#!/usr/bin/env python3 # -*- coding: utf-8 -*- import os, glob, re, torch, numpy as np from typing import List, Tuple from safetensors.torch import load_file import gradio as gr import plotly.graph_objects as go import torch.nn as nn from huggingface_hub import hf_hub_download, list_repo_files # ========================= # Config # ========================= LOCAL_FILE_DEFAULT = "assets/scene0073_00.safetensors" # local safetensors file PM_KEY_DEFAULT = "point_map" TOPK_VIEWS_DEFAULT = 3 VOXEL_SIZE = 0.02 DOWNSAMPLE_N_MAX = 600_000 POINT_SIZE_MIN = 1.2 POINT_SIZE_MAX = 2.0 CLR_RED = "rgba(230,40,40,0.98)" BG_COLOR = "#f7f9fb" GRID_COLOR = "#e6ecf2" BOX_COLOR = "rgba(80,80,80,0.6)" TOP_VIEW_IMAGE_PATH = "assets/scene0073_00.png" DEFAULT_CAM = dict( eye=dict(x=1.35, y=1.35, z=0.95), up=dict(x=0, y=0, z=1), center=dict(x=0, y=0, z=0), projection=dict(type="perspective"), ) def _merge_safetensors_dicts(paths: List[str]): merged = {} for p in paths: sd = load_file(p, device="cpu") merged.update(sd) return merged def _local_all_under(path: str) -> List[str]: out = [] if os.path.isfile(path): return [path] for root, _, files in os.walk(path): for f in files: out.append(os.path.join(root, f)) return sorted(out) # ========================= # Load pretrained (your existing loader) # ========================= def load_pretrain( model: torch.nn.Module, ckpt_path: str, # e.g. "assets/ckpt_100.pth" or "assets/model.safetensors" repo_id: str = "MatchLab/poma3d-demo", revision: str = "main", allow_local_fallback: bool = True, ): if allow_local_fallback and (os.path.isfile(ckpt_path) or os.path.isdir(ckpt_path)): print(f"📂 Using local checkpoint(s): {ckpt_path}") local_files = _local_all_under(ckpt_path) else: # 2) REMOTE: resolve file list from Space print(f"📦 Resolving from HF Space: {repo_id}/{ckpt_path} (rev={revision})") files = list_repo_files(repo_id=repo_id, repo_type='model', revision=revision) # Exact file hit? if ckpt_path in files: to_fetch = [ckpt_path] else: # Treat ckpt_path as a folder prefix (ensure trailing slash for matching) prefix = ckpt_path if ckpt_path.endswith("/") else ckpt_path + "/" to_fetch = [f for f in files if f.startswith(prefix)] if not to_fetch: preview = "\n".join(files[:100]) raise FileNotFoundError( f"'{ckpt_path}' not found in Space '{repo_id}' (rev='{revision}').\n" f"Files present (first 100):\n{preview}" ) # Download all matching files locally local_files = [] for rel in to_fetch: lp = hf_hub_download(repo_id=repo_id, filename=rel, repo_type='model', revision=revision) local_files.append(lp) local_files.sort() # Filter by types we know how to load safes = [p for p in local_files if p.endswith(".safetensors")] pths = [p for p in local_files if re.search(r"\.(?:pth|pt)$", p)] if safes: print(f"🧩 Found {len(safes)} .safetensors shard(s); merging…") state = _merge_safetensors_dicts(safes) elif pths: # pick the largest .pth/.pt to avoid optimizer/state variants pths_sorted = sorted(pths, key=lambda p: os.path.getsize(p), reverse=True) pick = pths_sorted[0] print(f"🧩 Using .pth/.pt: {os.path.basename(pick)} (largest of {len(pths)} candidates)") state = torch.load(pick, map_location="cpu") # strip common prefixes if isinstance(state, dict) and any(k.startswith(("model.", "target_model.")) for k in state.keys()): state = { (k.split(".", 1)[1] if k.startswith(("model.", "target_model.")) else k): v for k, v in state.items() } # nested 'state_dict' if isinstance(state, dict) and "state_dict" in state and isinstance(state["state_dict"], dict): state = state["state_dict"] else: raise FileNotFoundError( "No loadable checkpoint found. Expecting one or more of: " ".safetensors or .pth/.pt under the given path." ) # Load into model result = model.load_state_dict(state, strict=False) # Report weight_keys = set(state.keys()) if isinstance(state, dict) else set() model_keys = set(model.state_dict().keys()) loaded_keys = model_keys.intersection(weight_keys) print("✅ Weights loaded") print(f" • Loaded keys: {len(loaded_keys)}") print(f" • Missing keys: {len(result.missing_keys)}") print(f" • Unexpected keys: {len(result.unexpected_keys)}") return result # ========================= # Representation model (fg-clip-base + LoRA) # ========================= def build_model(device: torch.device): from transformers import AutoModelForCausalLM, AutoTokenizer, AutoConfig from peft import LoraConfig, get_peft_model class RepModel(torch.nn.Module): def __init__(self, model_root="qihoo360/fg-clip-base"): super().