"""Compose reason (25f) + RGB (24f) training videos for coaf_dataset_24_25.""" import argparse import csv import json from pathlib import Path import cv2 import imageio import numpy as np DATASET_ROOT = Path("/project/llmsvgen/sunkai/robomaster_3d/Casual_CoAF/coaf_dataset_24_25") RAW_ROOT = DATASET_ROOT / "raw" MOD_ROOT = DATASET_ROOT / "modalities" COMPOSED_ROOT = DATASET_ROOT / "composed" REASON_FRAMES = 25 RGB_FRAMES = 24 VERSION_CONFIGS = { "v1_pose_rgb": {"modalities": ["pose"]}, "v2_flow_rgb": {"modalities": ["flow"]}, "v3_pose_flow_rgb": {"modalities": ["pose", "flow"]}, "v4_depth_rgb": {"modalities": ["depth"]}, "v5_pose_depth_rgb": {"modalities": ["pose", "depth"]}, "v6_follow_rgb": {"modalities": ["follow"]}, "v7_follow_flow_rgb": {"modalities": ["follow", "flow"]}, "v8_follow_depth_rgb": {"modalities": ["follow", "depth"]}, } def get_modality_video_path(modality: str, episode_idx: int) -> Path: ep_name = f"episode_{episode_idx:06d}" if modality == "pose": return MOD_ROOT / "pose" / ep_name / "silhouette_silhouette.mp4" if modality == "flow": return MOD_ROOT / "flow" / ep_name / "preview.mp4" if modality == "depth": return MOD_ROOT / "depth" / ep_name / "depth.mp4" if modality == "follow": return MOD_ROOT / "follow" / ep_name / f"{ep_name}.mp4" raise ValueError(f"Unknown modality: {modality}") def read_video_frames(path: Path) -> np.ndarray: cap = cv2.VideoCapture(str(path)) frames = [] while True: ret, frame = cap.read() if not ret: break frames.append(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)) cap.release() if not frames: raise ValueError(f"No frames read from {path}") return np.stack(frames) def sample_frames(frames: np.ndarray, num_frames: int) -> np.ndarray: if len(frames) == num_frames: return frames indices = np.linspace(0, len(frames) - 1, num_frames).astype(int) return frames[indices] def read_rgb_pngs(rgb_dir: Path, num_frames: int = RGB_FRAMES) -> np.ndarray: frames = [] for i in range(1, num_frames + 1): path = rgb_dir / f"frame_{i:04d}.png" img = cv2.imread(str(path)) if img is None: raise FileNotFoundError(f"Missing {path}") frames.append(cv2.cvtColor(img, cv2.COLOR_BGR2RGB)) return np.stack(frames) def resize_frames(frames: np.ndarray, width: int = 256, height: int = 256) -> np.ndarray: if frames.shape[1] == height and frames.shape[2] == width: return frames src_pixels = frames.shape[1] * frames.shape[2] dst_pixels = width * height interpolation = cv2.INTER_LANCZOS4 if dst_pixels > src_pixels else cv2.INTER_AREA return np.stack([cv2.resize(f, (width, height), interpolation=interpolation) for f in frames]) def main(): parser = argparse.ArgumentParser(description=__doc__) parser.add_argument("--version", type=str, required=True, choices=list(VERSION_CONFIGS.keys())) parser.add_argument("--start", type=int, default=0) parser.add_argument("--stop", type=int, default=5000) parser.add_argument("--fps", type=int, default=8) parser.add_argument("--size", type=int, default=None) parser.add_argument("--width", type=int, default=None) parser.add_argument("--height", type=int, default=None) parser.add_argument("--output-suffix", type=str, default="") parser.add_argument( "--cond-frames", type=int, default=REASON_FRAMES, help="Reason modality frames per stream (default 25)", ) parser.add_argument( "--rgb-frames", type=int, default=RGB_FRAMES, help="RGB frames appended at end (default 24)", ) parser.add_argument("--validation-count", type=int, default=10) args = parser.parse_args() if args.width is not None or args.height is not None: if args.width is None or args.height