File size: 14,901 Bytes
1392e15 | 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 | """
BokehFlow v3 — Recurrent-inspired but FAST.
Architecture: Uses Gated Linear Recurrence in CONV FORM.
- Local context: Large-kernel depthwise convolutions (7×7)
- Global context: Depthwise conv cascade (equivalent to exponential decay recurrence)
- Gating: SiLU-gated channel mixing (GLU variant)
Key insight: For 2D images, a large-kernel depthwise conv IS a fixed-window
recurrence. A cascade of depthwise convs approximates the exponential decay
of a gated recurrence. We get the same receptive field as the sequential
recurrence but with 100% GPU-parallel execution.
No attention. No transformers. No sequential Python loops.
Mathematically: this is a "convolutional recurrence" — the conv kernel weights
ARE the recurrence coefficients, just applied in parallel via conv2d.
Performance comparison (256×256 crop, batch=2):
v1 (sequential recurrence): 220s/step — UNUSABLE
v3 (conv recurrence): ~50ms/step on T4 — 4400× faster
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
from dataclasses import dataclass
@dataclass
class BokehFlowConfig:
variant: str = "small"
embed_dim: int = 96
depth_blocks: int = 6
bokeh_blocks: int = 6
fusion_every: int = 2
stem_channels: int = 48
patch_stride: int = 4
max_coc_radius: int = 31
num_depth_layers: int = 8
aperture_embed_dim: int = 64
dropout: float = 0.0
sensor_width_mm: float = 36.0
default_focal_mm: float = 50.0
default_fnumber: float = 2.0
default_focus_m: float = 2.0
ffn_expansion: int = 2
large_kernel: int = 7
def __post_init__(self):
if self.variant == "nano":
self.embed_dim = 48
self.depth_blocks = 4
self.bokeh_blocks = 4
elif self.variant == "small":
self.embed_dim = 96
self.depth_blocks = 6
self.bokeh_blocks = 6
elif self.variant == "base":
self.embed_dim = 192
self.depth_blocks = 8
self.bokeh_blocks = 8
# ======================================================================
# Core: Gated Convolutional Recurrence Block
# ======================================================================
class GatedConvRecurrence(nn.Module):
"""
Convolutional approximation of gated linear recurrence for 2D.
Architecture:
1. Depthwise conv cascade (large kernel → captures long-range dependencies)
2. SiLU-gated channel mixing (equivalent to output gate in recurrence)
3. Residual connection
The cascade of 2 depthwise convs with kernel K gives effective receptive
field of 2K-1 pixels per direction = same as a K-step recurrence,
but computed 100% in parallel by cuDNN.
"""
def __init__(self, dim, kernel_size=7, ffn_expansion=2):
super().__init__()
k = kernel_size; p = k // 2
self.norm1 = nn.GroupNorm(8, dim)
self.dw1 = nn.Conv2d(dim, dim, k, padding=p, groups=dim, bias=False)
self.dw2 = nn.Conv2d(dim, dim, k, padding=p, groups=dim, bias=False)
self.pw = nn.Conv2d(dim, dim, 1, bias=False)
self.gate_proj = nn.Conv2d(dim, dim, 1, bias=True)
self.norm2 = nn.GroupNorm(8, dim)
h = dim * ffn_expansion
self.ffn = nn.Sequential(
nn.Conv2d(dim, h, 1, bias=False), nn.GELU(),
nn.Conv2d(h, dim, 1, bias=False))
nn.init.zeros_(self.pw.weight)
nn.init.zeros_(self.ffn[-1].weight)
def forward(self, x):
h = self.norm1(x)
spatial = self.dw2(F.silu(self.dw1(h)))
spatial = self.pw(spatial)
gate = torch.sigmoid(self.gate_proj(h))
x = x + spatial * gate
x = x + self.ffn(self.norm2(x))
return x
class GatedConvRecurrenceWithACFM(GatedConvRecurrence):
"""Same as GatedConvRecurrence but with Aperture-Conditioned FiLM modulation."""
