Spaces:
Running
Running
File size: 15,897 Bytes
da49528 b365ca5 a043ce4 b365ca5 a043ce4 b365ca5 a043ce4 b365ca5 a043ce4 b669e6f 8860868 b365ca5 b669e6f b365ca5 b669e6f 8860868 b365ca5 8860868 b365ca5 b669e6f a043ce4 b669e6f b365ca5 b669e6f a043ce4 b365ca5 b669e6f b365ca5 b669e6f b365ca5 b669e6f b365ca5 b669e6f b365ca5 b669e6f b365ca5 b669e6f 8860868 b365ca5 b669e6f b365ca5 66e1339 b365ca5 66e1339 c18eaea b365ca5 8860868 b669e6f 8860868 b365ca5 b669e6f b365ca5 14c61c5 b365ca5 b669e6f b365ca5 b669e6f b365ca5 b669e6f b365ca5 0b30d62 8860868 0b30d62 8860868 b365ca5 b669e6f b365ca5 b669e6f b365ca5 b669e6f b365ca5 b669e6f b365ca5 8860868 b365ca5 8860868 b365ca5 b669e6f b365ca5 97fd6e5 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 | # patch gradio_client bug β "const" in bool raises TypeError
import gradio_client.utils as _gcu
_original_get_type = _gcu.get_type
def _patched_get_type(schema):
if not isinstance(schema, dict):
return "Any"
return _original_get_type(schema)
_gcu.get_type = _patched_get_type
# now import gradio normally
import gradio as gr
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision.transforms.functional as TF
import math
import numpy as np
from PIL import Image
class PatchEmbedding(nn.Module):
def __init__(self, in_channels=3, embed_dim=192, patch_size=4, img_size=64):
super().__init__()
self.in_channels = in_channels
self.embed_dim = embed_dim
num_patches = (img_size // patch_size) ** 2
self.proj = nn.Conv2d(in_channels, embed_dim, kernel_size=patch_size, stride=patch_size)
self.pos_embed = nn.Parameter(torch.randn(1, num_patches, embed_dim) * 0.02)
nn.init.trunc_normal_(self.proj.weight, std=0.02)
nn.init.zeros_(self.proj.bias)
def forward(self, x):
x = self.proj(x) # [B,3,64,64] β [B,192,16,16]
x = x.flatten(2) # β [B,192,256]
x = x.transpose(1, 2) # β [B,256,192]
x = x + self.pos_embed
return x
class MultiHeadSelfAttention(nn.Module):
def __init__(self, embed_dim=192, num_heads=6, attn_drop=0.1):
super().__init__()
assert embed_dim % num_heads == 0, \
f"embed_dim {embed_dim} must be divisible by num_heads {num_heads}"
self.embed_dim = embed_dim
self.num_heads = num_heads
self.head_dim = embed_dim // num_heads
self.q_proj = nn.Linear(embed_dim, embed_dim)
self.k_proj = nn.Linear(embed_dim, embed_dim)
self.v_proj = nn.Linear(embed_dim, embed_dim)
self.out_proj = nn.Linear(embed_dim, embed_dim)
self.attn_drop = nn.Dropout(attn_drop)
def forward(self, x):
q = self.q_proj(x)
k = self.k_proj(x)
v = self.v_proj(x)
B, N, D = q.shape
q = q.reshape(B, N, self.num_heads, self.head_dim).transpose(1, 2)
k = k.reshape(B, N, self.num_heads, self.head_dim).transpose(1, 2)
v = v.reshape(B, N, self.num_heads, self.head_dim).transpose(1, 2)
scores = (q @ k.transpose(2, 3)) / math.sqrt(self.head_dim)
weights = self.attn_drop(F.softmax(scores, dim=-1))
output = (weights @ v).transpose(1, 2).contiguous()
output = output.reshape(B, N, self.embed_dim)
output = self.out_proj(output)
return output
class GDFN(nn.Module):
def __init__(self, embed_dim, ffn_expansion = 4, grid_size = 16):
super().__init__()
self.embed_dim = embed_dim
self.ffn_expansion = ffn_expansion
self.grid_size = grid_size
self.hidden = embed_dim * ffn_expansion
self.g_proj = nn.Linear(embed_dim, 2*self.hidden)
self.path_1 = nn.Conv2d(self.hidden, self.hidden, 3, padding=1, groups=self.hidden)
self.path_2 = nn.Conv2d(self.hidden, self.hidden, 3, padding=1, groups=self.hidden)
self.out_proj = nn.Linear(self.hidden, embed_dim)
