File size: 15,571 Bytes
2335bf1 | 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 | """
Shared components for all MPKNet model variants.
Contains building blocks used across V1, V2, V3, V4 and detection models:
- RGCLayer: Biologically accurate retinal ganglion cell preprocessing
- BinocularPreMPK: Legacy retinal preprocessing (deprecated, use RGCLayer)
- StereoDisparity: Stereo disparity simulation
- OcularDominanceConv: Convolution with ocular dominance channels
- BinocularMPKPathway: Pathway with binocular processing
- MonocularPathwayBlock: Pathway keeping eyes separate
- StridedMonocularBlock: Strided pathway for V4
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Tuple
class RGCLayer(nn.Module):
"""
Biologically accurate Retinal Ganglion Cell layer.
Based on Kim et al. 2021 "Retinal Ganglion Cells—Diversity of Cell Types
and Clinical Relevance" (Front. Neurol. 12:661938).
Models three main RGC types that feed the P/K/M pathways:
1. MIDGET RGCs (~70% of RGCs):
- Small receptive field (5-10 μm dendritic field)
- Center-surround via Difference of Gaussians (DoG)
- Red-Green color opponency (L-M or M-L)
- Feeds PARVOCELLULAR (P) pathway
- High spatial acuity, low temporal resolution
2. PARASOL RGCs (~10% of RGCs):
- Large receptive field (30-300 μm dendritic field)
- Center-surround DoG on luminance
- Achromatic (no color, L+M pooled)
- Feeds MAGNOCELLULAR (M) pathway
- Motion detection, high temporal resolution
3. SMALL BISTRATIFIED RGCs (~5-8% of RGCs):
- Medium receptive field
- S-cone ON center, (L+M) OFF surround
- Blue-Yellow opponency
- Feeds KONIOCELLULAR (K) pathway
- Color context, particularly blue
Key biological details implemented:
- DoG (Difference of Gaussians) for center-surround RF
- RF size ratios: Midget < Bistratified < Parasol
- Surround ~3-6x larger than center (we use 3x)
- ON-center and OFF-center populations (we use ON-center)
"""
def __init__(
self,
midget_sigma: float = 0.8, # Small RF for fine detail
parasol_sigma: float = 2.5, # Large RF for motion/gist
bistrat_sigma: float = 1.2, # Medium RF for color context
surround_ratio: float = 3.0, # Surround is 3x center
):
super().__init__()
self.midget_sigma = midget_sigma
self.parasol_sigma = parasol_sigma
self.bistrat_sigma = bistrat_sigma
self.surround_ratio = surround_ratio
# Create DoG kernels for each cell type
self.register_buffer('midget_center', self._make_gaussian(midget_sigma))
self.register_buffer('midget_surround', self._make_gaussian(midget_sigma * surround_ratio))
self.register_buffer('parasol_center', self._make_gaussian(parasol_sigma))
self.register_buffer('parasol_surround', self._make_gaussian(parasol_sigma * surround_ratio))
self.register_buffer('bistrat_center', self._make_gaussian(bistrat_sigma))
self.register_buffer('bistrat_surround', self._make_gaussian(bistrat_sigma * surround_ratio))
# Store kernel sizes for padding calculation
self.midget_ks = self.midget_surround.shape[-1]
self.parasol_ks = self.parasol_surround.shape[-1]
self.bistrat_ks = self.bistrat_surround.shape[-1]
def _make_gaussian(self, sigma: float) -> torch.Tensor:
"""Create a normalized 2D Gaussian kernel."""
ks = int(6 * sigma + 1) | 1 # Ensure odd, cover 3 sigma each side
ax = torch.arange(ks, dtype=torch.float32) - ks // 2
xx, yy = torch.meshgrid(ax, ax, indexing='ij')
kernel = torch.exp(-(xx**2 + yy**2) / (2 * sigma**2))
kernel = kernel / kernel.sum() # Normalize
return kernel.unsqueeze(0).unsqueeze(0) # (1, 1, H, W)
def _apply_dog(
self,
x: torch.Tensor,
center_kernel: torch.Tensor,
surround_kernel: torch.Tensor,
kernel_size: int
) -> torch.Tensor:
"""Apply Difference of Gaussians (center - surround)."""