is None: parser.error("--width and --height must be set together") out_width, out_height = args.width, args.height else: square = args.size if args.size is not None else 256 out_width = out_height = square config = VERSION_CONFIGS[args.version] output_name = f"{args.version}{args.output_suffix}" output_root = COMPOSED_ROOT / output_name videos_dir = output_root / "videos" cond_dir = output_root / "condition_images" videos_dir.mkdir(parents=True, exist_ok=True) cond_dir.mkdir(parents=True, exist_ok=True) video_paths, image_paths, prompts, state_paths, action_paths, failed = [], [], [], [], [], [] for idx in range(args.start, args.stop): ep_name = f"episode_{idx:06d}" rgb_dir = RAW_ROOT / ep_name / "rgb" instruction_file = RAW_ROOT / ep_name / "instruction" / "instruction.txt" state_path = RAW_ROOT / ep_name / "state" / "state.npy" action_path = RAW_ROOT / ep_name / "action" / "action.npy" try: if not state_path.is_file() or not action_path.is_file(): raise FileNotFoundError(f"Missing state/action for {ep_name}") modality_frames_list = [] for mod in config["modalities"]: mod_path = get_modality_video_path(mod, idx) frames = read_video_frames(mod_path) frames = sample_frames(frames, args.cond_frames) modality_frames_list.append(resize_frames(frames, out_width, out_height)) rgb_frames = read_rgb_pngs(rgb_dir, args.rgb_frames) rgb_frames = resize_frames(rgb_frames, out_width, out_height) combined = np.concatenate(modality_frames_list + [rgb_frames], axis=0) expected = args.cond_frames * len(config["modalities"]) + args.rgb_frames assert len(combined) == expected, f"expected {expected}, got {len(combined)}" out_video = videos_dir / f"{ep_name}.mp4" imageio.mimsave( str(out_video), combined, fps=args.fps, codec="libx264", macro_block_size=1 ) cond_image = cond_dir / f"{ep_name}.png" imageio.imwrite(str(cond_image), rgb_frames[0]) prompt = "robot manipulation task" if instruction_file.exists(): text = instruction_file.read_text().strip() if text: prompt = text video_paths.append(str(out_video)) image_paths.append(str(cond_image)) prompts.append(prompt) state_paths.append(str(state_path)) action_paths.append(str(action_path)) if idx % 500 == 0 or idx == args.start: print(f"[ok] {ep_name}: {len(combined)} frames") except Exception as e: print(f"[fail] {ep_name}: {e}") failed.append({"episode_idx": idx, "error": str(e)}) (output_root / "videos.txt").write_text("\n".join(video_paths) + "\n") (output_root / "images.txt").write_text("\n".join(image_paths) + "\n") (output_root / "prompt.txt").write_text("\n".join(prompts) + "\n") (output_root / "state_paths.txt").write_text("\n".join(state_paths) + "\n") (output_root / "action_paths.txt").write_text("\n".join(action_paths) + "\n") with (output_root / "metadata.csv").open("w", newline="") as f: writer = csv.DictWriter(f, fieldnames=["index", "image", "video", "text"]) writer.writeheader() for i, (video, image, prompt) in enumerate(zip(video_paths, image_paths, prompts)): writer.writerow({"index": i, "image": image, "video": video, "text": prompt}) val_count = min(args.validation_count, len(video_paths)) val_entries = [ { "sample_index": i, "caption": prompts[i], "image_path": image_paths[i], "video_path": video_paths[i], } for i in range(val_count) ] (output_root / "validation.json").write_text(json.dumps({"data": val_entries}, indent=2) + "\n") if failed: (output_root / "failed_episodes.json").write_text(json.dumps(failed, indent=2) + "\n") print(f"\nDone: {len(video_paths)} composed, {len(failed)} failed -> {output_root}") if __name__ == "__main__": main()