def forward(self, x):
x = self.g_proj(x)
x1, x2 = x.chunk(2, dim=-1)
B, N, C = x1.shape
x1 = x1.transpose(1,2).reshape(B, C, self.grid_size, self.grid_size)
x2 = x2.transpose(1,2).reshape(B, C, self.grid_size, self.grid_size)
x1 = self.path_1(x1)
x2 = self.path_2(x2)
x1 = x1.flatten(2).transpose(1,2)
x2 = x2.flatten(2).transpose(1,2)
x = x1 * F.gelu(x2)
x = self.out_proj(x)
return x
class TransformerBlock(nn.Module):
def __init__(self, embed_dim=192, num_heads=6, attn_drop=0.1):
super().__init__()
self.norm1 = nn.LayerNorm(embed_dim)
self.norm2 = nn.LayerNorm(embed_dim)
self.attn = MultiHeadSelfAttention(embed_dim, num_heads, attn_drop)
self.ffn = GDFN(embed_dim)
self.gamma1 = nn.Parameter(1e-4 * torch.ones(embed_dim))
self.gamma2 = nn.Parameter(1e-4 * torch.ones(embed_dim))
def forward(self, x):
x = x + self.gamma1 * self.attn(self.norm1(x))
x = x + self.gamma2 * self.ffn(self.norm2(x))
return x
class ImageSRTransformer(nn.Module):
def __init__(self,
embed_dims=[192, 256, 288, 384],
num_heads=[6, 8, 6, 8],
depths=[3, 3, 3, 3],
patch_size=4,
img_size=64):
super().__init__()
self.embed_dims = embed_dims
self.grid_size = img_size // patch_size
self.patch_embed = PatchEmbedding(embed_dim=embed_dims[0], patch_size=patch_size, img_size=img_size)
self.stages = nn.ModuleList([
nn.ModuleList([TransformerBlock(embed_dims[i], num_heads[i])
for _ in range(depths[i])]) for i in range(len(embed_dims))])
self.projections = nn.ModuleList([
nn.Linear(embed_dims[i], embed_dims[i+1])
for i in range(len(embed_dims) - 1)
])
self.head = nn.Sequential(
nn.Conv2d(1120, 192, 3, padding=1),
nn.LeakyReLU(0.2),
nn.Conv2d(192, 768, 3, padding=1),
nn.PixelShuffle(2),
nn.Conv2d(192, 768, 3, padding=1),
nn.PixelShuffle(2),
nn.Conv2d(192, 768, 3, padding=1),
nn.PixelShuffle(2),
nn.Conv2d(192, 768, 3, padding=1),
nn.PixelShuffle(2),
nn.Conv2d(192, 3, 3, padding=1)
)
def forward(self, x):
lr = x
x = self.patch_embed(x)
stage_outputs = []
for i, stage in enumerate(self.stages):
for block in stage:
x = block(x)
stage_outputs.append(x)
if i < len(self.stages) -1:
x = self.projections[i](x)
x = torch.cat(stage_outputs, dim=-1)
B, N, C = x.shape
x = x.transpose(1,2).reshape(B, C, self.grid_size, self.grid_size)
x = self.head(x)
base = F.interpolate(lr, size=(256, 256), mode="bicubic", align_corners=False)
return base + x
def psnr(pred, target):
mse = torch.mean((pred - target) ** 2)
return 10 * torch.log10(1.0 / (mse + 1e-8))
def ssim(pred, target, window_size=11):
C1 = 0.01 ** 2
C2 = 0.03 ** 2
mu1 = F.avg_pool2d(pred, window_size, 1, window_size // 2)
mu2 = F.avg_pool2d(target, window_size, 1, window_size // 2)
mu1_sq = mu1 ** 2
mu2_sq = mu2 ** 2
mu1_mu2 = mu1 * mu2
s1 = F.avg_pool2d(pred ** 2, window_size, 1, window_size // 2) - mu1_sq
s2 = F.avg_pool2d(target ** 2, window_size, 1, window_size // 2) - mu2_sq
s12 = F.avg_pool2d(pred * target, window_size, 1, window_size // 2) - mu1_mu2
num = (2 * mu1_mu2 + C1) * (2 * s12 + C2)
den = (mu1_sq + mu2_sq + C1) * (s1 + s2 + C2)
return (num / den).mean().item()
# ββ Load model ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
model = ImageSRTransformer()
checkpoint = torch.load(
"sr_best_v4_resumed.pt",
map_location="cpu",
weights_only=False
)
model.load_state_dict(checkpoint["model_state_dict"])
model.eval()
print(f"Model loaded β val PSNR: {checkpoint['val_psnr']:.2f} dB")
# ββ Inference βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def run_sr(img_pil):
device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device)