B, C, H, W = x.shape
padding = kernel_size // 2
# Expand kernels for all channels
center_k = center_kernel.expand(C, 1, -1, -1)
surround_k = surround_kernel.expand(C, 1, -1, -1)
# Pad surround kernel to match size if needed
c_size = center_k.shape[-1]
s_size = surround_k.shape[-1]
if c_size < s_size:
pad_amt = (s_size - c_size) // 2
center_k = F.pad(center_k, (pad_amt, pad_amt, pad_amt, pad_amt))
# Apply center and surround
center_response = F.conv2d(x, center_k, padding=padding, groups=C)
surround_response = F.conv2d(x, surround_k, padding=padding, groups=C)
# DoG: ON-center response (center - surround)
return center_response - surround_response
def forward(
self,
x_left: torch.Tensor,
x_right: torch.Tensor
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor,
torch.Tensor, torch.Tensor, torch.Tensor]:
"""
Process left and right eye inputs through RGC populations.
Returns:
P_left, P_right: Midget RGC output (R-G opponency) -> P pathway
M_left, M_right: Parasol RGC output (luminance DoG) -> M pathway
K_left, K_right: Bistratified RGC output (S vs L+M) -> K pathway
"""
# ========== MIDGET RGCs -> P pathway ==========
# Red-Green opponency: L-cone vs M-cone
# Approximate: R channel vs G channel
# DoG on the opponent signal
# Extract R and G channels (approximating L and M cones)
R_left, G_left = x_left[:, 0:1], x_left[:, 1:2]
R_right, G_right = x_right[:, 0:1], x_right[:, 1:2]
# L-M opponency (R-G) with small receptive field DoG
rg_left = R_left - G_left
rg_right = R_right - G_right
P_left = self._apply_dog(rg_left, self.midget_center, self.midget_surround, self.midget_ks)
P_right = self._apply_dog(rg_right, self.midget_center, self.midget_surround, self.midget_ks)
# Expand back to 3 channels for compatibility
P_left = P_left.expand(-1, 3, -1, -1)
P_right = P_right.expand(-1, 3, -1, -1)
# ========== PARASOL RGCs -> M pathway ==========
# Achromatic: pool L+M (approximate as luminance)
# Large RF DoG for motion sensitivity
lum_left = 0.299 * x_left[:, 0:1] + 0.587 * x_left[:, 1:2] + 0.114 * x_left[:, 2:3]
lum_right = 0.299 * x_right[:, 0:1] + 0.587 * x_right[:, 1:2] + 0.114 * x_right[:, 2:3]
M_left = self._apply_dog(lum_left, self.parasol_center, self.parasol_surround, self.parasol_ks)
M_right = self._apply_dog(lum_right, self.parasol_center, self.parasol_surround, self.parasol_ks)
# Expand to 3 channels
M_left = M_left.expand(-1, 3, -1, -1)
M_right = M_right.expand(-1, 3, -1, -1)
# ========== SMALL BISTRATIFIED RGCs -> K pathway ==========
# S-cone ON center, (L+M) OFF surround
# Blue-Yellow opponency: S vs (L+M)
# S-cone approximated by B channel
# (L+M) approximated by (R+G)/2
S_left = x_left[:, 2:3] # Blue
S_right = x_right[:, 2:3]
LM_left = (x_left[:, 0:1] + x_left[:, 1:2]) / 2
LM_right = (x_right[:, 0:1] + x_right[:, 1:2]) / 2
# S - (L+M) opponency with medium RF
by_left = S_left - LM_left
by_right = S_right - LM_right
K_left = self._apply_dog(by_left, self.bistrat_center, self.bistrat_surround, self.bistrat_ks)
K_right = self._apply_dog(by_right, self.bistrat_center, self.bistrat_surround, self.bistrat_ks)
# Expand to 3 channels
K_left = K_left.expand(-1, 3, -1, -1)
K_right = K_right.expand(-1, 3, -1, -1)
return P_left, M_left, K_left, P_right, M_right, K_right
class BinocularPreMPK(nn.Module):
"""
Simulates retinal + LGN preprocessing for both eyes.