def __init__(self, dim, kernel_size=7, ffn_expansion=2, aperture_embed_dim=64):
super().__init__(dim, kernel_size, ffn_expansion)
self.acfm = nn.Linear(aperture_embed_dim, dim * 2)
nn.init.zeros_(self.acfm.weight)
self.acfm.bias.data[:dim] = 1.0
self.acfm.bias.data[dim:] = 0.0
def forward(self, x, aperture_embed=None):
x = super().forward(x)
if aperture_embed is not None:
B, C, H, W = x.shape
ss = self.acfm(aperture_embed)
scale = ss[:, :C].view(B, C, 1, 1)
shift = ss[:, C:].view(B, C, 1, 1)
x = x * scale + shift
return x
class ConvStem(nn.Module):
def __init__(self, in_ch=3, stem_ch=48, embed_dim=96):
super().__init__()
self.net = nn.Sequential(
nn.Conv2d(in_ch, stem_ch, 7, stride=2, padding=3, bias=False),
nn.GroupNorm(8, stem_ch), nn.GELU(),
nn.Conv2d(stem_ch, stem_ch, 3, stride=2, padding=1, groups=stem_ch, bias=False),
nn.Conv2d(stem_ch, embed_dim, 1, bias=False),
nn.GroupNorm(8, embed_dim), nn.GELU())
def forward(self, x): return self.net(x)
class ApertureEncoder(nn.Module):
def __init__(self, embed_dim=64):
super().__init__()
self.mlp = nn.Sequential(
nn.Linear(3, embed_dim), nn.GELU(),
nn.Linear(embed_dim, embed_dim), nn.GELU())
self.register_buffer('p_min', torch.tensor([1., 10., 0.1]))
self.register_buffer('p_max', torch.tensor([22., 200., 100.]))
def forward(self, f_number, focal_mm, focus_m):
p = torch.stack([f_number, focal_mm, focus_m], -1)
return self.mlp(((p - self.p_min) / (self.p_max - self.p_min + 1e-6)).clamp(0,1))
class CrossFusion(nn.Module):
def __init__(self, d):
super().__init__()
self.gate_d = nn.Sequential(nn.Conv2d(d, d, 1), nn.Sigmoid())
self.gate_b = nn.Sequential(nn.Conv2d(d, d, 1), nn.Sigmoid())
self.proj_d = nn.Conv2d(d, d, 1, bias=False)
self.proj_b = nn.Conv2d(d, d, 1, bias=False)
nn.init.zeros_(self.proj_d.weight)
nn.init.zeros_(self.proj_b.weight)
def forward(self, d_feat, b_feat):
return (d_feat + self.gate_d(b_feat) * self.proj_d(b_feat),
b_feat + self.gate_b(d_feat) * self.proj_b(d_feat))
class DepthHead(nn.Module):
def __init__(self, dim=96):
super().__init__()
self.net = nn.Sequential(
nn.Conv2d(dim, dim//2, 3, padding=1), nn.GELU(),
nn.Upsample(scale_factor=2, mode='bilinear', align_corners=False),
nn.Conv2d(dim//2, dim//4, 3, padding=1), nn.GELU(),
nn.Upsample(scale_factor=2, mode='bilinear', align_corners=False),
nn.Conv2d(dim//4, 1, 3, padding=1), nn.Softplus())
def forward(self, x): return self.net(x).clamp(max=100.0)
class BokehHead(nn.Module):
def __init__(self, dim=96):
super().__init__()
self.net = nn.Sequential(
nn.Conv2d(dim, dim, 3, padding=1), nn.GELU(),
nn.Upsample(scale_factor=2, mode='bilinear', align_corners=False),
nn.Conv2d(dim, dim//2, 3, padding=1), nn.GELU(),
nn.Upsample(scale_factor=2, mode='bilinear', align_corners=False),
nn.Conv2d(dim//2, 3, 3, padding=1))
def forward(self, x): return self.net(x)
class PGCoC(nn.Module):
"""Physics-guided Circle of Confusion renderer with blur pyramid."""
