"""Preprocess 5000 episodes with 24-frame RGB and 25-frame aligned reason data. Per episode (same episode list as coaf_dataset/splits/train_5k.json): rgb/ frame_0001.png .. frame_0024.png (256x256, 24 frames) rgb_align/ frame_0001.png .. frame_0025.png (same timesteps as reason) state/state.npy (25, 7) — indices match rgb_align / depth / pose / flow / follow action/action.npy (25, 7) instruction/instruction.txt manifest.json records both index arrays and shapes """ from __future__ import annotations import argparse import json import sys import time from pathlib import Path import cv2 import numpy as np import tensorflow_datasets as tfds SCRIPT_DIR = Path(__file__).resolve().parent sys.path.insert(0, str(SCRIPT_DIR)) from sampling import RGB_FRAMES, REASON_FRAMES, reason_indices, rgb_indices TFDS_DIR = "/project/llmsvgen/sunkai/robomaster_3d/CoAF/data/bridge_v_full/1.0.0" DATASET_ROOT = Path("/project/llmsvgen/sunkai/robomaster_3d/Casual_CoAF/coaf_dataset_24_25") SPLIT_FILE = DATASET_ROOT / "splits" / "train_5k.json" OUTPUT_ROOT = DATASET_ROOT / "raw" IMAGE_SIZE = 256 MIN_RAW_FRAMES = max(RGB_FRAMES, REASON_FRAMES) def save_rgb_frames(frames: np.ndarray, out_dir: Path, image_size: int) -> None: out_dir.mkdir(parents=True, exist_ok=True) for i, frame in enumerate(frames): if frame.shape[0] != image_size or frame.shape[1] != image_size: frame = cv2.resize( frame, (image_size, image_size), interpolation=cv2.INTER_LANCZOS4 ) cv2.imwrite( str(out_dir / f"frame_{i + 1:04d}.png"), cv2.cvtColor(frame, cv2.COLOR_RGB2BGR), ) def main() -> None: parser = argparse.ArgumentParser(description=__doc__) parser.add_argument("--tfds-dir", type=str, default=TFDS_DIR) parser.add_argument("--split-file", type=Path, default=SPLIT_FILE) parser.add_argument("--output-root", type=Path, default=OUTPUT_ROOT) parser.add_argument("--image-size", type=int, default=IMAGE_SIZE) parser.add_argument("--skip-existing", action="store_true") parser.add_argument( "--start", type=int, default=0, help="Minimum dataset output index (episode_000000 -> 0)", ) parser.add_argument( "--stop", type=int, default=None, help="Exclusive max dataset output index (default: all in split)", ) args = parser.parse_args() episode_ids = json.loads(args.split_file.read_text()) if args.stop is not None: episode_ids = episode_ids[args.start : args.stop] else: episode_ids = episode_ids[args.start :] print(f"Loaded {len(episode_ids)} episode IDs from {args.split_file} " f"(dataset_idx {args.start}..{args.stop if args.stop is not None else 'end'})") builder = tfds.builder_from_directory(args.tfds_dir) dataset = builder.as_dataset(split="train") args.output_root.mkdir(parents=True, exist_ok=True) target_set = set(episode_ids) target_sorted = sorted(episode_ids) id_to_out = {eid: i for i, eid in enumerate(target_sorted)} start_time = time.time() processed = 0 failed = [] print(f"RGB frames={RGB_FRAMES}, reason-aligned frames={REASON_FRAMES}") print(f"Episode ID range: {target_sorted[0]} ~ {target_sorted[-1]}") for episode_idx, episode in enumerate(dataset): if episode_idx > target_sorted[-1]: break if episode_idx not in target_set: continue out_idx = id_to_out[episode_idx] episode_dir = args.output_root / f"episode_{out_idx:06d}" done_marker = episode_dir / "rgb" / f"frame_{RGB_FRAMES:04d}.png" if args.skip_existing and done_marker.exists(): processed += 1 continue try: steps = list(episode["steps"].as_numpy_iterator()) num_steps = len(steps) if num_steps < MIN_RAW_FRAMES: raise ValueError(f"num_steps={num_steps} < {MIN_RAW_FRAMES}") states_raw = np.stack([s["observation"]["state"] for s in steps]) actions_raw = np.stack([s["action"] for s in steps]) rgb_raw = np.stack([s["observation"]["image_0"] for s in steps]) instruction = steps[0]["language_instruction"] if isinstance(instruction, bytes): instruction = instruction.decode("utf-8", errors="replace") instruction = instruction.strip() idx_rgb = rgb_indices(num_steps) idx_reason = reason_indices(num_steps) states = states_raw[idx_reason] actions = actions_raw[idx_reason] rgb_frames = rgb_raw[idx_rgb] rgb_align_frames = rgb_raw[idx_reason] assert states.shape == (REASON_FRAMES, 7) assert actions.shape == (REASON_FRAMES, 7) assert len(rgb_frames) == RGB_FRAMES assert len(rgb_align_frames) == REASON_FRAMES save_rgb_frames(rgb_frames, episode_dir / "rgb", args.image_size) save_rgb_frames(rgb_align_frames, episode_dir / "rgb_align", args.image_size) state_dir = episode_dir / "state" action_dir = episode_dir / "action" instr_dir = episode_dir / "instruction" for d in (state_dir, action_dir, instr_dir): d.mkdir(parents=True, exist_ok=True) np.save(str(state_dir / "state.npy"), states) np.save(str(action_dir / "action.npy"), actions) (instr_dir / "instruction.txt").write_text(instruction, encoding="utf-8") manifest = { "original_episode_idx": episode_idx, "dataset_idx": out_idx, "num_raw_frames": num_steps, "instruction": instruction, "rgb_frames": RGB_FRAMES, "reason_frames": REASON_FRAMES, "rgb_indices": idx_rgb.tolist(), "reason_indices": idx_reason.tolist(), "state_shape": list(states.shape), "action_shape": list(actions.shape), "image_size": args.image_size, "sampling_note": ( "rgb uses rgb_indices; rgb_align/state/action/reason modalities " "share reason_indices" ), } (episode_dir / "manifest.json").write_text( json.dumps(manifest, indent=2) + "\n" ) processed += 1 if processed % 200 == 0: elapsed = time.time() - start_time eps = processed / elapsed remaining = (len(episode_ids) - processed) / eps print( f" [{processed}/{len(episode_ids)}] episode_idx={episode_idx}, " f"{elapsed:.0f}s elapsed, ~{remaining:.0f}s remaining" ) except Exception as e: print(f" [FAIL] episode_idx={episode_idx}: {e}") failed.append({"episode_idx": episode_idx, "error": str(e)}) elapsed = time.time() - start_time print(f"\nDone! Processed {processed}/{len(episode_ids)} episodes in {elapsed:.0f}s") if failed: fail_path = args.output_root / "preprocess_failures.json" fail_path.write_text(json.dumps(failed, indent=2) + "\n") print(f"Failures saved to {fail_path}") if __name__ == "__main__": main()