chess-sim/code/sim/make_train_view.py

"""Training copy of a phase 2 dataset with only the cameras the policy uses.

SmolVLA is trained on the overhead and wrist views; the dataset keeps the rover view
for later. LeRobot's remove_feature writes a new dataset without the dropped cameras,
so training never decodes their videos. The original is untouched.

Run:  .venv/bin/python sim/make_train_view.py data/so101_chess_sim data/so101_chess_sim_train
"""
from __future__ import annotations

import argparse


def main():
    from lerobot.datasets.dataset_tools import remove_feature
    from lerobot.datasets.lerobot_dataset import LeRobotDataset

    ap = argparse.ArgumentParser(description=__doc__)
    ap.add_argument("src")
    ap.add_argument("dst")
    ap.add_argument("--drop", nargs="+", default=["observation.images.rover"])
    args = ap.parse_args()
    ds = LeRobotDataset("local/so101_chess_sim", root=args.src)
    out = remove_feature(ds, feature_names=args.drop, output_dir=args.dst, repo_id="local/so101_chess_sim_train")
    print(f"{args.dst}: {out.num_episodes} episodes, {out.num_frames} frames, features {list(out.meta.features)}")


if __name__ == "__main__":
    main()