Each eye gets its own center-surround filtering.
Biological motivation:
- Retinal ganglion cells have center-surround receptive fields
- M cells respond to luminance changes (motion/gist)
- P cells respond to color/detail (high-pass filtered)
"""
def __init__(self, sigma: float = 1.0):
super().__init__()
self.sigma = sigma
ks = int(4 * sigma + 1) | 1 # ensure odd
ax = torch.arange(ks, dtype=torch.float32) - ks // 2
xx, yy = torch.meshgrid(ax, ax, indexing='ij')
kernel = torch.exp(-(xx**2 + yy**2) / (2 * sigma**2))
kernel = kernel / kernel.sum()
self.register_buffer('gauss', kernel.unsqueeze(0).unsqueeze(0))
self.ks = ks
def _blur(self, x: torch.Tensor) -> torch.Tensor:
B, C, H, W = x.shape
kernel = self.gauss.expand(C, 1, self.ks, self.ks)
return F.conv2d(x, kernel, padding=self.ks // 2, groups=C)
def forward(self, x_left: torch.Tensor, x_right: torch.Tensor) -> Tuple[torch.Tensor, ...]:
"""
Returns (P_left, M_left, P_right, M_right)
P = high-pass (center - surround) for detail
M = low-pass luminance for motion/gist
"""
# Left eye
blur_L = self._blur(x_left)
P_left = x_left - blur_L # high-pass (Parvo-like)
lum_L = x_left.mean(dim=1, keepdim=True)
M_left = self._blur(lum_L).expand(-1, 3, -1, -1) # low-pass luminance (Magno-like)
# Right eye
blur_R = self._blur(x_right)
P_right = x_right - blur_R
lum_R = x_right.mean(dim=1, keepdim=True)
M_right = self._blur(lum_R).expand(-1, 3, -1, -1)
return P_left, M_left, P_right, M_right
class StereoDisparity(nn.Module):
"""
Creates stereo disparity by horizontally shifting left/right views.
Simulates the slight positional difference between two eyes.
disparity_range: maximum pixel shift (positive = crossed disparity)
"""
def __init__(self, disparity_range: int = 2):
super().__init__()
self.disparity_range = disparity_range
def forward(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
"""
Takes single image, returns (left_view, right_view) with disparity.
For training, uses random disparity. For inference, uses fixed small disparity.
"""
B, C, H, W = x.shape
if self.training:
d = torch.randint(-self.disparity_range, self.disparity_range + 1, (1,)).item()
else:
d = 1
if d == 0:
return x, x
if d > 0:
x_left = F.pad(x[:, :, :, d:], (0, d, 0, 0), mode='replicate')
x_right = F.pad(x[:, :, :, :-d], (d, 0, 0, 0), mode='replicate')
else:
d = -d
x_left = F.pad(x[:, :, :, :-d], (d, 0, 0, 0), mode='replicate')
x_right = F.pad(x[:, :, :, d:], (0, d, 0, 0), mode='replicate')
return x_left, x_right
class OcularDominanceConv(nn.Module):
"""
Convolution with ocular dominance - channels are assigned to left/right eye
with graded mixing (some purely monocular, some binocular).
Inspired by V1 ocular dominance columns but applied at LGN stage
for computational efficiency.
