File size: 15,386 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 | #!/usr/bin/env python3
"""
VLAC prefix-robustness β full batch (sampling-path perturbation).
Experiment philosophy (same as Robometer / TopReward): the same physical
target frame should get roughly the same accumulated progress value no matter
how the frames leading up to it were sampled. Large spread across sampling
paths = not robust (Prefix Range > 20 pts).
VLAC is an InternVL2-8B pairwise critic: for an adjacent sampled-frame pair
[prev, cur] it emits the progress INCREMENT of cur vs prev; the increments are
accumulated along the sampled sequence into a 0-100 absolute value curve
(evo_vlac/utils/model_utils.py: get_trajectory_critic + critic_to_value_simple).
The value at a frame therefore depends on which intermediate frames were
sampled on the way there -- exactly the robustness axis under test.
For every episode we compress the source video with VLAC's own preprocessing
(5 fps, 448x448 -- the model-side fixed pipeline, NOT changed to 3 fps) into a
frame sequence `seq` of length N, take 4 target frames (1/4, 2/4, 3/4, end),
and for each of 5 sampling-path modes build a frame sequence that starts at 0,
ends at the target t, and only changes which intermediate frames are kept.
Each path is accumulated with the exact baseline critic call
(get_trajectory_critic, ref_num=0 zero-shot, skip=1); the value read is the
accumulated value at t (= last element of the value curve for that path).
Modes (5 paths to the same target t):
dense_all keep every frame in [0, t] (baseline, skip=1)
stride2 every 2nd frame from 0 to t
stride4 every 4th frame from 0 to t
front_dense [0, t/2] dense, (t/2, t] stride4
back_dense [0, t/2) stride4, [t/2, t] dense
Output layout (resume-safe: a mode .json that already exists is skipped):
<out-dir>/episode_results/<chunk>_<episode>/<mode>.json
Run (VLAC .venv, GPU 7):
export VLAC_REPO=/home/vcj9002/jianshu/workspace/code_keliang/Current_Baseline/verify/VLAC
export VLAC_MODEL=/home/vcj9002/jianshu/workspace/code_keliang/Current_Baseline/VLAC/models/VLAC-8b
export PYTHONPATH=$VLAC_REPO
/home/vcj9002/jianshu/workspace/code_keliang/Current_Baseline/VLAC/.venv/bin/python \
run_batch.py --gpu 7
"""
from __future__ import annotations
import argparse
import json
import os
import sys
import tempfile
import time
import traceback
from pathlib import Path
def parse_args():
p = argparse.ArgumentParser(description="VLAC prefix-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("--vlac-repo",
default=os.environ.get(
"VLAC_REPO",
"/home/vcj9002/jianshu/workspace/code_keliang/Current_Baseline/verify/VLAC"),
help="VLAC checkout that makes `evo_vlac` importable (has source .py)")
p.add_argument("--bench-dir",
default="/home/vcj9002/jianshu/workspace/code_keliang/eval/vlac",
help="Dir with benchmark_progress_mark_vlac.py (reused compression)")
p.add_argument("--model-path",
default=os.environ.get(
"VLAC_MODEL",
"/home/vcj9002/jianshu/workspace/code_keliang/Current_Baseline/VLAC/models/VLAC-8b"),
help="VLAC-8b (InternVL2-8B) weights dir")
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 camera as the other baselines")
p.add_argument("--compress-fps", type=int, default=5,
help="VLAC fixed preprocessing fps (do NOT change; model-side)")
p.add_argument("--target-size", type=int, default=448,
help="VLAC fixed preprocessing square size (do NOT change)")
p.add_argument("--batch-num", type=int, default=5,
help="Pairs scored per model batch (VLAC baseline default)")
p.add_argument("--gpu", default=None,
help="GPU id -> CUDA_VISIBLE_DEVICES; model uses cuda:0 within it")
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 / evo_vlac import βββββββββββββββββββ
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
VLAC_REPO = Path(ARGS.vlac_repo).resolve()
BENCH_DIR = Path(ARGS.bench_dir).resolve()
sys.path.insert(0, str(VLAC_REPO))
sys.path.insert(0, str(BENCH_DIR))
os.environ.setdefault("VLAC_REPO", str(VLAC_REPO))
