File size: 13,738 Bytes
90038de |
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 |
"""
CondensatePose Model Architecture
=================================
EfficientNetV2 encoder with Feature Pyramid Network and Style Modulation
for detecting biomolecular condensates in fluorescence microscopy images.
Architecture Components:
- Encoder: EfficientNetV2 (pretrained, adapted for grayscale)
- Decoder: Multi-scale FPN with style-based feature modulation
- Outputs: Binary mask + flow fields for instance segmentation
Paper: [Add your paper link here]
GitHub: [Add your GitHub link here]
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
from timm import create_model
from typing import Dict
class SparseAttention(nn.Module):
"""Spatial attention module for focusing on sparse condensate regions."""
def __init__(self, kernel_size=11):
super().__init__()
self.attention = nn.Sequential(
nn.Conv2d(2, 1, kernel_size=kernel_size, padding=kernel_size//2, bias=False),
nn.Sigmoid()
)
def forward(self, x):
avg_out = torch.mean(x, dim=1, keepdim=True)
max_out, _ = torch.max(x, dim=1, keepdim=True)
attention_input = torch.cat([avg_out, max_out], dim=1)
attention_map = self.attention(attention_input)
return x * attention_map
class NormProjection(nn.Module):
"""Normalized 1x1 convolution for feature projection."""
def __init__(self, in_channels, out_channels):
super().__init__()
self.bn = nn.BatchNorm2d(in_channels)
self.conv = nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False)
def forward(self, x):
return self.conv(self.bn(x))
class ConvBlock(nn.Module):
"""Standard convolution block with normalization and activation."""
def __init__(self, in_channels, out_channels):
super().__init__()
self.bn = nn.BatchNorm2d(in_channels)
self.swish = nn.SiLU(inplace=True)
self.conv = nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1, bias=False)
def forward(self, x):
return self.conv(self.swish(self.bn(x)))
class StyleModulatedConv(nn.Module):
"""Convolution with style-based feature modulation."""
def __init__(self, channels, style_dim):
super().__init__()
self.conv_block = ConvBlock(channels, channels)
self.style_projection = nn.Linear(style_dim, channels)
def forward(self, x, style_vector):
feat = self.conv_block(x)
style_bias = self.style_projection(style_vector).unsqueeze(-1).unsqueeze(-1)
return feat + style_bias
class DualResidualBlock(nn.Module):
"""Two-stage residual block for deep feature fusion with style conditioning."""
def __init__(self, channels, style_dim):
super().__init__()
self.style_conv1 = StyleModulatedConv(channels, style_dim)
self.style_conv2 = StyleModulatedConv(channels, style_dim)
self.style_conv3 = StyleModulatedConv(channels, style_dim)
self.projection = NormProjection(channels, channels)
self.initial_conv = ConvBlock(channels, channels)
def forward(self, x, lateral, style_vector):
combined = self.initial_conv(x) + lateral
x_intermediate = self.style_conv1(combined, style_vector) + self.projection(x)
refined = self.style_conv2(x_intermediate, style_vector)
output = self.style_conv3(refined, style_vector) + x_intermediate
return output
class MultiScaleEncoder(nn.Module):
"""EfficientNetV2-based encoder for multi-scale feature extraction."""
def __init__(self, variant='rw_s', pyramid_channels=[24, 48, 64, 160]):
super().__init__()
encoder_name = f"efficientnetv2_{variant}"
self.base_encoder = create_model(
encoder_name,
features_only=True,
pretrained=True,
in_chans=1,
out_indices=[0, 1, 2, 3]
)
enc_channels = self.base_encoder.feature_info.channels()
self.channel_adapters = nn.ModuleList([
nn.Sequential(
nn.Conv2d(enc_ch, target_ch, kernel_size=1, bias=False),
nn.BatchNorm2d(target_ch)
)
for enc_ch, target_ch in zip(enc_channels[:4], pyramid_channels)
])
def forward(self, x):
features = self.base_encoder(x)[:4]
adapted_features = []
for feat, adapter in zip(features, self.channel_adapters):
adapted_features.append(adapter(feat))
return adapted_features
class CondensatePoseModel(nn.Module):
"""
CondensatePose: Multi-scale segmentation model for biomolecular condensates.
