File size: 18,597 Bytes
45e45cb | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 | #!/usr/bin/env python3
"""
ProgressLM demo-robustness β full batch (ORIGINAL ProgressLM-3B-RL).
Mirrors the finished Robometer prefix-robustness experiment, but for the
DEMO-based ProgressLM. ProgressLM has NO history prefix: it builds a labelled
visual demonstration (N reference frames tagged 0% .. 100%) and scores ONE
current frame (stage_to_estimate) against that demo. So the perturbation is the
DEMO ORGANISATION (how the reference frames are sampled/arranged), while the
target/current frame is held fixed.
For every episode and each of 5 demo-organisation modes, we score the 4
checkpoint frames (pool 1/4, 2/4, 3/4, end) against a self-demo built from the
SAME episode's pool. Result: 5 scores per checkpoint per episode.
demo5_uniform (baseline) : 5 frames, anchors 0/25/50/75/100% (total_steps=4)
demo3_sparse : 3 frames, anchors 0/50/100% (total_steps=2)
demo9_dense : 9 frames, anchors every 12.5% (total_steps=8)
demo5_jitterA : 5 frames, middle anchors jittered +/-5% (seed=0),
labels RECOMPUTED from the true jittered position
demo5_jitterB : same, seed=1
RED LINE (v4): demo labels stay honest β a demo frame's % label is its true
temporal position in the pool. Uniform/sparse/dense anchors sit at exact
uniform fractions, so ProgressLM's own uniform labels are honest. Jitter anchors
move, so their labels are recomputed to the true position (rounded to integer %).
Model call is REUSED from the parity-verified RMBench wrapper
(progresslm_src: core.model.Qwen2VLChat + prompts.visual_demo_prompt), pointed at
the original ProgressLM-3B-RL. We do not rewrite the model call; for the jitter
modes we only substitute an honest progress-shift label string into the exact
same prompt structure (the stock builder can only emit uniform labels).
Output (resume-safe: an existing <mode>.json is skipped):
<out-dir>/episode_results/<chunk>_<episode>/<mode>.json
<out-dir>/progresslm_refs/<chunk>_<episode>/f####.png (demo + target frames)
Env (A6000 box): conda easyr1 (torch + transformers 4.57 + qwen-vl-utils + av + flash_attn)
/home/vcj9002/miniconda3/envs/easyr1/bin/python run_batch.py --gpu 8
"""
from __future__ import annotations
import argparse
import importlib.util as ilu
import json
import os
import re
import sys
import time
import traceback
import zlib
from pathlib import Path
def parse_args():
here = Path(__file__).resolve()
p = argparse.ArgumentParser(description="ProgressLM demo-robustness 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("--progresslm-code",
default="/home/vcj9002/jianshu/workspace/code_keliang/autodl_upload/test/"
"RMBench/Vanilla_Baseline/ProgressLM/progresslm/progresslm_src",
help="Parity-verified ProgressLM runtime (core/ prompts/ datasets/)")
p.add_argument("--model-path",
default="/home/vcj9002/jianshu/workspace/code_keliang/Current_Baseline/"
"ProgressLM/models/ProgressLM-3B-RL-qwen25vl",
help="ORIGINAL ProgressLM-3B-RL (7.6G). Points at a symlink whose "
"name carries 'qwen25' so the wrapper's own model-class fallback "
"(listinstr) selects Qwen2_5_VL under transformers>=4.57, which "
"reports model_type='qwen2_5_vl_text'. Same checkpoint, same "
"inference path β no wrapper edits.")
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 Robometer experiment")
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")
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()
# ββ GPU choice must happen before torch import βββββββββββββββββββββββββββββ
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
import numpy as np # noqa: E402
import av # noqa: E402
from PIL import Image # noqa: E402
# ββ ProgressLM parity wrapper: REUSE model call + prompt builder βββββββββββ
_CODE = str(Path(ARGS.progresslm_code).resolve())
sys.path.insert(0, _CODE)
from core.model import Qwen2VLChat # noqa: E402 (healthy package)
# prompts/__init__ imports deleted modules and crashes -> load the file directly
_spec = ilu.spec_from_file_location("plm_vdp", os.path.join(_CODE, "prompts/visual_demo_prompt.py"))
_vdp = ilu.module_from_spec(_spec)
sys.modules["plm_vdp"] = _vdp
_spec.loader.exec_module(_vdp)
build_visual_demo_prompt_from_item = _vdp.build_visual_demo_prompt_from_item
VISUAL_DEMO_SYSTEM_PROMPT = _vdp.VISUAL_DEMO_SYSTEM_PROMPT
MODEL_PATH = ARGS.model_path
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"
REFS_DIR = OUT_DIR / "progresslm_refs"
EP_DIR.mkdir(parents=True, exist_ok=True)
ERR_PATH = OUT_DIR / "errors.log"
MODES = ["demo5_uniform", "demo3_sparse", "demo9_dense", "demo5_jitterA", "demo5_jitterB"]
FRACS = ["1/4", "2/4", "3/4", "end"]
# ββ pool sampling: COPIED VERBATIM from Robometer so pools are byte-identical β
def load_all_video_frames(video_path: Path):
all_frames = []
with av.open(str(video_path)) as container:
stream = container.streams.video[0]
rate = stream.average_rate or stream.guessed_rate
native_fps = float(rate) if rate is not None else 1.0
for frame in container.decode(stream):
all_frames.append(frame.to_ndarray(format="rgb24"))
if not all_frames:
raise RuntimeError(f"Could not extract frames from video: {video_path}")
return all_frames, native_fps
def sample_video_frames_with_indices(all_frames, *, native_fps, fps, max_frames):
"""required_frames=[] branch of Robometer's sampler (identical pools)."""