import cv2 # noqa: E402
# Reuse the EXACT baseline compression (5 fps / 448, pyav) and the exact frame
# loader the baseline critic uses internally -- so `seq` matches the VLAC
# baseline frame-for-frame.
from benchmark_progress_mark_vlac import compress_video_with_pyav # noqa: E402
from evo_vlac import GAC_model # noqa: E402
from evo_vlac.utils.video_tool import images_get_from_video # noqa: E402
# init_model hardcodes attn_impl='flash_attn' and VLAC-8b's config forces
# flash_attention_2. When flash_attn is not installed (e.g. this box runs
# torch 2.11+cu13, which has no prebuilt flash-attn wheel) we fall back to
# 'eager'. This is exactly the fallback InternVL itself picks when flash_attn
# is missing: modeling_internvl_chat.py sets llm_config.attn_implementation
# ='eager' and the vision tower uses naive attention (modeling_intern_vit.py).
# It is a numerical attention-kernel choice only; it changes neither the critic
# prompt, the pairwise scoring, nor the critic->value accumulation.
try:
import flash_attn # noqa: F401
_HAS_FLASH = True
except Exception:
_HAS_FLASH = False
if not _HAS_FLASH:
import evo_vlac.utils.model_utils as _mu # noqa: E402
def _force_eager(_orig):
def wrapped(*a, **k):
k["attn_impl"] = "eager"
return _orig(*a, **k)
return wrapped
_mu.get_model_tokenizer = _force_eager(_mu.get_model_tokenizer)
print("[attn] flash_attn not installed -> loading with attn_impl='eager'")
MODEL_PATH = ARGS.model_path
VIDEOS_ROOT = Path(ARGS.videos_root)
CAMERA_DIR = f"observation.images.{ARGS.camera}"
TARGET_SIZE = (ARGS.target_size, ARGS.target_size)
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 = ["dense_all", "stride2", "stride4", "front_dense", "back_dense"]
REFERENCE_MODE = "dense_all"
FRACS = ["1/4", "2/4", "3/4", "end"]
# ββ target frames + sampling paths βββββββββββββββββββββββββββββββββββββββββ
def checkpoints_of(pool_n: int) -> list[int]:
"""seq indices at 1/4, 2/4, 3/4, end (same convention as render/robometer)."""
return [int((pool_n - 1) * k / 4) for k in (1, 2, 3, 4)]
def build_sequence(t: int, mode: str) -> list[int]:
"""Frame indices in [0, t] for `mode`; always starts at 0 and ends at t."""
if t <= 0:
return [0]
if mode == "dense_all":
idx = list(range(0, t + 1))
elif mode == "stride2":
idx = list(range(0, t + 1, 2))
elif mode == "stride4":
idx = list(range(0, t + 1, 4))
elif mode == "front_dense":
half = t // 2
idx = list(range(0, half + 1)) + list(range(half, t + 1, 4))
elif mode == "back_dense":
half = t // 2
idx = list(range(0, half + 1, 4)) + list(range(half, t + 1))
else:
raise ValueError(f"unknown mode: {mode}")
idx = sorted(set(idx))
if idx[0] != 0:
idx = [0] + idx
if idx[-1] != t:
idx = idx + [t]
return idx
# ββ scoring ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def accumulate_path(critic, task, seq, idx, batch_num):
"""Run the baseline pairwise critic over the sampled frame subsequence and
accumulate to a 0-100 value curve. Returns (critic_list, value_curve).