"""
def __init__(self, in_ch: int, out_ch: int, kernel_size: int,
monocular_ratio: float = 0.5):
super().__init__()
self.out_ch = out_ch
self.monocular_ratio = monocular_ratio
n_mono = int(out_ch * monocular_ratio)
n_mono_per_eye = n_mono // 2
n_bino = out_ch - 2 * n_mono_per_eye
self.n_left = n_mono_per_eye
self.n_right = n_mono_per_eye
self.n_bino = n_bino
self.conv_left = nn.Conv2d(in_ch, n_mono_per_eye, kernel_size, padding=kernel_size//2)
self.conv_right = nn.Conv2d(in_ch, n_mono_per_eye, kernel_size, padding=kernel_size//2)
self.conv_bino_L = nn.Conv2d(in_ch, n_bino, kernel_size, padding=kernel_size//2)
self.conv_bino_R = nn.Conv2d(in_ch, n_bino, kernel_size, padding=kernel_size//2)
self.bn = nn.BatchNorm2d(out_ch)
def forward(self, x_left: torch.Tensor, x_right: torch.Tensor) -> torch.Tensor:
left_only = self.conv_left(x_left)
right_only = self.conv_right(x_right)
bino = self.conv_bino_L(x_left) + self.conv_bino_R(x_right)
out = torch.cat([left_only, right_only, bino], dim=1)
return F.relu(self.bn(out))
class BinocularMPKPathway(nn.Module):
"""
Single pathway (M, P, or K) with binocular processing.
Receives left and right eye inputs, produces fused output.
"""
def __init__(self, in_ch: int, out_ch: int, kernel_sizes: list,
monocular_ratio: float = 0.5):
super().__init__()
layers = []
ch = in_ch
for i, ks in enumerate(kernel_sizes):
is_first = (i == 0)
if is_first:
layers.append(OcularDominanceConv(ch, out_ch, ks, monocular_ratio))
else:
layers.append(nn.Sequential(
nn.Conv2d(out_ch if i > 0 else ch, out_ch, ks, padding=ks//2),
nn.BatchNorm2d(out_ch),
nn.ReLU(inplace=True)
))
ch = out_ch
self.first_layer = layers[0]
self.rest = nn.Sequential(*layers[1:]) if len(layers) > 1 else nn.Identity()
def forward(self, x_left: torch.Tensor, x_right: torch.Tensor) -> torch.Tensor:
x = self.first_layer(x_left, x_right)
return self.rest(x)
class MonocularPathwayBlock(nn.Module):
"""
Single pathway block that keeps left/right eyes separate.
Used for LGN processing where eye segregation persists.
"""
def __init__(self, in_ch: int, out_ch: int, kernel_size: int):
super().__init__()
self.conv_left = nn.Sequential(
nn.Conv2d(in_ch, out_ch, kernel_size, padding=kernel_size//2),
nn.BatchNorm2d(out_ch),
nn.ReLU(inplace=True)
)
self.conv_right = nn.Sequential(
nn.Conv2d(in_ch, out_ch, kernel_size, padding=kernel_size//2),
nn.BatchNorm2d(out_ch),
nn.ReLU(inplace=True)
)
def forward(self, x_left: torch.Tensor, x_right: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
return self.conv_left(x_left), self.conv_right(x_right)
class StridedMonocularBlock(nn.Module):
"""
Monocular pathway block with configurable stride.
Keeps left/right eyes separate, uses stride to control spatial sampling.
Used in V4 for stride-based pathway differentiation.
"""
def __init__(self, in_ch: int, out_ch: int, kernel_size: int = 3, stride: int = 1):
super().__init__()
padding = kernel_size // 2
self.conv_left = nn.Sequential(
nn.Conv2d(in_ch, out_ch, kernel_size, stride=stride, padding=padding),
nn.BatchNorm2d(out_ch),
nn.ReLU(inplace=True)
)
self.conv_right = nn.Sequential(
nn.Conv2d(in_ch, out_ch, kernel_size, stride=stride, padding=padding),
nn.BatchNorm2d(out_ch),
nn.ReLU(inplace=True)
)
def forward(self, x_left: torch.Tensor, x_right: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
return self.conv_left(x_left), self.conv_right(x_right)
def count_params(model: nn.Module) -> int:
"""Count total trainable parameters."""
return sum(p.numel() for p in model.parameters())
|