exp2/code/select_frames.py

"""Stage 1: frame selection per arm -> keeps/<arm>/ep{N}.npy + keeps/<arm>/stats.json
    python select.py --trajs T --arm raw
    python select.py --trajs T --arm uni3
    python select.py --trajs T --arm isr_d4 --isr --d 4 --lam 0.002 --gthr 2.3
    python select.py --trajs T --arm rand_d4 --match isr_d4      (random subset, per-episode count matched)"""
import argparse, json, sys
from pathlib import Path
import numpy as np
from scipy.ndimage import gaussian_filter1d
sys.path.insert(0, "/workspace/isr/src/teleop_std_poc"); from isr_resample import isr_resample, _gripper_forced, time_uniform  # noqa

ap = argparse.ArgumentParser(); ap.add_argument("--trajs", required=True); ap.add_argument("--arm", required=True); ap.add_argument("--out", default="/workspace/isr/keeps")
ap.add_argument("--isr", action="store_true"); ap.add_argument("--d", type=float); ap.add_argument("--lam", type=float); ap.add_argument("--gthr", type=float); ap.add_argument("--ldist", type=float, default=1.0)
ap.add_argument("--uniform", type=int, default=0); ap.add_argument("--match", default=None); ap.add_argument("--seed", type=int, default=0); ap.add_argument("--smooth", type=float, default=0, help="gaussian sigma (frames) applied to positions BEFORE info computation (selection still indexes raw frames)")
a = ap.parse_args(); out = Path(a.out) / a.arm; out.mkdir(parents=True, exist_ok=True); rng = np.random.default_rng(a.seed)
stats = {"arm": a.arm, "knobs": vars(a), "episodes": {}}
for f in sorted(Path(a.trajs).glob("ep*.npz")):
    d = np.load(f); n = len(d["positions"]); forced = 0
    if a.isr:
        P = gaussian_filter1d(d["positions"], a.smooth, axis=0) if a.smooth > 0 else d["positions"]
        keep = isr_resample(P, d["timestamps"], d["gripper"], a.d, a.ldist, a.lam, a.gthr); forced = len(_gripper_forced(d["gripper"], n, a.gthr))
    elif a.uniform: keep = time_uniform(n, a.uniform)
    elif a.match:
        m = len(np.load(Path(a.out) / a.match / f"{f.stem}.npy")); mid = rng.choice(np.arange(1, n - 1), max(m - 2, 0), replace=False)
        keep = sorted(set([0, n - 1] + mid.tolist()))
    else: keep = list(range(n))
    np.save(out / f"{f.stem}.npy", np.array(keep, dtype=int)); stats["episodes"][f.stem] = {"in": n, "out": len(keep), "forced": forced}
r = [e["out"] / e["in"] for e in stats["episodes"].values()]
stats["summary"] = {"episodes": len(r), "frames_in": sum(e["in"] for e in stats["episodes"].values()), "frames_out": sum(e["out"] for e in stats["episodes"].values()),
                    "mean_kept": float(np.mean(r)), "min_kept": float(np.min(r)), "max_kept": float(np.max(r))}
(out / "stats.json").write_text(json.dumps(stats, indent=1)); print(a.arm, json.dumps(stats["summary"]))