robometer/scripts/run_batch.py

#!/usr/bin/env python3
"""
Robometer prefix-robustness — full batch, dense curves.

For every episode and each of 5 prefix-sampling modes, run a full dense
per-frame scoring pass over the whole (optionally downsampled) video:
at every pool position t, pick 8 frames from [0, t] by the mode's rule
(always including frame 0 and frame t) and score with Robometer.
Result: 5 complete progress curves per episode.

Modes: uniform (= original benchmark), front_biased, back_biased,
       random_seed0, random_seed1.

Output layout (resume-safe: a mode .json that already exists is skipped):
  <out-dir>/episode_results/<chunk>_<episode>/<mode>.json

Local run (A6000 box):
  conda run -n robometer python run_batch.py

AutoDL (paths differ, 80G card, no downsampling):
  python run_batch.py --videos-root ... --robometer-repo ... --model-path ... \
      --fps 0 --max-frames 0 --batch-size 16
"""
from __future__ import annotations

import argparse
import json
import os
import sys
import time
import traceback
from pathlib import Path


def parse_args():
    p = argparse.ArgumentParser(description="Robometer prefix-robustness dense batch")
    p.add_argument("--videos-root",
                   default="/home/vcj9002/jianshu/workspace/code_keliang/Videos",
                   help="Dir containing chunk-*_filtered/ with episode_tasks.json")
    p.add_argument("--robometer-repo",
                   default="/home/vcj9002/jianshu/workspace/code_keliang/Current_Baseline/Robometer/robometer",
                   help="Robometer repo dir (has scripts/ and the robometer package)")
    p.add_argument("--model-path", default=None,
                   help="Robometer-4B dir (default: <robometer-repo>/../models/Robometer-4B)")
    p.add_argument("--out-dir", default=None,
                   help="Default: <this file>/../results_full")
    p.add_argument("--camera", default="wrist_image_left",
                   help="wrist_image_left = same as the original Robometer benchmark")
    p.add_argument("--fps", type=float, default=3.0,
                   help="Temporal downsample fps; 0 = keep native fps")
    p.add_argument("--max-frames", type=int, default=128,
                   help="Cap on pool size; 0 = no cap (needs big GPU/time)")
    p.add_argument("--batch-size", type=int, default=4,
                   help="Positions scored per model batch")
    p.add_argument("--gpu", default=None,
                   help="GPU id; default: auto-pick card with least used memory")
    p.add_argument("--limit", type=int, default=None,
                   help="Only process first N remaining episodes (smoke test)")
    return p.parse_args()


ARGS = parse_args()

# ── GPU choice must happen before torch import ─────────────────────────────
if "CUDA_VISIBLE_DEVICES" not in os.environ:
    if ARGS.gpu is not None:
        os.environ["CUDA_VISIBLE_DEVICES"] = str(ARGS.gpu)
    else:
        import subprocess
        try:
            out = subprocess.check_output(
                ["nvidia-smi", "--query-gpu=index,memory.used",
                 "--format=csv,noheader,nounits"], text=True)
            idx = min((l.split(",") for l in out.strip().splitlines()),
                      key=lambda x: int(x[1]))[0].strip()
        except Exception:
            idx = "0"
        os.environ["CUDA_VISIBLE_DEVICES"] = idx

ROBOMETER_REPO = Path(ARGS.robometer_repo).resolve()
sys.path.insert(0, str(ROBOMETER_REPO))
sys.path.insert(0, str(ROBOMETER_REPO / "scripts"))

import numpy as np  # noqa: E402

from benchmark_progress_mark_local import (  # noqa: E402
    RobometerLocalRunner,
    load_video_frames_with_indices,
    load_all_video_frames,
)
from robometer.data.dataset_types import ProgressSample, Trajectory  # noqa: E402

MODEL_PATH = ARGS.model_path or str(ROBOMETER_REPO.parent / "models" / "Robometer-4B")
VIDEOS_ROOT = Path(ARGS.videos_root)
CAMERA_DIR = f"observation.images.{ARGS.camera}"