total_frames = len(all_frames)
if fps <= 0:
fps = native_fps
if native_fps > 0:
desired_frames = int(round(total_frames * (fps / native_fps)))
else:
desired_frames = total_frames
desired_frames = max(1, min(desired_frames, total_frames, max_frames))
if desired_frames == total_frames:
sampled_indices = list(range(total_frames))
else:
base_indices = np.linspace(0, total_frames - 1, desired_frames, dtype=int).tolist()
sampled_indices = sorted(set(base_indices))
if len(sampled_indices) < desired_frames:
for idx in base_indices:
if idx not in sampled_indices:
sampled_indices.append(idx)
if len(sampled_indices) == desired_frames:
break
if len(sampled_indices) < desired_frames:
for idx in range(total_frames):
if idx not in sampled_indices:
sampled_indices.append(idx)
if len(sampled_indices) == desired_frames:
break
sampled_indices = sorted(sampled_indices[:desired_frames])
sampled_frames = np.stack([all_frames[idx] for idx in sampled_indices], axis=0)
return sampled_frames, sampled_indices
def load_pool(video_path: Path):
"""Returns (pool[N,H,W,3], total_raw_frames, native_fps). Same as Robometer."""
all_frames, native_fps = load_all_video_frames(video_path)
if ARGS.fps <= 0 and ARGS.max_frames <= 0:
pool = np.stack(all_frames, axis=0)
return pool, len(all_frames), float(native_fps)
fps = ARGS.fps if ARGS.fps > 0 else 10_000.0
max_frames = ARGS.max_frames if ARGS.max_frames > 0 else 10 ** 9
pool, _idx = sample_video_frames_with_indices(
all_frames, native_fps=native_fps, fps=fps, max_frames=max_frames)
return pool, len(all_frames), float(native_fps)
# ββ checkpoints (target frames) β identical to Robometer's render ββββββββββ
def checkpoints_of(pool_n: int) -> list[int]:
return [int((pool_n - 1) * k / 4) for k in (1, 2, 3, 4)]
# ββ demo organisation ββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def _ep_seed(ep_key: str, mode_seed: int) -> int:
"""Deterministic per (episode, mode) seed. Mirrors Robometer's
`seed * 1_000_003 + <varying-unit>`; here the varying unit is the episode
(each episode gets one demo per mode). Reproducible + resume-safe."""
return mode_seed * 1_000_003 + int(zlib.crc32(ep_key.encode()))
def build_demo(mode: str, n: int, ep_key: str):
"""Return (demo_idx, labels, total_steps_or_None).
total_steps is set for the uniform-spacing modes (uniform/sparse/dense): the
stock ProgressLM builder emits its own uniform labels from total_steps, which
are honest because the anchors sit at exact uniform fractions. For jitter,
total_steps is None and `labels` are the honest recomputed integer percents.
"""
if mode == "demo5_uniform":
nd = 5
elif mode == "demo3_sparse":
nd = 3
elif mode == "demo9_dense":
nd = 9
elif mode in ("demo5_jitterA", "demo5_jitterB"):
seed = 0 if mode.endswith("A") else 1
rng = np.random.default_rng(_ep_seed(ep_key, seed))
base = [0.0, 0.25, 0.5, 0.75, 1.0]
fracs = [base[0]]
for f in base[1:-1]: # jitter middle anchors only
fracs.append(min(1.0, max(0.0, f + float(rng.uniform(-0.05, 0.05)))))
fracs.append(base[-1]) # endpoints fixed at 0 / 100%
demo_idx = [max(0, min(n - 1, int(round(f * (n - 1))))) for f in fracs]
labels = [round(i / max(n - 1, 1) * 100) for i in demo_idx] # honest
return demo_idx, labels, None
else:
raise ValueError(f"unknown mode {mode}")
# uniform-spacing modes: int(linspace) demo sampling == RMBench bundle parity
demo_idx = [int(x) for x in np.linspace(0, n - 1, nd)]
total_steps = nd - 1
labels = [round(i / total_steps * 100) for i in range(nd)]
return demo_idx, labels, total_steps
def build_prompt_custom_labels(task_goal, demo_paths, labels, target_path):
"""Replicate build_visual_demo_prompt EXACTLY, but with an explicit honest
progress-shift label string (used only by the jitter modes)."""