img_pil = img_pil.convert("RGB")
w, h = img_pil.size
# ensure minimum 256Γ256 for centre crop
if w < 256 or h < 256:
scale_up = max(256 / w, 256 / h)
img_pil = img_pil.resize(
(int(w * scale_up), int(h * scale_up)), Image.BICUBIC)
w, h = img_pil.size
# centre crop 256Γ256 β ground truth
left = (w - 256) // 2
top = (h - 256) // 2
gt = img_pil.crop((left, top, left + 256, top + 256))
# bicubic downscale β 64Γ64 LR
lr = gt.resize((64, 64), Image.BICUBIC)
lr_t = TF.to_tensor(lr).unsqueeze(0).to(device)
gt_t = TF.to_tensor(gt).unsqueeze(0).to(device)
with torch.no_grad():
with torch.autocast(device_type="cuda", dtype=torch.bfloat16,
enabled=(device == "cuda")):
sr_t = model(lr_t)
sr_t = sr_t.float().clamp(0, 1)
lr_display_t = F.interpolate(
lr_t.float(), size=(256, 256), mode="bilinear", align_corners=False)
psnr_lr = psnr(lr_display_t, gt_t).item()
ssim_lr = ssim(lr_display_t, gt_t)
psnr_sr = psnr(sr_t, gt_t).item()
ssim_sr = ssim(sr_t, gt_t)
def to_pil(t):
return TF.to_pil_image(t.squeeze(0).cpu())
metrics = (
f"**LR baseline** β PSNR: {psnr_lr:.2f} dB | SSIM: {ssim_lr:.4f}\n\n"
f"**SR output** β PSNR: {psnr_sr:.2f} dB | SSIM: {ssim_sr:.4f}\n\n"
f"**Improvement** β ΞPSNR: +{psnr_sr - psnr_lr:.2f} dB | "
f"ΞSSIM: +{ssim_sr - ssim_lr:.4f}"
)
return to_pil(lr_display_t), to_pil(sr_t), gt, metrics
# ββ Example images ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# 6 examples chosen to showcase V4 strengths:
EXAMPLES = [
["examples/urban.png"],
["examples/aerial.png"],
["examples/architecture.png"],
["examples/nature.png"],
["examples/portrait.png"],
["examples/texture.png"]
]
# ββ CSS βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
CSS = """
body, .gradio-container {
background: #0a0c10 !important;
color: #c9d1d9 !important;
font-family: 'Inter', system-ui, sans-serif !important;
}
.header-block {
text-align: center;
padding: 2rem 1rem 1rem;
border-bottom: 1px solid #1e2128;
margin-bottom: 1.5rem;
}
.header-title {
font-size: 1.8rem;
font-weight: 700;
color: #2dd4bf;
letter-spacing: -0.02em;
margin-bottom: 0.25rem;
}
.header-sub {
font-size: 0.85rem;
color: #6e7681;
}
.hint-text {
font-size: 0.78rem;
color: #6e7681;
margin-top: 8px;
line-height: 1.5;
}
.metric-box textarea, .metric-box .prose {
background: #111318 !important;
border: 1px solid #1e2128 !important;
color: #c9d1d9 !important;
border-radius: 8px !important;
font-family: 'JetBrains Mono', monospace !important;
font-size: 0.82rem !important;
padding: 12px !important;
}
.run-btn {
background: #2dd4bf !important;
color: #0a0c10 !important;
border: none !important;
font-weight: 700 !important;
font-size: 0.9rem !important;
border-radius: 8px !important;
padding: 0.6rem 2rem !important;
}
.run-btn:hover { background: #5eead4 !important; }
.accordion {
background: #111318 !important;
border: 1px solid #1e2128 !important;
border-radius: 8px !important;
margin-top: 1.5rem !important;
}
.accordion .label-wrap {
color: #6e7681 !important;
font-size: 0.8rem !important;
letter-spacing: 0.08em !important;
text-transform: uppercase !important;
}
.image-frame img {
border-radius: 8px !important;
border: 1px solid #1e2128 !important;
}
"""
# ββ Architecture info (collapsible) ββββββββββββββββββββββββββββββββββββββββββ
ARCH_INFO = """
> "Isotropic constant-resolution hierarchical ViT with inter-stage dense feature aggregation β eliminating spatial bottlenecks while preserving coordinate integrity throughout all processing stages."
