chess-sim/code/sim/set_norm_stats.py

"""Make a fine-tune keep a checkpoint's normalization.

When `lerobot-train` fine-tunes from a checkpoint it replaces the checkpoint's saved
state/action normalization with the training dataset's statistics
(lerobot_train.py passes them as processor overrides, and those take precedence).
Merging correction data into a dataset changes those statistics, which by itself
shifts the policy's commands by millimetres at the grasp (sim/reports/pipeline_check).

This writes the checkpoint's saved statistics for `observation.state` and `action`
into the dataset's meta/stats.json (the original is kept as meta/stats_original.json),
so the override hands the checkpoint its own normalization back. Image statistics are
left alone (SmolVLA does not normalize images).

Run:  .venv/bin/python sim/set_norm_stats.py data/merged_v2 models/baseline
"""
from __future__ import annotations

import argparse
import json
import os
import shutil
from pathlib import Path


def saved_stats(checkpoint: Path) -> dict:
    from safetensors.torch import load_file

    f = next(checkpoint.glob("policy_preprocessor_step_*_normalizer_processor.safetensors"))
    out: dict = {}
    for key, tensor in load_file(str(f)).items():
        feature, stat = key.rsplit(".", 1)
        out.setdefault(feature, {})[stat] = tensor.float().numpy().tolist()
    return out


def main():
    ap = argparse.ArgumentParser(description=__doc__)
    ap.add_argument("dataset")
    ap.add_argument("checkpoint")
    args = ap.parse_args()
    meta = Path(args.dataset) / "meta"
    stats_path = meta / "stats.json"
    backup = meta / "stats_original.json"
    if not backup.exists():
        shutil.copyfile(stats_path, backup)
    stats = json.loads(backup.read_text())
    saved = saved_stats(Path(args.checkpoint))
    for feature in ("observation.state", "action"):
        for stat, value in saved[feature].items():
            if stat in stats[feature]:
                stats[feature][stat] = value
    tmp = meta / "stats.json.tmp"          # a new file, not an in-place edit: hard-linked copies stay intact
    tmp.write_text(json.dumps(stats, indent=2))
    os.replace(tmp, stats_path)
    for feature in ("observation.state", "action"):
        print(feature, {s: [round(x, 3) for x in stats[feature][s]] for s in ("mean", "std")})


if __name__ == "__main__":
    main()