| |
| """Locate torchcodec failures while reading the canonical RECAP Value dataset. |
| |
| This is a bounded diagnostic: it never writes dataset files and prints the |
| video path supplied to the decoder when a sample fails. It intentionally |
| uses the same ValueDataset implementation as the official Value worker. |
| """ |
|
|
| from __future__ import annotations |
|
|
| import argparse |
| import sys |
| from pathlib import Path |
|
|
| import torch |
| from torch.utils.data import DataLoader |
|
|
| SCRIPT_DIR = Path(__file__).resolve().parent |
| PROJECT_ROOT = SCRIPT_DIR.parent |
| if str(SCRIPT_DIR) not in sys.path: |
| sys.path.insert(0, str(SCRIPT_DIR)) |
|
|
|
|
| def parse_args() -> argparse.Namespace: |
| parser = argparse.ArgumentParser() |
| parser.add_argument("--dataset-path", type=Path, default=PROJECT_ROOT / "datasets/recap_libero10_task0/libero10_task0_train") |
| parser.add_argument("--tag", default="recap_v1_fail300_g1") |
| parser.add_argument("--samples-per-seed", type=int, default=100) |
| parser.add_argument("--seeds", type=int, nargs="+", default=[0, 1, 2, 3, 4, 5, 6, 7]) |
| return parser.parse_args() |
|
|
|
|
| def main() -> int: |
| args = parse_args() |
| dataset_path = args.dataset_path.resolve() |
| if PROJECT_ROOT not in dataset_path.parents: |
| raise ValueError(f"dataset path must be inside {PROJECT_ROOT}: {dataset_path}") |
|
|
| import lerobot.common.datasets.lerobot_dataset as lerobot_dataset |
|
|
| original_decode = lerobot_dataset.decode_video_frames |
|
|
| def traced_decode(video_path, timestamps, tolerance_s, backend=None): |
| try: |
| return original_decode(video_path, timestamps, tolerance_s, backend) |
| except Exception as exc: |
| print( |
| f"DECODE_FAILURE\t{video_path}\t{type(exc).__name__}: {exc}", |
| flush=True, |
| ) |
| raise |
|
|
| lerobot_dataset.decode_video_frames = traced_decode |
|
|
| from rlinf.data.datasets.recap.value_model import ValueDataset |
|
|
| dataset = ValueDataset( |
| dataset_path=dataset_path, |
| robot_type="libero", |
| model_type="pi05", |
| action_horizon=10, |
| gamma=1.0, |
| return_min=-779.0, |
| return_max=0.0, |
| normalize_to_minus_one_zero=True, |
| split="train", |
| tag=args.tag, |
| ) |
| print(f"DATASET_LENGTH\t{len(dataset)}", flush=True) |
|
|
| for seed in args.seeds: |
| generator = torch.Generator() |
| generator.manual_seed(seed) |
| loader = DataLoader( |
| dataset, |
| batch_size=1, |
| shuffle=True, |
| generator=generator, |
| num_workers=0, |
| collate_fn=lambda batch: batch[0], |
| ) |
| try: |
| for index, _sample in enumerate(loader): |
| if index + 1 >= args.samples_per_seed: |
| break |
| print(f"SEED_PASS\t{seed}\t{args.samples_per_seed}", flush=True) |
| except Exception as exc: |
| print(f"SEED_STOP\t{seed}\t{index}\t{type(exc).__name__}: {exc}", flush=True) |
|
|
| return 0 |
|
|
|
|
| if __name__ == "__main__": |
| raise SystemExit(main()) |
|
|
