File size: 10,902 Bytes
33569f9 | 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 | """Zero-shot validation for R1 / R3 binary probing assumptions.
Goal: BEFORE committing to a multi-day GRPO training run with binary probing
reward, verify that Qwen2.5-VL actually distinguishes forgery boundaries from
generic "smooth" video positions.
What it tests
-------------
For each test video with multi-segment forgery GT, we probe at three kinds
of boundary points:
- forgery_start : t = GT segment start
- forgery_end : t = GT segment end
- control : a random t far from any GT boundary (Δ_safe seconds)
At each boundary, we run BOTH R1 (3 window probes: pre/post/cross coherence)
and R3 (4 point probes: forgery-classification at t±1).
Output
------
A JSON with per-class distribution statistics (mean / std / quantiles) and a
GO/MARGINAL/NO-GO recommendation per reward variant. Use this to decide
whether to add `binary_probing` to the v10 reward stack.
Run
---
python scripts/probe_zero_shot.py \
--annot_dir /mnt/local-fast/zhangt/annot/annot \
--video_root /mnt/local-fast/zhangt/video \
--preprocessed_data_path /mnt/local-fast/zhangt/forensics_grpo_cache_uniform3584_fps2.0 \
--model_path /mnt/local-fast/zhangt/Qwen2.5-VL-7B-Instruct \
--n_per_class 100 \
--out_json probe_zero_shot_results.json
"""
import argparse
import json
import os
import random
import sys
from collections import defaultdict
import numpy as np
import torch
from tqdm import tqdm
# Allow execution from anywhere inside the repo.
HERE = os.path.dirname(os.path.abspath(__file__))
sys.path.insert(0, os.path.dirname(HERE))
from src.open_r1.data_loader import build_examples, TEST_GENERATORS # noqa: E402
from src.open_r1.binary_prober import BinaryProber, slice_video_by_time # noqa: E402
from src.open_r1.reward import ( # noqa: E402
R1_COHERENCE_QUESTION,
R3_FORGERY_QUESTION,
)
def parse_args():
p = argparse.ArgumentParser()
p.add_argument("--annot_dir", default="/mnt/local-fast/zhangt/annot/annot")
p.add_argument("--video_root", default="/mnt/local-fast/zhangt/video")
p.add_argument("--preprocessed_data_path", required=True,
help="Forensics cache root (output of preprocess_forensics.py)")
p.add_argument("--model_path", required=True,
help="Path to Qwen2.5-VL checkpoint used as frozen prober")
p.add_argument("--n_per_class", type=int, default=100,
help="Cap samples per boundary class (forgery_start/end, control)")
p.add_argument("--delta_s", type=float, default=2.0)
p.add_argument("--point_window_s", type=float, default=1.0)
p.add_argument("--safe_band_s", type=float, default=3.0,
help="Control points must be at least this many seconds "
"from any GT boundary")
p.add_argument("--seed", type=int, default=42)
p.add_argument("--out_json", required=True)
return p.parse_args()
def _enumerate_boundaries(examples, safe_band, rng):
"""Build (example, t_anchor, label) entries for each boundary class."""
by_label = defaultdict(list)
for ex in examples:
if not ex.get("preprocessed_path"):
continue
sol = ex["solution"]
duration = ex["durations"]
if not sol or not duration or duration < 2 * safe_band + 2:
continue
for (s, e) in sol:
if safe_band <= s <= duration - safe_band:
by_label["forgery_start"].append((ex, float(s)))
if safe_band <= e <= duration - safe_band:
by_label["forgery_end"].append((ex, float(e)))
# One control point per video (random, away from any GT boundary).
for _ in range(20):
t = float(rng.uniform(safe_band, duration - safe_band))
far_enough = all(
min(abs(t - s), abs(t - e)) > safe_band for (s, e) in sol
)
if far_enough:
by_label["control"].append((ex, t))
break
return by_label
def _load_video(ex):
"""Return (video_tensor, fps, duration) from a forensics example."""
pdir = ex["preprocessed_path"]
vi_path = os.path.join(pdir, "video_inputs.pt")
vk_path = os.path.join(pdir, "video_kwargs.json")
if not (os.path.exists(vi_path) and os.path.exists(vk_path)):
return None, None, None
vi = torch.load(vi_path, map_location="cpu", weights_only=False)
with open(vk_path) as f:
vk = json.load(f)
if isinstance(vi, list):
vi = vi[0]
fps = vk.get("fps")
if isinstance(fps, list):
fps = fps[0]
return vi, float(fps), float(ex["durations"])
def _r1_window_probes(t, delta, duration):
"""Return [(s_s, s_e, expected), ...] for R1 window probes around `t`."""