Architecture:
- Encoder: EfficientNetV2 with channel adaptation
- Decoder: Feature Pyramid Network with:
* Style-based feature modulation
* Dual residual blocks for feature fusion
* Multi-scale upsampling refinement
- Attention: Spatial attention for sparse object detection
- Outputs: Binary mask logits + flow field vectors
Args:
encoder_variant (str): EfficientNetV2 variant (default: 'rw_s')
pyramid_channels (list): Channel dimensions for pyramid levels
use_spatial_attention (bool): Enable spatial attention module
spatial_kernel_size (int): Kernel size for spatial attention
dropout_rate (float): Dropout rate in decoder
"""
def __init__(
self,
encoder_variant='rw_s',
pyramid_channels=[24, 48, 64, 160],
use_spatial_attention=True,
spatial_kernel_size=11,
dropout_rate=0.15
):
super().__init__()
self.use_spatial_attention = use_spatial_attention
self.pyramid_channels = pyramid_channels
# Multi-scale encoder
self.encoder = MultiScaleEncoder(
variant=encoder_variant,
pyramid_channels=pyramid_channels
)
# Style vector dimension
style_dim = pyramid_channels[-1] # Default: 160
pyramid_dim = 32
# Pyramid processing blocks
self.pyramid_block4 = DualResidualBlock(pyramid_dim, style_dim)
self.pyramid_block3 = DualResidualBlock(pyramid_dim, style_dim)
self.pyramid_block2 = DualResidualBlock(pyramid_dim, style_dim)
self.pyramid_block1 = DualResidualBlock(pyramid_dim, style_dim)
# Channel reduction for all levels
self.lateral_conv4 = nn.Conv2d(pyramid_channels[3], pyramid_dim, kernel_size=1, bias=False)
self.lateral_conv3 = nn.Conv2d(pyramid_channels[2], pyramid_dim, kernel_size=1, bias=False)
self.lateral_conv2 = nn.Conv2d(pyramid_channels[1], pyramid_dim, kernel_size=1, bias=False)
self.lateral_conv1 = nn.Conv2d(pyramid_channels[0], pyramid_dim, kernel_size=1, bias=False)
# Upsampling refinement blocks
self.upsample_blocks = nn.ModuleList([
DualResidualBlock(pyramid_dim, style_dim) for _ in range(3)
])
# Final feature projection
self.output_projection = NormProjection(pyramid_dim, pyramid_dim)
# Spatial attention
if use_spatial_attention:
self.spatial_attention = SparseAttention(spatial_kernel_size)
# Dropout
self.dropout = nn.Dropout2d(dropout_rate)
# Output heads
self.mask_head = nn.Conv2d(pyramid_dim, 1, kernel_size=1) # Binary segmentation
self.flow_head = nn.Conv2d(pyramid_dim, 2, kernel_size=1) # Flow vectors
self._init_weights()
def _init_weights(self):
"""Initialize weights for sparse condensate segmentation."""
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
if m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.BatchNorm2d):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.Linear):
nn.init.xavier_normal_(m.weight)
if m.bias is not None:
nn.init.constant_(m.bias, 0)
# Mask head initialization
nn.init.xavier_uniform_(self.mask_head.weight, gain=0.1)
if self.mask_head.bias is not None:
nn.init.constant_(self.mask_head.bias, -2.0) # Bias toward background
# Flow head: initialize to zero
nn.init.zeros_(self.flow_head.weight)
if self.flow_head.bias is not None:
nn.init.zeros_(self.flow_head.bias)
def upsample_and_refine(self, x, style_vector, block):
"""Upsample features and apply refinement block."""
x_up = F.interpolate(x, scale_factor=2, mode='nearest')
return block(x_up, torch.zeros_like(x_up), style_vector)
def forward(self, x: torch.Tensor) -> Dict[str, torch.Tensor]:
"""
Forward pass.
