File size: 8,409 Bytes
63f8a47 | 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 | """Evaluate a trained V-PVP head on the AIGVDBench Open-Source split.
Computes per-generator AUC, mean-of-generators AUC (mga AUC), and overall
AUC on the test set. Each backbone is loaded once (frozen, fp16 autocast),
patch features ``z`` of shape ``(B, P, T', D)`` are extracted, the velocity
field ``v = z[:, :, 1:] - z[:, :, :-1]`` is computed, and the head produces
a scalar logit per video.
Usage:
python -m src.eval_opensource \
--backbone videomae_k400 \
--backbone_weights <path-or-hub-id> \
--head_ckpt checkpoints/pvp_heads/videomae_k400_vpvp.pt \
--data_root <path-to-aigvdbench> \
--num_frames 16 \
--batch_size 32 \
--out_json results/videomae_open_source.json
Expected mga AUC on the Open-Source split with the released checkpoint
(20 generators, 3000 real + 60000 fake videos):
videomae_k400 -> 0.9528 (V-PVP row of Table 1)
"""
import argparse
import json
import os
import sys
import time
import numpy as np
import torch
import torch.nn as nn
from sklearn.metrics import roc_auc_score
from torch.utils.data import DataLoader
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
from backbones import BACKBONE_REGISTRY
from dataset import OpenSourceConfig, OpenSourceTest, collate
from pvp_head import PVPHead
# Frame count the released checkpoint was trained at.
DEFAULT_NUM_FRAMES = {
"videomae_k400": 16,
}
# Dataset-side normalization applied by ``OpenSourceTest`` (mean 0.45 /
# std 0.225). The eval loop undoes it and re-applies the backbone's own
# statistics below; this is the exact preprocessing path the released
# checkpoint was trained and evaluated with.
DS_MEAN = (0.45, 0.45, 0.45)
DS_STD = (0.225, 0.225, 0.225)
def renormalize(clips, ds_mean, ds_std, bb_mean, bb_std):
"""Undo dataset normalization, re-apply backbone's. clips: (B, C, T, H, W)."""
x = clips * ds_std + ds_mean
return (x - bb_mean) / bb_std
def compute_metrics(scores, labels, cats, threshold=0.5):
"""mga AUC = mean-of-generators AUC over the per-(real, generator) pools."""
cats = np.asarray(cats)
labels = np.asarray(labels)
scores = np.asarray(scores)
real = (labels == 0)
preds = (scores > threshold).astype(int)
auc_per_gen, acc_per_gen = {}, {}
for c in sorted(set(cats[~real].tolist())):
gen_mask = (cats == c) & (labels == 1)
if gen_mask.sum() < 2 or real.sum() < 2:
continue
y = np.r_[np.zeros(real.sum()), np.ones(gen_mask.sum())]
s = np.r_[scores[real], scores[gen_mask]]
p = np.r_[preds[real], preds[gen_mask]]
auc_per_gen[c] = float(roc_auc_score(y, s))
acc_per_gen[c] = float((p == y).mean())
return {
"auc_per_gen": auc_per_gen,
"acc_per_gen": acc_per_gen,
"mga_auc": float(np.mean(list(auc_per_gen.values()))),
"mga_acc": float(np.mean(list(acc_per_gen.values()))),
"overall_auc": float(roc_auc_score(labels, scores)),
"overall_acc": float((preds == labels).mean()),
"n_real": int(real.sum()),
"n_fake": int((~real).sum()),
}
@torch.no_grad()
def evaluate(backbone, head, loader, device, ds_norm, bb_norm):
head.eval()
backbone.eval()
scores_all, labels_all, cats_all = [], [], []
n_done = 0
t0 = time.time()
for clips, labels, cats in loader:
clips = clips.to(device, non_blocking=True)
x = renormalize(clips, *ds_norm, *bb_norm)
with torch.autocast(device_type="cuda", dtype=torch.float16):
z = backbone.extract_patch_features(x)
z = z.to(torch.float32)
if z.size(2) < 2:
raise RuntimeError(f"backbone returned T'={z.size(2)} < 2; need >=2 frames")
v = z[:, :, 1:, :] - z[:, :, :-1, :]
logit = head(z, v)
scores_all.append(torch.sigmoid(logit).cpu().numpy())
labels_all.append(labels.numpy())
cats_all.extend(cats)
n_done += clips.size(0)