return [
(max(0.0, t - delta), t, "yes"), # pre
(t, min(duration, t + delta), "yes"), # post
(max(0.0, t - delta / 2), min(duration, t + delta / 2), "no"), # cross
]
def _r3_point_probes(t, point_window, duration):
half = point_window / 2
return [
(max(0.0, t - 1 - half), max(0.0, t - 1 + half), "no"),
(max(0.0, t + 1 - half), min(duration, t + 1 + half), "yes"),
]
def main():
args = parse_args()
random.seed(args.seed)
rng = np.random.default_rng(args.seed)
examples = build_examples(
annot_dir=args.annot_dir,
video_root=args.video_root,
generators=TEST_GENERATORS,
split_prefix="test",
preprocessed_data_path=args.preprocessed_data_path,
require_video_exists=False,
)
print(f"Loaded {len(examples)} test examples")
by_label = _enumerate_boundaries(examples, args.safe_band_s, rng)
print({k: len(v) for k, v in by_label.items()})
# Cap each class to n_per_class.
for label in list(by_label.keys()):
items = by_label[label]
if args.n_per_class > 0 and len(items) > args.n_per_class:
idx = rng.choice(len(items), args.n_per_class, replace=False)
by_label[label] = [items[i] for i in idx]
print(f" {label}: {len(by_label[label])} kept")
prober = BinaryProber(model_path=args.model_path)
# Result store: results[label][probe_kind][expected] -> list of P(expected)
results: dict = defaultdict(lambda: defaultdict(list))
def _run_probes(label, ex, t):
vi, fps, duration = _load_video(ex)
if vi is None:
return
# R1 probes (3 per boundary).
r1 = _r1_window_probes(t, args.delta_s, duration)
clips, fpss, qs, expecteds, probe_keys = [], [], [], [], []
for (s, e, expected) in r1:
clip = slice_video_by_time(vi, fps, s, e)
if clip is None:
continue
clips.append(clip)
fpss.append(fps)
qs.append(R1_COHERENCE_QUESTION)
expecteds.append(expected)
probe_keys.append(("R1", expected))
# R3 probes (2 per anchor; original spec is 4 around (t1, t2), but
# in zero-shot we treat each boundary point in isolation).
r3 = _r3_point_probes(t, args.point_window_s, duration)
for (s, e, expected) in r3:
clip = slice_video_by_time(vi, fps, s, e)
if clip is None:
continue
clips.append(clip)
fpss.append(fps)
qs.append(R3_FORGERY_QUESTION)
expecteds.append(expected)
probe_keys.append(("R3", expected))
if not clips:
return
# Batch in small chunks to avoid OOM on long videos.
out = []
BS = 8
for i in range(0, len(clips), BS):
out.extend(prober.probe_batch(clips[i:i + BS],
fpss[i:i + BS],
qs[i:i + BS]))
for (kind, expected), (p_yes, p_no) in zip(probe_keys, out):
results[label][f"{kind}_{expected}_Pexp"].append(
p_yes if expected == "yes" else p_no
)
results[label][f"{kind}_{expected}_Pyes"].append(p_yes)
for label, items in by_label.items():
for (ex, t) in tqdm(items, desc=label):
_run_probes(label, ex, t)
# Summarise + decision.
summary = {"args": vars(args), "stats": {}, "decision": {}}
def stat_of(arr):
a = np.asarray(arr)
if a.size == 0:
return {"n": 0}
return {
"n": int(a.size),
"mean": float(a.mean()),
"std": float(a.std()),
"median": float(np.median(a)),
"q25": float(np.percentile(a, 25)),
"q75": float(np.percentile(a, 75)),
}
for label, kinds in results.items():
summary["stats"][label] = {k: stat_of(v) for k, v in kinds.items()}
# R1 decision: cross-window P(no) at forgery boundary vs control.
def _mean(label, key):
vs = results.get(label, {}).get(key, [])
return float(np.mean(vs)) if vs else None
forg_cross_pno = np.mean(
(1 - np.array(results.get("forgery_start", {}).get("R1_no_Pyes", []) or [1])).tolist()
+ (1 - np.array(results.get("forgery_end", {}).get("R1_no_Pyes", []) or [1])).tolist()
)
ctrl_cross_pno = 1 - np.mean(results.get("control", {}).get("R1_no_Pyes", []) or [1.0])
delta_r1 = float(forg_cross_pno - ctrl_cross_pno)
summary["decision"]["R1"] = {
"forgery_cross_P_no_mean": float(forg_cross_pno),
"control_cross_P_no_mean": float(ctrl_cross_pno),
"delta": delta_r1,
"verdict": (
"GO (delta>0.20)" if delta_r1 > 0.20 else
"MARGINAL (0.10<delta<=0.20)" if delta_r1 > 0.10 else
"NO-GO (delta<=0.10)"
),
}
# R3 decision: P(yes-is-forgery) at boundary+1 vs control+1.
forg_yes_after = np.mean(
(results.get("forgery_start", {}).get("R3_yes_Pexp", []) or []) +
(results.get("forgery_end", {}).get("R3_yes_Pexp", []) or [])
)
ctrl_yes_after = np.mean(results.get("control", {}).get("R3_yes_Pexp", []) or [0.0])
delta_r3 = float(forg_yes_after - ctrl_yes_after)
summary["decision"]["R3"] = {
"forgery_inside_P_forged_mean": float(forg_yes_after),
"control_inside_P_forged_mean": float(ctrl_yes_after),
"delta": delta_r3,
"verdict": (
"GO (delta>0.20)" if delta_r3 > 0.20 else
"MARGINAL (0.10<delta<=0.20)" if delta_r3 > 0.10 else
"NO-GO (delta<=0.10)"
),
}
os.makedirs(os.path.dirname(os.path.abspath(args.out_json)) or ".", exist_ok=True)
with open(args.out_json, "w") as f:
json.dump(summary, f, indent=2)
print("\n=== DECISION ===")
print(json.dumps(summary["decision"], indent=2))
print(f"\nFull stats written to {args.out_json}")
if __name__ == "__main__":
main()
|