File size: 5,645 Bytes
b7d26f4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/bin/bash
# Run pMF reproduction inside HF Job
set -e

VARIANT=${1:-pMF-H/16}
NUM_SAMPLES=${2:-50000}

echo "=== pMF Reproduction: $VARIANT ==="
echo "Samples: $NUM_SAMPLES"

pip install -q torch torchvision --index-url https://download.pytorch.org/whl/cu118
pip install -q diffusers transformers scipy numpy clean-fid Pillow huggingface_hub

# Map variant to subfolder
case "$VARIANT" in
    pMF-B/16) SF=pMF-B-16; FID_URL=https://huggingface.co/Lyy0725/pMF/resolve/main/imagenet_256_fid_stats.npz ;;
    pMF-L/16) SF=pMF-L-16; FID_URL=https://huggingface.co/Lyy0725/pMF/resolve/main/imagenet_256_fid_stats.npz ;;
    pMF-H/16) SF=pMF-H-16; FID_URL=https://huggingface.co/Lyy0725/pMF/resolve/main/imagenet_256_fid_stats.npz ;;
    pMF-B/32) SF=pMF-B-32; FID_URL=https://huggingface.co/Lyy0725/pMF/resolve/main/imagenet_512_fid_stats.npz ;;
    pMF-L/32) SF=pMF-L-32; FID_URL=https://huggingface.co/Lyy0725/pMF/resolve/main/imagenet_512_fid_stats.npz ;;
    pMF-H/32) SF=pMF-H-32; FID_URL=https://huggingface.co/Lyy0725/pMF/resolve/main/imagenet_512_fid_stats.npz ;;
    *) echo "Unknown variant: $VARIANT"; exit 1 ;;
esac

# Download variant subfolder locally
echo "Downloading variant $SF from BiliSakura/pMF-diffusers..."
python3 -c "
from huggingface_hub import snapshot_download
snapshot_download('BiliSakura/pMF-diffusers', allow_patterns=f'$SF/**', local_dir='./model')
"
MODEL_DIR="./model/$SF"
echo "Model directory: $MODEL_DIR"
ls -la "$MODEL_DIR"

# Download FID stats
echo "Downloading FID stats..."
curl -sL "$FID_URL" -o /tmp/stats.npz
echo "FID stats: $(wc -c < /tmp/stats.npz) bytes"

# Run reproduction
python3 << 'PYEOF'
import json, os, time
import numpy as np
import torch
from pathlib import Path
from diffusers import DiffusionPipeline
from scipy import linalg

VARIANT = os.environ.get("VARIANT", "pMF-H/16")
NUM = int(os.environ.get("NUM_SAMPLES", "50000"))
MODEL_DIR = os.environ.get("MODEL_DIR")
FID_STATS = "/tmp/stats.npz"

def compute_fid_from_stats(mu1, sigma1, mu2, sigma2, eps=1e-6):
    diff = mu1 - mu2
    covmean, _ = linalg.sqrtm(sigma1.dot(sigma2), disp=False)
    if not np.isfinite(covmean).all():
        offset = np.eye(sigma1.shape[0]) * eps
        covmean = linalg.sqrtm((sigma1+offset).dot(sigma2+offset), disp=False)
    if np.iscomplexobj(covmean):
        covmean = covmean.real
    return float(diff.dot(diff) + np.trace(sigma1 + sigma2 - 2.0 * covmean))

def compute_fid(imgs, ref_path, bs=100):
    from cleanfid.features import build_feature_extractor, get_image_features
    ref = np.load(ref_path)
    mu_ref, sigma_ref = ref["mu"], ref["sigma"]
    device = "cuda" if torch.cuda.is_available() else "cpu"
    model = build_feature_extractor("inception_v3", device=device)
    model.eval()
    feats = []
    for i in range(0, len(imgs), bs):
        b = imgs[i:i+bs]
        bt = torch.from_numpy(b).permute(0,3,1,2).float()/255.0
        feats.append(get_image_features(model, bt.to(device)).cpu().numpy())
    all_f = np.concatenate(feats, axis=0)
    mu = np.mean(all_f, axis=0)
    sg = np.cov(all_f, rowvar=False)
    return compute_fid_from_stats(mu, sg, mu_ref, sigma_ref)

