_policy_update_period: int = 100, target_critic_update_period: int = 100, min_replay_size: int = 1000, max_replay_size: int = 1000000, samples_per_insert: float = 32.0, policy_loss_module: Optional[snt.Module] = None, policy_optimizer: Optional[snt.Optimizer] = None, critic_optimizer: Optional[snt.Optimizer] = None, n_step: int = 5, num_samples: int = 20, clipping: bool = True, logger: Optional[loggers.Logger] = None, counter: Optional[counting.Counter] = None, checkpoint: bool = True, replay_table_name: str = adders.DEFAULT_PRIORITY_TABLE): """Initialize the agent. Args: environment_spec: description of the actions, observations, etc. policy_network: the online (optimized) policy. critic_network: the online critic. observation_network: optional network to transform the observations before they are fed into any network. discount: discount to use for TD updates. batch_size: batch size for updates. prefetch_size: size to prefetch from replay. target_policy_update_period: number of updates to perform before updating the target policy network. target_critic_update_period: number of updates to perform before updating the target critic network. min_replay_size: minimum replay size before updating. max_replay_size: maximum replay size. samples_per_insert: number of samples to take from replay for every insert that is made. policy_loss_module: configured MPO loss function for the policy optimization; defaults to sensible values on the control suite. See `acme/tf/losses/mpo.py` for more details. policy_optimizer: optimizer to be used on the policy. critic_optimizer: optimizer to be used on the critic. n_step: number of steps to squash into a single transition. num_samples: number of actions to sample when doing a Monte Carlo integration with respect to the policy. clipping: whether to clip gradients by global norm. logger: logging object used to write to logs. counter: counter object used to keep track of steps. checkpoint: boolean indicating whether to checkpoint the learner. replay_table_name: string indicating what name to give the replay table. """ # Create a replay server to add data to. replay_table = reverb.Table( name=adders.DEFAULT_PRIORITY_TABLE, sampler=reverb.selectors.Uniform(), remover=reverb.selectors.Fifo(), max_size=max_replay_size, rate_limiter=reverb.rate_limiters.MinSize(min_size_to_sample=1), signature=adders.NStepTransitionAdder.signature(environment_spec)) self._server = reverb.Server([replay_table], port=None) # The adder is used to insert observations into replay. address = f'localhost:{self._server.port}' adder = adders.NStepTransitionAdder( client=reverb.Client(address), n_step=n_step, discount=discount) # The dataset object to learn from. dataset = datasets.make_reverb_dataset( table=replay_table_name, server_address=address, batch_size=batch_size, prefetch_size=prefetch_size) # Make sure observation network is a Sonnet Module. observation_network = tf2_utils.to_sonnet_module(observation_network) # Create target networks before creating online/target network variables. target_policy_network = copy.deepcopy(policy_network) target_critic_network = copy.deepcopy(critic_network) target_observation_network = copy.deepcopy(observation_network) # Get observation and action specs. act_spec = environment_spec.actions obs_spec = environment_spec.observations emb_spec = tf2_utils.create_variables(observation_network, [obs_spec]) # Create the behavior policy. behavior_network = snt.Sequential([ observation_network, policy_network, networks.StochasticSamplingHead(), ]) # Create variables. tf2_utils.create_variables(policy_network, [emb_spec]) tf2_utils.create_variables(critic_network, [emb_spec, act_spec]) tf2_utils.create_variables(target_policy_network, [emb_spec]) tf2_utils.create_variables(target_critic_network, [emb_spec, act_spec]) tf2_utils.create_variables(target_observation_network, [obs_spec]) # Create the actor which defines how we take actions. actor = actors.FeedForwardActor( policy_network=behavior_network, adder=adder) # Create optimizers. policy_optimizer = policy_optimizer or snt.optimizers.Adam(1e-4) critic_optimizer = critic_optimizer or snt.optimizers.Adam(1e-4) # The learner updates the parameters (and initializes them). learner = learning.DistributionalMPOLearner( policy_network=policy_network, critic_network=critic_network, observation_network=observation_network, target_policy_network=target_policy_network, target_critic_network=target_critic_network, target_observation_network=target_observation_network, policy_loss_module=policy_loss_module, policy_optimizer=policy_optimizer, critic_optimizer=critic_optimizer, clipping=clipping, discount=discount, num_samples=num_samples, target_policy_update_period=target_policy_update_period, target_critic_update_period=target_critic_update_period, dataset=dataset, logger=logger, counter=counter, checkpoint=checkpoint) super().__init__( actor=actor, learner=learner, min_observations=max(batch_size, min_replay_size), observations_per_step=float(batch_size) / samples_per_insert)
acme-master
acme/agents/tf/dmpo/agent.py