shifts = " ".join(f"<image> {lab}%" for lab in labels)
msgs = [
{"type": "text", "value": VISUAL_DEMO_SYSTEM_PROMPT},
{"type": "text", "value": f"The overall task goal is {task_goal}"},
{"type": "text", "value": _vdp.VISUAL_DEMO_INSTRUCTION_PART1},
]
for dp in demo_paths:
msgs.append({"type": "image", "value": dp})
msgs.append({"type": "text",
"value": f"The progress shifts across all given visual demos is: {shifts}"})
msgs.append({"type": "text", "value": _vdp.VISUAL_DEMO_INSTRUCTION_PART2})
msgs.append({"type": "image", "value": target_path})
msgs.append({"type": "text", "value": _vdp.VISUAL_DEMO_INSTRUCTION_PART3})
return msgs
def parse_visual_demo_response(response: str):
"""Extract <score> as 0..1 (or 'n/a'/None). Mirrors ProgressLM's parser."""
if not response:
return {"score": None}
m = re.search(r"<score>(.*?)</score>", response, re.DOTALL)
if not m:
return {"score": None}
s = m.group(1).strip()
if s.lower() in ("n/a", "na"):
return {"score": "n/a"}
try:
v = float(s[:-1]) / 100.0 if s.endswith("%") else float(s)
if v > 1.0:
v = v / 100.0
return {"score": max(0.0, min(1.0, v))}
except ValueError:
return {"score": None}
def save_frame(pool, idx: int, refs_dir: Path) -> str:
refs_dir.mkdir(parents=True, exist_ok=True)
p = refs_dir / f"f{int(idx):04d}.png"
if not p.exists():
Image.fromarray(pool[int(idx)]).save(p)
return str(p)
# ββ episode enumeration (identical to Robometer) βββββββββββββββββββββββββββ
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_key(ep) -> str:
return f"{ep['chunk']}_{ep['episode'].replace('.mp4', '')}"
def episode_dir(ep) -> Path:
return EP_DIR / episode_key(ep)
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}")
print(f"Modes: {MODES}")
print(f"Episodes: total={len(episodes)} todo={len(todo)}")
if not todo:
print("Nothing to do.")
return
model = Qwen2VLChat(model_path=str(MODEL_PATH), system_prompt=VISUAL_DEMO_SYSTEM_PROMPT)
for i, ep in enumerate(todo, 1):
ep_key = episode_key(ep)
ep_out = episode_dir(ep)
ep_out.mkdir(parents=True, exist_ok=True)
refs_dir = REFS_DIR / ep_key
print(f"[{i}/{len(todo)}] {ep['chunk']}/{ep['episode']}", flush=True)
try:
pool, total_raw, native_fps = load_pool(ep["video"])
n = len(pool)
target_idx = checkpoints_of(n)
print(f" pool={n} frames (raw={total_raw}, fps={native_fps:.2f}) "
f"targets={target_idx}", flush=True)
# cache the 4 target frames once (shared across modes)
target_paths = [save_frame(pool, t, refs_dir) for t in target_idx]
for mode in MODES:
mode_path = ep_out / f"{mode}.json"
if mode_path.exists():
continue
t0 = time.time()
demo_idx, labels, total_steps = build_demo(mode, n, ep_key)
demo_paths = [save_frame(pool, di, refs_dir) for di in demo_idx]
scores_100, scores_raw, raw_responses = [], [], []
for tp in target_paths:
if total_steps is not None: # uniform/sparse/dense
item = {"task_goal": ep["task"], "visual_demo": demo_paths,
"total_steps": total_steps, "stage_to_estimate": tp}
msg = build_visual_demo_prompt_from_item(item)
else: # jitter: honest labels
msg = build_prompt_custom_labels(ep["task"], demo_paths, labels, tp)
resp = model.generate(msg)
s = parse_visual_demo_response(resp).get("score")
raw_responses.append(resp)
if s in (None, "n/a"):
scores_raw.append(None)
scores_100.append(None)
else:
scores_raw.append(round(float(s), 6))
scores_100.append(round(float(s) * 100.0, 4))
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,
"n_demo": len(demo_idx),
"total_steps": (total_steps if total_steps is not None
else len(demo_idx) - 1),
"demo_frame_indices": demo_idx,
"demo_labels": labels,
"checkpoint_fracs": FRACS,
"target_frame_indices": target_idx,
"scores_raw": scores_raw,
"scores_100": scores_100,
"raw_responses": raw_responses,
}
tmp = mode_path.with_suffix(".json.tmp")
tmp.write_text(json.dumps(payload))
tmp.rename(mode_path) # atomic: resume never sees half a file
sv = ["n/a" if s is None else f"{s:.0f}" for s in scores_100]
print(f" {mode}: scores={sv} 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()
|