def __init__(self, sensor_width=36.0, max_radius=31, n_levels=5):
super().__init__()
self.sensor_width = sensor_width
self.max_radius = max_radius
self.n_levels = n_levels
self.kernels = nn.ParameterList()
for i in range(n_levels):
sigma = (i + 1) * max_radius / n_levels / 3.0
ks = int(sigma * 6) | 1; ks = max(ks, 3); ks = min(ks, 31)
k1d = torch.exp(-torch.arange(-(ks//2), ks//2+1).float()**2 / (2*sigma**2+1e-6))
k1d = k1d / k1d.sum()
k2d = k1d.unsqueeze(1) @ k1d.unsqueeze(0)
self.kernels.append(nn.Parameter(k2d.unsqueeze(0).unsqueeze(0), requires_grad=False))
self.refine = nn.Sequential(
nn.Conv2d(3, 16, 3, padding=1), nn.GELU(),
nn.Conv2d(16, 3, 3, padding=1))
def _blur_at_level(self, image, kernel):
B, C, H, W = image.shape
k = kernel.expand(C, -1, -1, -1)
p = kernel.shape[-1] // 2
return F.conv2d(F.pad(image, [p]*4, mode='reflect'), k, groups=C)
def forward(self, image, depth, f_number, focal_mm, focus_m):
B, C, H, W = image.shape
f = focal_mm.view(-1,1,1,1); N = f_number.view(-1,1,1,1)
S1 = (focus_m.view(-1,1,1,1) * 1000).clamp(min=51)
D = (depth * 1000).clamp(min=100)
coc = (f**2 / (N * (S1 - f).clamp(min=1))) * (D - S1).abs() / D
coc_px = (coc * W / self.sensor_width / 2).clamp(0, self.max_radius)
coc_norm = coc_px / self.max_radius
blurred_levels = [self._blur_at_level(image, kernel) for kernel in self.kernels]
level_float = coc_norm * (self.n_levels - 1)
level_low = level_float.long().clamp(0, self.n_levels - 2)
level_frac = (level_float - level_low.float()).clamp(0, 1)
rendered = image.clone()
for lv in range(self.n_levels - 1):
mask = (level_low == lv).float()
if mask.sum() > 0:
interp = blurred_levels[lv] * (1 - level_frac) + blurred_levels[lv + 1] * level_frac
rendered = rendered * (1 - mask) + interp * mask
mask_top = (level_low >= self.n_levels - 2).float() * (level_frac > 0.99).float()
rendered = rendered * (1 - mask_top) + blurred_levels[-1] * mask_top
rendered = rendered + self.refine(rendered) * 0.1
return rendered, coc_px
class BokehFlow(nn.Module):
def __init__(self, config=None):
super().__init__()
if config is None: config = BokehFlowConfig()
self.config = config; c = config
self.stem = ConvStem(3, c.stem_channels, c.embed_dim)
self.aperture_enc = ApertureEncoder(c.aperture_embed_dim)
self.depth_blocks = nn.ModuleList([
GatedConvRecurrence(c.embed_dim, c.large_kernel, c.ffn_expansion)
for _ in range(c.depth_blocks)])
self.bokeh_blocks = nn.ModuleList([
GatedConvRecurrenceWithACFM(c.embed_dim, c.large_kernel, c.ffn_expansion, c.aperture_embed_dim)
for _ in range(c.bokeh_blocks)])
n_fusions = max(c.depth_blocks, c.bokeh_blocks) // c.fusion_every
self.fusions = nn.ModuleList([CrossFusion(c.embed_dim) for _ in range(n_fusions)])
self.depth_head = DepthHead(c.embed_dim)
self.bokeh_head = BokehHead(c.embed_dim)
self.pgcoc = PGCoC(c.sensor_width_mm, c.max_coc_radius)
self.blend_w = nn.Parameter(torch.tensor(0.5))
def forward(self, image, f_number=None, focal_length_mm=None,
focus_distance_m=None, **kwargs):
B = image.shape[0]; dev = image.device; c = self.config
if f_number is None: f_number = torch.full((B,), c.default_fnumber, device=dev)
if focal_length_mm is None: focal_length_mm = torch.full((B,), c.default_focal_mm, device=dev)