# Copyright 2018 DeepMind Technologies Limited. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Distributional MPO learner implementation.""" import time from typing import List, Optional import acme from acme import types from acme.tf import losses from acme.tf import networks from acme.tf import savers as tf2_savers from acme.tf import utils as tf2_utils from acme.utils import counting from acme.utils import loggers import numpy as np import sonnet as snt import tensorflow as tf class DistributionalMPOLearner(acme.Learner): """Distributional MPO learner.""" def __init__( self, policy_network: snt.Module, critic_network: snt.Module, target_policy_network: snt.Module, target_critic_network: snt.Module, discount: float, num_samples: int, target_policy_update_period: int, target_critic_update_period: int, dataset: tf.data.Dataset, observation_network: types.TensorTransformation = tf.identity, target_observation_network: types.TensorTransformation = tf.identity, policy_loss_module: Optional[snt.Module] = None, policy_optimizer: Optional[snt.Optimizer] = None, critic_optimizer: Optional[snt.Optimizer] = None, dual_optimizer: Optional[snt.Optimizer] = None, clipping: bool = True, counter: Optional[counting.Counter] = None, logger: Optional[loggers.Logger] = None, checkpoint: bool = True, ): # Store online and target networks. self._policy_network = policy_network self._critic_network = critic_network self._target_policy_network = target_policy_network self._target_critic_network = target_critic_network # Make sure observation networks are snt.Module's so they have variables. self._observation_network = tf2_utils.to_sonnet_module(observation_network) self._target_observation_network = tf2_utils.to_sonnet_module( target_observation_network) # General learner book-keeping and loggers. self._counter = counter or counting.Counter() self._logger = logger or loggers.make_default_logger('learner') # Other learner parameters. self._discount = discount self._num_samples = num_samples self._clipping = clipping # Necessary to track when to update target networks. self._num_steps = tf.Variable(0, dtype=tf.int32) self._target_policy_update_period = target_policy_update_period self._target_critic_update_period = target_critic_update_period # Batch dataset and create iterator. # TODO(b/155086959): Fix type stubs and remove. self._iterator = iter(dataset) # pytype: disable=wrong-arg-types self._policy_loss_module = policy_loss_module or losses.MPO( epsilon=1e-1, epsilon_penalty=1e-3, epsilon_mean=2.5e-3, epsilon_stddev=1e-6, init_log_temperature=10., init_log_alpha_mean=10., init_log_alpha_stddev=1000.) # Create the optimizers. self._critic_optimizer = critic_optimizer or snt.optimizers.Adam(1e-4) self._policy_optimizer = policy_optimizer or snt.optimizers.Adam(1e-4) self._dual_optimizer = dual_optimizer or snt.optimizers.Adam(1e-2) # Expose the variables. policy_network_to_expose = snt.Sequential( [self._target_observation_network, self._target_policy_network]) self._variables = { 'critic': self._target_critic_network.variables, 'policy': policy_network_to_expose.variables, } # Create a checkpointer and snapshotter object. self._checkpointer = None self._snapshotter = None if checkpoint: self._checkpointer = tf2_savers.Checkpointer( subdirectory='dmpo_learner', objects_to_save={ 'counter': self._counter, 'policy': self._policy_network, 'critic': self._critic_network, 'observation': self._observation_network, 'target_policy': self._target_policy_network, 'target_critic': self._target_critic_network, 'target_observation': self._target_observation_network, 'policy_optimizer': self._policy_optimizer, 'critic_optimizer': self._critic_optimizer, 'dual_optimizer': self._dual_optimizer, 'policy_loss_module': self._policy_loss_module, 'num_steps': self._num_steps, }) self._snapshotter = tf2_savers.Snapshotter( objects_to_save={ 'policy': snt.Sequential([ self._target_observation_network, self._target_policy_network ]), }) # Do not record timestamps until after the first learning step is done. # This is to avoid including the time it takes for actors to come online and # fill the replay buffer. self._timestamp = None @tf.function def _step(self) -> types.NestedTensor: # Update target network. online_policy_variables = self._policy_network.variables target_policy_variables = self._target_policy_network.variables online_critic_variables = ( *self._observation_network.variables, *self._critic_network.variables, ) target_critic_variables = ( *self._target_observation_network.variables, *self._target_critic_network.variables, ) # Make online policy -> target policy network update ops. if tf.math.mod(self._num_steps, self._target_policy_update_period) == 0: for src, dest in zip(online_policy_variables, target_policy_variables): dest.assign(src) # Make online critic -> target critic network update ops. if tf.math.mod(self._num_steps, self._target_critic_update_period) == 0: for src, dest in zip(online_critic_variables, target_critic_variables): dest.assign(src) self._num_steps.assign_add(1) # Get data from replay (dropping extras if any). Note there is no # extra data here because we do not insert any into