Identical call to the VLAC baseline: ref_image_list=None -> ref_num=0
(zero-shot), skip=1, frame_skip=True, think=False. get_trajectory_critic
scores each adjacent pair [seq[idx[k-1]], seq[idx[k]]] and folds the
increments via critic_to_value_simple."""
subframes = [seq[i] for i in idx]
if len(subframes) < 2:
return [], [0.0]
critic_list, value_curve = critic.get_trajectory_critic(
task=task,
image_list=subframes,
ref_image_list=None,
batch_num=batch_num,
ref_num=0,
think=False,
skip=1,
rich=False,
reverse_eval=False,
frame_skip=True,
)
critic_list = [float(c) for c in critic_list]
value_curve = [float(v) for v in value_curve]
return critic_list, value_curve
# ββ episode enumeration ββββββββββββββββββββββββββββββββββββββββββββββββββββ
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 probe_native(video_path: Path):
"""(native_fps, total_raw_frames) of the source video."""
cap = cv2.VideoCapture(str(video_path))
fps = float(cap.get(cv2.CAP_PROP_FPS) or 0.0)
nfr = int(cap.get(cv2.CAP_PROP_FRAME_COUNT) or 0)
cap.release()
return fps, nfr
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"Repo : {VLAC_REPO}")
print(f"Out : {EP_DIR}")
print(f"Preproc: compress_fps={ARGS.compress_fps} size={TARGET_SIZE} "
f"camera={ARGS.camera} batch_num={ARGS.batch_num}")
print(f"Modes: {MODES}")
print(f"Episodes: total={len(episodes)} todo={len(todo)}")
if not todo:
print("Nothing to do.")
return
critic = GAC_model(tag="critic")
critic.init_model(model_path=str(MODEL_PATH), model_type="internvl2",
device_map="cuda:0")
critic.temperature = 0.5
critic.top_k = 1
critic.set_config()
critic.set_system_prompt()
for i, ep in enumerate(todo, 1):
ep_out = episode_dir(ep)
ep_out.mkdir(parents=True, exist_ok=True)
modes_todo = [m for m in MODES if not (ep_out / f"{m}.json").exists()]
if not modes_todo:
continue
print(f"[{i}/{len(todo)}] {ep['chunk']}/{ep['episode']}", flush=True)
try:
native_fps, total_raw = probe_native(ep["video"])
with tempfile.TemporaryDirectory() as td:
comp_path, comp_fps, orig_idx = compress_video_with_pyav(
ep["video"], Path(td) / "input_fps5_448.mp4",
target_size=TARGET_SIZE, fps=ARGS.compress_fps)
seq = images_get_from_video(str(comp_path))
n = len(seq)
orig_idx = list(orig_idx)[:n]
cps = checkpoints_of(n)
print(f" seq={n} frames (raw={total_raw}, native_fps={native_fps:.2f}, "
f"comp_fps={comp_fps:.2f})", flush=True)
for mode in modes_todo:
mode_path = ep_out / f"{mode}.json"
if mode_path.exists():
continue
t0 = time.time()
checkpoints = {}
values = []
for frac, t in zip(FRACS, cps):
idx = build_sequence(t, mode)
critic_list, value_curve = accumulate_path(
critic, ep["task"], seq, idx, ARGS.batch_num)
value = round(value_curve[-1], 4)
values.append(value)
checkpoints[frac] = {
"target_t": int(t),
"value": value,
"seq_indices": [int(k) for k in idx],
"orig_frames": [int(orig_idx[k]) if k < len(orig_idx) else -1
for k in idx],
"critic_list": [round(c, 6) for c in critic_list],
"value_curve": [round(v, 4) for v in value_curve],
}
payload = {
"model": "VLAC-8b",
"chunk": ep["chunk"], "episode": ep["episode"],
"task": ep["task"], "camera": ARGS.camera,
"mode": mode,
"compress_fps_arg": ARGS.compress_fps,
"compressed_fps": round(float(comp_fps), 4),
"target_size": list(TARGET_SIZE),
"native_fps": round(native_fps, 3),
"total_raw_frames": total_raw,
"pool_n": n,
"sampled_original_frame_indices": [int(x) for x in orig_idx],
"fracs": FRACS,
"target_frames": [int(t) for t in cps],
"values": values,
"checkpoints": checkpoints,
}
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}: 4 targets, values={values} "
f"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()
|