### Key architectural decisions
**Isotropic token grid** β constant 16Γ16 spatial resolution across all 4 transformer stages. Zero patch merging, zero token downsampling. Every token maps to the same 4Γ4 pixel region from input to output.
**Hierarchical embed dims [192 β 256 β 288 β 384]** β representational capacity scales with feature complexity. Early stages learn local edges and textures (192-dim is sufficient). Deep stages reason about global scene semantics (384-dim is necessary).
**Inter-stage macro concatenation** β outputs from all 4 stages concatenated directly to the reconstruction head: `cat([h1, h2, h3, h4]) β [B, 256, 1120]`. The head receives low-level edge maps (h1) and high-level semantic context (h4) simultaneously.
**GDFN feed-forward** β replaces standard MLPs with Gated Depthwise Feed-Forward Networks. Each token sees its 3Γ3 spatial neighborhood during the MLP step. Local spatial context injected at every attention layer.
**Bilinear skip connection** β `output = F.interpolate(lr, 256Γ256) + vit_residual`. Model learns residual correction only, not full reconstruction from scratch.
### Results
| Benchmark | Avg PSNR | Avg SSIM |
|-----------|------|------|
| DIV2K validation | 25.20 dB | 0.8298 |
"""
# ββ Gradio UI βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
with gr.Blocks(css=CSS, title="Dense-Iso-ViT SR") as demo:
gr.HTML("""
<div class="header-block">
<div class="header-title">Dense-Iso-ViT</div>
<div class="header-sub">
Constant-Resolution Hierarchical Vision Transformer for Γ4 Image Super-Resolution
</div>
</div>
""")
with gr.Row():
with gr.Column(scale=1):
input_img = gr.Image(
type="pil",
label="Upload any image",
elem_classes=["image-frame"],
)
run_btn = gr.Button(
"Run Γ4 Super-Resolution",
variant="primary",
elem_classes=["run-btn"],
)
gr.Markdown(
"π‘ For best results, upload a low-resolution image "
"(256Γ256 or smaller). The model upscales it Γ4.",
elem_classes=["hint-text"],
)
with gr.Column(scale=3):
with gr.Row():
lr_out = gr.Image(
label="LR Input (bilinear upscaled for display)",
elem_classes=["image-frame"],
)
sr_out = gr.Image(
label="SR Output β Dense-Iso-ViT",
elem_classes=["image-frame"],
)
gt_out = gr.Image(
label="Ground Truth (original crop)",
elem_classes=["image-frame"],
)
metrics_out = gr.Markdown(elem_classes=["metric-box"])
gr.Examples(
examples=EXAMPLES,
inputs=[input_img],
label="Examples β showing V4 strengths",
)
with gr.Accordion(
"Architecture details β Dense-Iso-ViT",
open=False,
elem_classes=["accordion"],
):
gr.Markdown(ARCH_INFO)
run_btn.click(
fn=run_sr,
inputs=[input_img],
outputs=[lr_out, sr_out, gt_out, metrics_out],
)
input_img.change(
fn=run_sr,
inputs=[input_img],
outputs=[lr_out, sr_out, gt_out, metrics_out],
)
if __name__ == "__main__":
demo.launch(show_api=False) |