Reverb. inputs = next(self._iterator) transitions: types.Transition = inputs.data # Get batch size and scalar dtype. batch_size = transitions.reward.shape[0] # Cast the additional discount to match the environment discount dtype. discount = tf.cast(self._discount, dtype=transitions.discount.dtype) with tf.GradientTape(persistent=True) as tape: # Maybe transform the observation before feeding into policy and critic. # Transforming the observations this way at the start of the learning # step effectively means that the policy and critic share observation # network weights. o_tm1 = self._observation_network(transitions.observation) # This stop_gradient prevents gradients to propagate into the target # observation network. In addition, since the online policy network is # evaluated at o_t, this also means the policy loss does not influence # the observation network training. o_t = tf.stop_gradient( self._target_observation_network(transitions.next_observation)) # Get online and target action distributions from policy networks. online_action_distribution = self._policy_network(o_t) target_action_distribution = self._target_policy_network(o_t) # Sample actions to evaluate policy; of size [N, B, ...]. sampled_actions = target_action_distribution.sample(self._num_samples) # Tile embedded observations to feed into the target critic network. # Note: this is more efficient than tiling before the embedding layer. tiled_o_t = tf2_utils.tile_tensor(o_t, self._num_samples) # [N, B, ...] # Compute target-estimated distributional value of sampled actions at o_t. sampled_q_t_distributions = self._target_critic_network( # Merge batch dimensions; to shape [N*B, ...]. snt.merge_leading_dims(tiled_o_t, num_dims=2), snt.merge_leading_dims(sampled_actions, num_dims=2)) # Compute average logits by first reshaping them and normalizing them # across atoms. new_shape = [self._num_samples, batch_size, -1] # [N, B, A] sampled_logits = tf.reshape(sampled_q_t_distributions.logits, new_shape) sampled_logprobs = tf.math.log_softmax(sampled_logits, axis=-1) averaged_logits = tf.reduce_logsumexp(sampled_logprobs, axis=0) # Construct the expected distributional value for bootstrapping. q_t_distribution = networks.DiscreteValuedDistribution( values=sampled_q_t_distributions.values, logits=averaged_logits) # Compute online critic value distribution of a_tm1 in state o_tm1. q_tm1_distribution = self._critic_network(o_tm1, transitions.action) # Compute critic distributional loss. critic_loss = losses.categorical(q_tm1_distribution, transitions.reward, discount * transitions.discount, q_t_distribution) critic_loss = tf.reduce_mean(critic_loss) # Compute Q-values of sampled actions and reshape to [N, B]. sampled_q_values = sampled_q_t_distributions.mean() sampled_q_values = tf.reshape(sampled_q_values, (self._num_samples, -1)) # Compute MPO policy loss. policy_loss, policy_stats = self._policy_loss_module( online_action_distribution=online_action_distribution, target_action_distribution=target_action_distribution, actions=sampled_actions, q_values=sampled_q_values) # For clarity, explicitly define which variables are trained by which loss. critic_trainable_variables = ( # In this agent, the critic loss trains the observation network. self._observation_network.trainable_variables + self._critic_network.trainable_variables) policy_trainable_variables = self._policy_network.trainable_variables # The following are the MPO dual variables, stored in the loss module. dual_trainable_variables = self._policy_loss_module.trainable_variables # Compute gradients. critic_gradients = tape.gradient(critic_loss, critic_trainable_variables) policy_gradients, dual_gradients = tape.gradient( policy_loss, (policy_trainable_variables, dual_trainable_variables)) # Delete the tape manually because of the persistent=True flag. del tape # Maybe clip gradients. if self._clipping: policy_gradients = tuple(tf.clip_by_global_norm(policy_gradients, 40.)[0]) critic_gradients = tuple(tf.clip_by_global_norm(critic_gradients, 40.)[0]) # Apply gradients. self._critic_optimizer.apply(critic_gradients, critic_trainable_variables) self._policy_optimizer.apply(policy_gradients, policy_trainable_variables) self._dual_optimizer.apply(dual_gradients, dual_trainable_variables) # Losses to track. fetches = { 'critic_loss': critic_loss, 'policy_loss': policy_loss, } fetches.update(policy_stats) # Log MPO stats. return fetches def step(self): # Run the learning step. fetches = self._step() # Compute elapsed time. timestamp = time.time() elapsed_time = timestamp - self._timestamp if self._timestamp else 0 self._timestamp = timestamp # Update our counts and record it. counts = self._counter.increment(steps=1, walltime=elapsed_time) fetches.update(counts) # Checkpoint and attempt to write the logs. if self._checkpointer is not None: self._checkpointer.save() if self._snapshotter is not None: self._snapshotter.save() self._logger.write(fetches) def get_variables(self, names: List[str]) -> List[List[np.ndarray]]: return [tf2_utils.to_numpy(self._variables[name]) for name in names]
acme-master
acme/agents/tf/dmpo/learning.py