__init__() lora_cfg = LoraConfig( r=32, lora_alpha=64, target_modules=["q_proj","k_proj","v_proj","fc1","fc2"], lora_dropout=0.05, bias="none", task_type="FEATURE_EXTRACTION" ) cfg = AutoConfig.from_pretrained(model_root, trust_remote_code=True) base = AutoModelForCausalLM.from_config(cfg, trust_remote_code=True) self.target_model = get_peft_model(base, lora_cfg) self.tokenizer = AutoTokenizer.from_pretrained(model_root, trust_remote_code=True, use_fast=True) @torch.no_grad() def get_text_feature(self, texts, device): tok = self.tokenizer(texts, padding="max_length", truncation=True, max_length=248, return_tensors="pt").to(device) feats = self.target_model.get_text_features(tok["input_ids"], walk_short_pos=False) feats = torch.nn.functional.normalize(feats.float(), dim=-1) return feats @torch.no_grad() def get_image_feature(self, pm_batched): feats = self.target_model.get_image_features(pm_batched) feats = torch.nn.functional.normalize(feats.float(), dim=-1) return feats m = RepModel().to(device).eval() print("Using fg-clip-base RepModel.") return m # ========================= # Data loading & helpers # ========================= def load_scene_local(path: str, pm_key: str = PM_KEY_DEFAULT) -> torch.Tensor: if not os.path.exists(path): raise FileNotFoundError(f"Local file not found: {path}") sd = load_file(path) if pm_key not in sd: raise KeyError(f"Key '{pm_key}' not found in {list(sd.keys())}") pm = sd[pm_key] # (V,H,W,3) if pm.dim() != 4 or pm.shape[-1] != 3: raise ValueError(f"Invalid shape {tuple(pm.shape)}, expected (V,H,W,3)") return pm.permute(0, 3, 1, 2).contiguous() # -> (V,3,H,W) def _xyz_to_numpy(xyz: torch.Tensor) -> np.ndarray: pts = xyz.permute(1, 2, 0).reshape(-1, 3).cpu().numpy().astype(np.float32) mask = np.isfinite(pts).all(axis=1) return pts[mask] def stack_views(pm: torch.Tensor) -> Tuple[np.ndarray, np.ndarray]: pts_all, vid_all = [], [] for v in range(pm.shape[0]): pts = _xyz_to_numpy(pm[v]) if pts.size == 0: continue pts_all.append(pts) vid_all.append(np.full((pts.shape[0],), v, dtype=np.int32)) pts_all = np.concatenate(pts_all, axis=0) vid_all = np.concatenate(vid_all, axis=0) return pts_all, vid_all def voxel_downsample_with_ids(pts, vids, voxel: float): if pts.shape[0] == 0: return pts, vids grid = np.floor(pts / voxel).astype(np.int64) key = np.core.records.fromarrays(grid.T, names="x,y,z", formats="i8,i8,i8") _, uniq_idx = np.unique(key, return_index=True) return pts[uniq_idx], vids[uniq_idx] def hard_cap(pts, vids, cap: int): N = pts.shape[0] if N <= cap: return pts, vids idx = np.random.choice(N, size=cap, replace=False) return pts[idx], vids[idx] def adaptive_point_size(n: int) -> float: ps = 2.4 * (150_000 / max(n, 10)) ** 0.25 return float(np.clip(ps, POINT_SIZE_MIN, POINT_SIZE_MAX)) def scene_bbox(pts: np.ndarray): mn, mx = pts.min(axis=0), pts.max(axis=0) x0,y0,z0 = mn; x1,y1,z1 = mx corners = np.array([ [x0,y0,z0],[x1,y0,z0],[x1,y1,z0],[x0,y1,z0], [x0,y0,z1],[x1,y0,z1],[x1,y1,z1],[x0,y1,z1] ]) edges = [(0,1),(1,2),(2,3),(3,0),(4,5),(5,6),(6,7),(7,4),(0,4),(1,5),(2,6),(3,7)] xs,ys,zs=[],[],[] for a,b in edges: xs += [corners[a,0], corners[b,0], None] ys += [corners[a,1], corners[b,1], None] zs += [corners[a,2], corners[b,2], None] return xs,ys,zs @torch.no_grad() def rank_views_for_text(model, text, pm, device, topk: int): img_feats = model.get_image_feature(pm.float().to(device)) txt_feat = model.get_text_feature([text], device=device)[0] sims = torch.matmul(img_feats, txt_feat) order = torch.argsort(sims, descending=True)[:max(1, int(topk))] return