Args:
x: Input tensor of shape (B, 1, H, W) - grayscale microscopy image
Returns:
Dictionary with keys:
- 'mask': Shape (B, 1, H, W) - binary mask logits
- 'flows': Shape (B, 2, H, W) - flow field vectors (dy, dx)
"""
# Input validation
if torch.isnan(x).any() or torch.isinf(x).any():
raise ValueError("Input contains NaN or Inf values")
# Normalize input
x_mean = x.mean(dim=[2, 3], keepdim=True)
x_std = x.std(dim=[2, 3], keepdim=True) + 1e-8
x_normalized = (x - x_mean) / x_std
# Extract multi-scale features
features = self.encoder(x_normalized)
C1, C2, C3, C4 = features
# Compute global style vector
style_vector = F.adaptive_avg_pool2d(C4, 1).squeeze(-1).squeeze(-1)
style_vector = F.normalize(style_vector, p=2, dim=1)
# Build feature pyramid
C4_reduced = self.lateral_conv4(C4)
P4 = self.pyramid_block4(C4_reduced, C4_reduced, style_vector)
P4_up = F.interpolate(P4, size=C3.shape[2:], mode='nearest')
C3_reduced = self.lateral_conv3(C3)
P3 = self.pyramid_block3(P4_up, C3_reduced, style_vector)
P3_up = F.interpolate(P3, size=C2.shape[2:], mode='nearest')
C2_reduced = self.lateral_conv2(C2)
P2 = self.pyramid_block2(P3_up, C2_reduced, style_vector)
P2_up = F.interpolate(P2, size=C1.shape[2:], mode='nearest')
C1_reduced = self.lateral_conv1(C1)
P1 = self.pyramid_block1(P2_up, C1_reduced, style_vector)
# Multi-scale upsampling refinement
P4_refined = self.upsample_and_refine(P4, style_vector, self.upsample_blocks[0])
P4_refined = self.upsample_and_refine(P4_refined, style_vector, self.upsample_blocks[1])
P4_refined = self.upsample_and_refine(P4_refined, style_vector, self.upsample_blocks[2])
P3_refined = self.upsample_and_refine(P3, style_vector, self.upsample_blocks[0])
P3_refined = self.upsample_and_refine(P3_refined, style_vector, self.upsample_blocks[1])
P2_refined = self.upsample_and_refine(P2, style_vector, self.upsample_blocks[0])
# Combine all scales
combined = P1 + P2_refined + P3_refined + P4_refined
features_final = self.output_projection(combined)
# Apply spatial attention
if self.use_spatial_attention:
features_final = self.spatial_attention(features_final)
# Apply dropout
features_final = self.dropout(features_final)
# Generate outputs
mask_logits = self.mask_head(features_final)
flow_vectors = self.flow_head(features_final)
# Upsample to match input size
target_size = x.shape[2:]
mask_logits = F.interpolate(mask_logits, size=target_size, mode='bilinear', align_corners=False)
flow_vectors = F.interpolate(flow_vectors, size=target_size, mode='bilinear', align_corners=False)
return {
'mask': mask_logits,
'flows': flow_vectors
}
def load_condensatepose_model(
checkpoint_path: str,
device: str = 'cuda'
) -> CondensatePoseModel:
"""
Load a trained CondensatePose model from checkpoint.
Args:
checkpoint_path: Path to model checkpoint (.pth file)
device: Device to load model on ('cuda' or 'cpu')
Returns:
Loaded model in eval mode
Example:
>>> model = load_condensatepose_model('model_weights.pth', device='cuda')
>>> model.eval()
>>>
>>> # Run inference
>>> with torch.no_grad():
>>> outputs = model(image_tensor)
>>> mask_logits = outputs['mask']
>>> flows = outputs['flows']
"""
checkpoint = torch.load(checkpoint_path, map_location=device)
config = checkpoint.get('model_config', {})
model = CondensatePoseModel(
encoder_variant=config.get('encoder_variant', 'rw_s'),
pyramid_channels=config.get('pyramid_channels', [24, 48, 64, 160]),
use_spatial_attention=config.get('use_spatial_attention', True),
spatial_kernel_size=config.get('spatial_kernel_size', 11),
dropout_rate=config.get('dropout_rate', 0.15),
)
model.load_state_dict(checkpoint['model_state_dict'])
model.to(device)
model.eval()
return model
# Alias for compatibility
CondensateSegmentationNet = CondensatePoseModel
|