OUT_DIR = (Path(ARGS.out_dir) if ARGS.out_dir
           else Path(__file__).resolve().parent.parent / "results_full")
EP_DIR = OUT_DIR / "episode_results"
EP_DIR.mkdir(parents=True, exist_ok=True)
ERR_PATH = OUT_DIR / "errors.log"

MODES = ["uniform", "front_biased", "back_biased", "random_seed0", "random_seed1"]
N_SLOTS = 8  # frames fed to the model per scoring call (original benchmark setting)


# ── prefix construction ────────────────────────────────────────────────────

def _fill_to_slots(idxs: list[int]) -> list[int]:
    """Return exactly N_SLOTS sorted indices; duplicates allowed when the
    candidate set is smaller (mirrors the original linspace behaviour)."""
    idxs = sorted(int(i) for i in idxs)
    if len(idxs) == N_SLOTS:
        return idxs
    pos = np.linspace(0, len(idxs) - 1, N_SLOTS, dtype=int)
    return [int(idxs[i]) for i in pos]


def build_frame_indices(t: int, mode: str) -> list[int]:
    """8 sorted indices in [0, t], always containing 0 and t."""
    if t == 0:
        return [0] * N_SLOTS
    if mode == "uniform":
        # identical to the original benchmark: duplicates possible at small t
        return [int(x) for x in np.linspace(0, t, N_SLOTS, dtype=int)]
    if mode == "front_biased":
        half = max(t // 2, 1)
        cand = sorted(set([0] + np.linspace(0, half, 6, dtype=int).tolist() + [t]))
        return _fill_to_slots(cand)
    if mode == "back_biased":
        half = t // 2
        cand = sorted(set([0] + np.linspace(half, t, 6, dtype=int).tolist() + [t]))
        return _fill_to_slots(cand)
    if mode in ("random_seed0", "random_seed1"):
        seed = 0 if mode.endswith("0") else 1
        # deterministic per position so resume/re-runs are reproducible
        rng = np.random.default_rng(seed * 1_000_003 + t)
        avail = list(range(1, t))
        k = min(6, len(avail))
        drawn = sorted(rng.choice(avail, k, replace=False).tolist()) if k else []
        return _fill_to_slots(sorted(set([0] + drawn + [t])))
    raise ValueError(f"unknown mode: {mode}")


# ── scoring ────────────────────────────────────────────────────────────────

def make_sample(pool: np.ndarray, idxs: list[int], pool_n: int, task: str):
    frames = pool[idxs]
    traj = Trajectory(
        frames=frames, frames_shape=tuple(frames.shape), task=task, id="0",
        metadata={"subsequence_length": pool_n}, video_embeddings=None)
    return ProgressSample(trajectory=traj, sample_type="progress")


def run_batched(runner, samples, batch_size):
    """Score samples in batches; returns final-frame score per sample.
    Falls back to batch size 1 on CUDA OOM."""
    import torch
    out = []
    i = 0
    bs = max(1, batch_size)
    while i < len(samples):
        chunk = samples[i:i + bs]
        try:
            preds, _ = runner._run_progress_samples(chunk)
            for p in preds:
                out.append(float(np.asarray(p).reshape(-1)[-1]))
            i += len(chunk)
        except torch.cuda.OutOfMemoryError:
            torch.cuda.empty_cache()
            if bs == 1:
                raise
            bs = max(1, bs // 2)
            print(f"    [OOM] retrying with batch_size={bs}", flush=True)
    return out


def load_pool(video_path: Path):
    """Load frames per CLI sampling settings. Returns (pool, total_raw, fps)."""
    if ARGS.fps <= 0 and ARGS.max_frames <= 0:
        frames, native_fps = load_all_video_frames(video_path)
        pool = np.stack(frames, axis=0)
        return pool, len(frames), float(native_fps)
    fps = ARGS.fps if ARGS.fps > 0 else 10_000.0        # huge = keep native
    max_frames = ARGS.max_frames if ARGS.max_frames > 0 else 10 ** 9
    pool, _idx, total_raw, native_fps = load_video_frames_with_indices(
        video_path, fps=fps, max_frames=max_frames, required_frames=[])
    return pool, total_raw, float(native_fps)