order.tolist() # ========================= # Visualization # ========================= def depth_values(pts: np.ndarray) -> np.ndarray: z = pts[:, 2] z_min, z_max = z.min(), z.max() return (z - z_min) / (z_max - z_min + 1e-9) def base_figure_gray_depth(pts: np.ndarray, point_size: float, camera=DEFAULT_CAM) -> go.Figure: depth = depth_values(pts) fig = go.Figure(go.Scatter3d( x=pts[:,0], y=pts[:,1], z=pts[:,2], mode="markers", marker=dict(size=point_size, color=depth, colorscale="Greys", reversescale=True, opacity=0.50), hoverinfo="skip" )) bx,by,bz = scene_bbox(pts) fig.add_trace(go.Scatter3d(x=bx,y=by,z=bz,mode="lines",line=dict(color=BOX_COLOR,width=2),hoverinfo="skip")) fig.update_layout(scene=dict(aspectmode="data",camera=camera), margin=dict(l=0,r=0,b=0,t=0), paper_bgcolor=BG_COLOR, showlegend=False) return fig def highlight_views_3d(pts, view_ids, selected, point_size, camera=DEFAULT_CAM): depth = depth_values(pts) colors = np.stack([depth, depth, depth], axis=1) if selected: sel_mask = np.isin(view_ids, np.array(selected, dtype=np.int32)) colors[sel_mask] = np.array([1, 0, 0]) fig = go.Figure(go.Scatter3d( x=pts[:,0], y=pts[:,1], z=pts[:,2], mode="markers", marker=dict(size=point_size, color=[f"rgb({int(r*255)},{int(g*255)},{int(b*255)})" for r,g,b in colors], opacity=0.98), hoverinfo="skip" )) bx,by,bz = scene_bbox(pts) fig.add_trace(go.Scatter3d(x=bx,y=by,z=bz,mode="lines", line=dict(color=BOX_COLOR,width=2),hoverinfo="skip")) fig.update_layout(scene=dict(aspectmode="data",camera=camera), margin=dict(l=0,r=0,b=0,t=0), paper_bgcolor=BG_COLOR, showlegend=False) return fig # ========================= # App setup # ========================= device = 'cpu' model = build_model(device) load_pretrain(model, "assets/ckpt_100.pth") with gr.Blocks( title="POMA-3D: Text-conditioned 3D Scene Visualization", css="#plot3d, #img_ref {height: 450px !important;}" ) as demo: gr.Markdown("### POMA-3D: The Point Map Way to 3D Scene Understanding - Embodied Localization Demo\n" "Enter agent's situation text and choose **Top-K**; the most relevant views will turn **red**.") with gr.Row(): text_in = gr.Textbox(label="Text query", value="I am sleeping on the bed.", scale=4) topk_in = gr.Number(label="Top-K views", value=TOPK_VIEWS_DEFAULT, precision=0, minimum=1, maximum=12) submit_btn = gr.Button("Locate", variant="primary") with gr.Row(equal_height=True): with gr.Column(scale=1, min_width=500): plot3d = gr.Plot(label="3D Point Cloud (rotatable)", elem_id="plot3d") with gr.Column(scale=1, min_width=500): img_ref = gr.Image(label="Top-Down Reference View", value=TOP_VIEW_IMAGE_PATH, elem_id="img_ref") status = gr.Markdown() pm_state = gr.State(None) pts_state = gr.State(None) vids_state = gr.State(None) # Load scene automatically from LOCAL_FILE_DEFAULT def on_load(): pm = load_scene_local(LOCAL_FILE_DEFAULT) pts_all, vids_all = stack_views(pm) pts_vx, vids_vx = voxel_downsample_with_ids(pts_all, vids_all, VOXEL_SIZE) pts_vx, vids_vx = hard_cap(pts_vx, vids_vx, DOWNSAMPLE_N_MAX) ps = adaptive_point_size(pts_vx.shape[0]) fig3d = base_figure_gray_depth(pts_vx, ps, camera=DEFAULT_CAM) msg = f"✅ Loaded {os.path.basename(LOCAL_FILE_DEFAULT)} | Views: {pm.shape[0]} | Points: {pts_vx.shape[0]:,}" return fig3d, TOP_VIEW_IMAGE_PATH, msg, pm, pts_vx, vids_vx def on_submit(text, topk, pm, pts_vx, vids_vx): if pm is None: return gr.update(), TOP_VIEW_IMAGE_PATH, "⚠️ Scene not loaded yet." k = int(max(1, min(12, int(topk)))) if topk else TOPK_VIEWS_DEFAULT top_views = rank_views_for_text(model, text, pm, device, topk=k) ps = adaptive_point_size(pts_vx.shape[0]) fig = highlight_views_3d(pts_vx, vids_vx, top_views, ps, camera=DEFAULT_CAM) msg = f"Highlighted views (top-{k}): {top_views}" return fig, TOP_VIEW_IMAGE_PATH, msg demo.load(on_load, inputs=[], outputs=[plot3d, img_ref, status, pm_state, pts_state, vids_state]) submit_btn.click(on_submit, inputs=[text_in, topk_in, pm_state, pts_state, vids_state], outputs=[plot3d, img_ref, status]) if __name__ == "__main__": demo.launch()