cfg_params = {
    "pMF-B/16":{"gs":7.5,"tmi":0.1,"tma":0.8,"ns":1.0},
    "pMF-L/16":{"gs":7.0,"tmi":0.2,"tma":0.7,"ns":1.0},
    "pMF-H/16":{"gs":7.0,"tmi":0.2,"tma":0.6,"ns":2.0},
    "pMF-B/32":{"gs":6.5,"tmi":0.1,"tma":0.7,"ns":2.0},
    "pMF-L/32":{"gs":7.5,"tmi":0.2,"tma":0.6,"ns":4.0},
    "pMF-H/32":{"gs":5.5,"tmi":0.1,"tma":0.6,"ns":4.0},
}
paper = {"pMF-B/16":3.12,"pMF-L/16":2.52,"pMF-H/16":2.22,"pMF-B/32":3.70,"pMF-L/32":2.75,"pMF-H/32":2.48}

device = "cuda" if torch.cuda.is_available() else "cpu"
print(f"Device: {device}", flush=True)
c = cfg_params[VARIANT]
os.makedirs("outputs", exist_ok=True)

t0 = time.time()
model_path = Path(MODEL_DIR)
print(f"Loading from {model_path}...", flush=True)
pipe = DiffusionPipeline.from_pretrained(
    str(model_path),
    local_files_only=True,
    custom_pipeline=str(model_path / "pipeline.py"),
    trust_remote_code=True,
    torch_dtype=torch.float16,
).to(device)
pipe.set_progress_bar_config(disable=True)
print(f"Loaded in {time.time()-t0:.1f}s", flush=True)

classes = [i%1000 for i in range(NUM)]
gen = torch.Generator(device=device)

t1 = time.time()
all_imgs = []
for i in range(0, NUM, 50):
    gen.manual_seed(i)
    with torch.no_grad():
        imgs = pipe(
            class_labels=classes[i:i+50],
            num_inference_steps=1,
            guidance_scale=c["gs"],
            guidance_interval_min=c["tmi"],
            guidance_interval_max=c["tma"],
            noise_scale=c["ns"],
            generator=gen,
        ).images
    all_imgs.extend([np.array(img) for img in imgs])
    if (i+50)%1000==0 or i+50>=NUM:
        print(f"  {len(all_imgs)}/{NUM} ({time.time()-t1:.0f}s)", flush=True)

imgs_np = np.stack(all_imgs, axis=0)
gt = time.time()-t1
print(f"Gen: {gt:.0f}s ({gt/NUM:.4f}s/img)", flush=True)

npz_path = f"outputs/{VARIANT.replace(chr(47),chr(95))}_images.npz"
np.savez_compressed(npz_path, images=imgs_np)
print(f"Saved {npz_path}", flush=True)

t2 = time.time()
fid = round(compute_fid(imgs_np, FID_STATS), 4)
ft = time.time()-t2
print(f"FID: {ft:.0f}s", flush=True)

r = {
    "variant":VARIANT, "num_samples":NUM, "fid":fid,
    "fid_paper":paper[VARIANT], "fid_diff":round(fid-paper[VARIANT],4),
    "gen_time_s":round(gt,1), "fid_time_s":round(ft,1),
    "hardware":torch.cuda.get_device_name(0) if torch.cuda.is_available() else "cpu",
}
print(json.dumps(r,indent=2), flush=True)
rp = f"outputs/{VARIANT.replace(chr(47),chr(95))}_result.json"
with open(rp,"w") as f: json.dump(r,f,indent=2)
PYEOF