# ── episode enumeration ────────────────────────────────────────────────────

def list_episodes():
    eps = []
    for tasks_file in sorted(VIDEOS_ROOT.glob("chunk-*_filtered/episode_tasks.json")):
        meta = json.load(open(tasks_file))
        for e in meta["episodes"]:
            video = tasks_file.parent / CAMERA_DIR / e["episode"]
            if video.exists():
                eps.append({
                    "chunk": meta["chunk"],
                    "episode": e["episode"],
                    "task": " and ".join(e["tasks"]),
                    "video": video,
                })
    return eps


def episode_dir(ep) -> Path:
    stem = ep["episode"].replace(".mp4", "")
    return EP_DIR / f"{ep['chunk']}_{stem}"


def main():
    episodes = list_episodes()
    todo = [e for e in episodes
            if not all((episode_dir(e) / f"{m}.json").exists() for m in MODES)]
    if ARGS.limit:
        todo = todo[:ARGS.limit]
    print(f"GPU  : CUDA_VISIBLE_DEVICES={os.environ.get('CUDA_VISIBLE_DEVICES')}")
    print(f"Model: {MODEL_PATH}")
    print(f"Out  : {EP_DIR}")
    print(f"Sampling: fps={ARGS.fps or 'native'} max_frames={ARGS.max_frames or 'unlimited'} "
          f"camera={ARGS.camera} batch={ARGS.batch_size}")
    print(f"Episodes: total={len(episodes)} todo={len(todo)}")
    if not todo:
        print("Nothing to do.")
        return

    runner = RobometerLocalRunner(model_path=MODEL_PATH)

    for i, ep in enumerate(todo, 1):
        ep_out = episode_dir(ep)
        ep_out.mkdir(parents=True, exist_ok=True)
        print(f"[{i}/{len(todo)}] {ep['chunk']}/{ep['episode']}", flush=True)
        try:
            pool, total_raw, native_fps = load_pool(ep["video"])
            n = len(pool)
            print(f"  pool={n} frames (raw={total_raw}, fps={native_fps:.2f})", flush=True)
            for mode in MODES:
                mode_path = ep_out / f"{mode}.json"
                if mode_path.exists():
                    continue
                t0 = time.time()
                all_idxs = [build_frame_indices(t, mode) for t in range(n)]
                samples = [make_sample(pool, idxs, n, ep["task"]) for idxs in all_idxs]
                raw_scores = run_batched(runner, samples, ARGS.batch_size)
                scores_100 = [round(s * 100.0, 4) if s <= 2.0 else round(s, 4)
                              for s in raw_scores]
                payload = {
                    "chunk": ep["chunk"], "episode": ep["episode"],
                    "task": ep["task"], "camera": ARGS.camera,
                    "native_fps": round(native_fps, 3),
                    "total_raw_frames": total_raw, "pool_n": n,
                    "fps_arg": ARGS.fps, "max_frames_arg": ARGS.max_frames,
                    "mode": mode,
                    "scores_raw": [round(s, 6) for s in raw_scores],
                    "scores_100": scores_100,
                    "frame_indices": all_idxs,
                }
                tmp = mode_path.with_suffix(".json.tmp")
                tmp.write_text(json.dumps(payload))
                tmp.rename(mode_path)   # atomic: resume never sees half a file
                print(f"  {mode}: {n} positions in {time.time()-t0:.1f}s", flush=True)
        except Exception:
            with open(ERR_PATH, "a") as ef:
                ef.write(f"=== {ep['chunk']}/{ep['episode']} ===\n")
                ef.write(traceback.format_exc() + "\n")
            print(f"  ERROR (logged to {ERR_PATH.name}), continuing", flush=True)

    print("Done:", EP_DIR)


if __name__ == "__main__":
    main()