if n_done % 1024 == 0:
print(f" [{n_done}] {n_done/(time.time()-t0):.1f} v/s", flush=True)
return (np.concatenate(scores_all),
np.concatenate(labels_all),
np.array(cats_all))
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--backbone", required=True,
choices=list(BACKBONE_REGISTRY.keys()))
ap.add_argument("--backbone_weights", required=True,
help="path (or HuggingFace repo id) of the K400 backbone weights")
ap.add_argument("--head_ckpt", required=True,
help="path to the V-PVP head state_dict (.pt)")
ap.add_argument("--data_root", required=True,
help="AIGVDBench root containing Real/ and OpenSource/ folders")
ap.add_argument("--num_frames", type=int, default=0,
help="0 -> backbone default (16 for VideoMAE)")
ap.add_argument("--batch_size", type=int, default=32)
ap.add_argument("--num_workers", type=int, default=8)
ap.add_argument("--threshold", type=float, default=0.5)
ap.add_argument("--out_json", default="",
help="if set, write per-generator AUC + summary to this JSON")
args = ap.parse_args()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
nf = args.num_frames or DEFAULT_NUM_FRAMES[args.backbone]
print(f"backbone={args.backbone} num_frames={nf} bs={args.batch_size}", flush=True)
print(f"backbone_weights={args.backbone_weights}", flush=True)
print(f"head_ckpt={args.head_ckpt}", flush=True)
# ---- backbone (frozen) ----
bb_cls = BACKBONE_REGISTRY[args.backbone]
backbone = bb_cls(pretrained_model=args.backbone_weights,
num_frames=nf, img_size=224).to(device).eval()
for p in backbone.parameters():
p.requires_grad = False
D = backbone.embed_dim
print(f" P={backbone.num_patches_spatial} D={D} "
f"input_mean={backbone.input_mean} input_std={backbone.input_std}", flush=True)
# ---- head ----
head = PVPHead("velocity_gated", D=D).to(device)
sd = torch.load(args.head_ckpt, map_location=device)
if isinstance(sd, dict) and "state_dict" in sd:
sd = sd["state_dict"]
head.load_state_dict(sd)
head.eval()
n_head = sum(p.numel() for p in head.parameters())
print(f" head params: {n_head:,}", flush=True)
# ---- data ----
cfg = OpenSourceConfig(data_root=args.data_root, num_frames=nf,
crop_size=224, max_per_category=0)
ds = OpenSourceTest(cfg)
print(f" test N={len(ds)}", flush=True)
loader = DataLoader(ds, batch_size=args.batch_size, shuffle=False,
num_workers=args.num_workers, pin_memory=True,
collate_fn=collate)
# ---- normalization tensors ----
ds_mean = torch.tensor(DS_MEAN, device=device).view(1, 3, 1, 1, 1)
ds_std = torch.tensor(DS_STD, device=device).view(1, 3, 1, 1, 1)
bb_mean = torch.tensor(backbone.input_mean, device=device).view(1, 3, 1, 1, 1)
bb_std = torch.tensor(backbone.input_std, device=device).view(1, 3, 1, 1, 1)
# ---- run ----
t0 = time.time()
scores, labels, cats = evaluate(backbone, head, loader, device,
(ds_mean, ds_std), (bb_mean, bb_std))
elapsed = (time.time() - t0) / 60
print(f" inference done in {elapsed:.1f} min ({len(scores)} videos)", flush=True)
m = compute_metrics(scores, labels, cats, threshold=args.threshold)
print(f"\nOpen-Source split results ({m['n_real']} real / {m['n_fake']} fake):")
print(f" mga AUC = {m['mga_auc']:.4f}")
print(f" mga ACC = {m['mga_acc']:.4f}")
print(f" overall AUC = {m['overall_auc']:.4f}")
print(f" overall ACC = {m['overall_acc']:.4f}")
print("\nPer-generator AUC:")
for c, a in m["auc_per_gen"].items():
print(f" {c:<32s} {a*100:6.2f}")
if args.out_json:
os.makedirs(os.path.dirname(args.out_json) or ".", exist_ok=True)
out = {
"backbone": args.backbone,
"num_frames": nf,
"head_ckpt": args.head_ckpt,
**m,
}
with open(args.out_json, "w") as f:
json.dump(out, f, indent=2)
print(f"\nsaved {args.out_json}")
if __name__ == "__main__":
main()
|