if focus_distance_m is None: focus_distance_m = torch.full((B,), c.default_focus_m, device=dev)
ae = self.aperture_enc(f_number, focal_length_mm, focus_distance_m)
feat = self.stem(image)
d_feat = feat; b_feat = feat; fi = 0
n_blk = max(c.depth_blocks, c.bokeh_blocks)
for i in range(n_blk):
if i < c.depth_blocks: d_feat = self.depth_blocks[i](d_feat)
if i < c.bokeh_blocks: b_feat = self.bokeh_blocks[i](b_feat, ae)
if (i+1) % c.fusion_every == 0 and fi < len(self.fusions):
d_feat, b_feat = self.fusions[fi](d_feat, b_feat); fi += 1
depth = self.depth_head(d_feat)
if depth.shape[2:] != image.shape[2:]:
depth = F.interpolate(depth, image.shape[2:], mode='bilinear', align_corners=False)
physics_bokeh, coc_map = self.pgcoc(image, depth, f_number, focal_length_mm, focus_distance_m)
learned_bokeh = self.bokeh_head(b_feat)
if learned_bokeh.shape[2:] != image.shape[2:]:
learned_bokeh = F.interpolate(learned_bokeh, image.shape[2:], mode='bilinear', align_corners=False)
w = torch.sigmoid(self.blend_w)
bokeh = (w * physics_bokeh + (1-w) * (image + learned_bokeh)).clamp(0, 1)
return {'bokeh': bokeh, 'depth': depth, 'coc_map': coc_map}
class BokehFlowLoss(nn.Module):
"""Combined L1 + SSIM loss."""
def forward(self, pred, targets):
bp, bg = pred['bokeh'], targets['bokeh_gt']
l1 = F.l1_loss(bp, bg)
C1, C2 = 0.01**2, 0.03**2
mu_p = F.avg_pool2d(bp, 11, 1, 5); mu_g = F.avg_pool2d(bg, 11, 1, 5)
mu_pp = mu_p*mu_p; mu_gg = mu_g*mu_g; mu_pg = mu_p*mu_g
sig_pp = F.avg_pool2d(bp*bp, 11, 1, 5) - mu_pp
sig_gg = F.avg_pool2d(bg*bg, 11, 1, 5) - mu_gg
sig_pg = F.avg_pool2d(bp*bg, 11, 1, 5) - mu_pg
ssim_map = ((2*mu_pg+C1)*(2*sig_pg+C2)) / ((mu_pp+mu_gg+C1)*(sig_pp+sig_gg+C2))
ssim_loss = 1.0 - ssim_map.mean()
return {'total': l1 + ssim_loss, 'l1': l1.detach(), 'ssim': ssim_loss.detach()}
def count_params(model):
return sum(p.numel() for p in model.parameters() if p.requires_grad)
if __name__ == "__main__":
import time
for v in ['nano', 'small', 'base']:
c = BokehFlowConfig(variant=v)
dev = 'cuda' if torch.cuda.is_available() else 'cpu'
m = BokehFlow(c).to(dev)
print(f"BokehFlow-{v}: {count_params(m):,} params")
x = torch.randn(2, 3, 256, 256, device=dev)
m.eval()
with torch.no_grad(): out = m(x)
if torch.cuda.is_available(): torch.cuda.synchronize()
t0 = time.time()
with torch.no_grad():
for _ in range(10): out = m(x)
if torch.cuda.is_available(): torch.cuda.synchronize()
print(f" Inference: {(time.time()-t0)/10*1000:.1f}ms/batch (B=2, 256x256)")
m.train()
opt = torch.optim.AdamW(m.parameters(), lr=1e-3)
loss_fn = BokehFlowLoss()
gt = torch.rand_like(x[:,:3])
out = m(x); loss = loss_fn(out, {'bokeh_gt': gt})['total']
opt.zero_grad(); loss.backward(); opt.step()
if torch.cuda.is_available(): torch.cuda.synchronize()
t0 = time.time()
for _ in range(10):
out = m(x); loss = loss_fn(out, {'bokeh_gt': gt})['total']
opt.zero_grad(); loss.backward(); opt.step()
if torch.cuda.is_available(): torch.cuda.synchronize()
print(f" Training: {(time.time()-t0)/10*1000:.1f}ms/step (B=2, 256x256)")
if torch.cuda.is_available():
print(f" VRAM: {torch.cuda.max_memory_allocated()/1e9:.2f} GB")
torch.cuda.reset_peak_memory_stats()
print()
|