simson_modeling/create_augmented_dataset.py

import pandas as pd
from tqdm import tqdm
from rdkit import Chem, RDLogger
from datasets import load_dataset
from multiprocessing import Pool, cpu_count
import os

# Suppress RDKit console output for cleaner logs
RDLogger.DisableLog('rdApp.*')

class SmilesEnumerator:
    """
    A simple class to encapsulate the SMILES randomization logic.
    Needed for multiprocessing to work correctly with instance methods.
    """
    def randomize_smiles(self, smiles):
        """Generates a randomized SMILES string."""
        try:
            mol = Chem.MolFromSmiles(smiles)
            # Return a randomized, non-canonical SMILES string
            return Chem.MolToSmiles(mol, doRandom=True, canonical=False) if mol else smiles
        except:
            # If RDKit fails, return the original smiles string
            return smiles

def create_augmented_pair(smiles_string):
    """
    Worker function: takes one SMILES string and returns a tuple
    containing two different randomized versions of it.
    """
    enumerator = SmilesEnumerator()
    smiles_1 = enumerator.randomize_smiles(smiles_string)
    smiles_2 = enumerator.randomize_smiles(smiles_string)
    return smiles_1, smiles_2

def main():
    """
    Main function to run the parallel data preprocessing.
    """
    # --- Configuration ---
    # Load your desired dataset from Hugging Face
    dataset_name = 'jablonkagroup/pubchem-smiles-molecular-formula'
    # Specify the column containing the SMILES strings
    smiles_column_name = 'smiles'
    # Set the output file path
    output_path = 'data/pubchem_2_epoch_50M'

    # --- Data Loading ---
    print(f"Loading dataset '{dataset_name}'...")
    # Use streaming to avoid downloading the whole dataset if you only need a subset
    #dataset = pd.read_csv('/home/jovyan/simson_training_bolgov/data/PI1M_v2.csv')
    dataset = load_dataset(dataset_name)['train'].select(range(50_000_000))
    # Take the desired number of samples
    smiles_list = dataset[smiles_column_name]
    print(f"Successfully fetched {len(smiles_list)} SMILES strings.")

    # --- Parallel Processing ---
    # Use all available CPU cores for maximum speed
    num_workers = cpu_count()
    print(f"Starting SMILES augmentation with {num_workers} worker processes...")

    # A Pool of processes will run the `create_augmented_pair` function in parallel
    with Pool(num_workers) as p:
        # Use tqdm to create a progress bar for the mapping operation
        results = list(tqdm(p.imap(create_augmented_pair, smiles_list), total=len(smiles_list), desc="Augmenting Pairs"))

    # --- Saving Data ---
    print("Processing complete. Converting to DataFrame...")
    # Convert the list of tuples into a pandas DataFrame
    df = pd.DataFrame(results, columns=['smiles_1', 'smiles_2'])

    # Ensure the output directory exists
    os.makedirs(os.path.dirname(output_path), exist_ok=True)
    
    print(f"Saving augmented pairs to '{output_path}'...")
    # Save the DataFrame to a Parquet file for efficient storage and loading
    df.to_parquet(output_path)
    
    print("All done. Your pre-computed dataset is ready!")

if __name__ == '__main__':
    main()