Datasets:
File size: 14,764 Bytes
fe39cc9 | 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 | """
Model Factory for public architectures used in this release.
"""
from pathlib import Path
import timm
import torch
import torch.nn as nn
import torchvision.models as models
from torchvision.models import (
DenseNet121_Weights,
EfficientNet_B0_Weights,
ResNet18_Weights,
ResNet34_Weights,
ResNet50_Weights,
Swin_B_Weights,
)
from config.experiment_config import MODEL_DEFAULTS
class MultiTaskHead(nn.Module):
"""Legacy fallback head used when biomarker_config is not provided."""
def __init__(self, input_dim, num_binary_tasks=7, num_calcium_classes=4, num_regression_tasks=2, dropout=0.1):
super().__init__()
self.shared_layers = nn.Sequential(
nn.Linear(input_dim, 512),
nn.ReLU(inplace=True),
nn.Dropout(dropout),
nn.BatchNorm1d(512),
)
self.binary_head = nn.Linear(512, num_binary_tasks)
self.calcium_head = nn.Linear(512, num_calcium_classes)
self.regression_head = nn.Linear(512, num_regression_tasks)
def forward(self, x):
shared = self.shared_layers(x)
binary_out = self.binary_head(shared)
calcium_out = self.calcium_head(shared)
regression_out = self.regression_head(shared)
return torch.cat([binary_out, calcium_out, regression_out], dim=1)
class ModelFactory:
"""Factory class for supported public model architectures."""
SUPPORTED_ARCHITECTURES = tuple(MODEL_DEFAULTS.keys())
@staticmethod
def create_model(
architecture,
num_classes=13,
pretrained_weights=None,
fine_tuning_strategy="full",
dropout=0.1,
biomarker_config=None,
single_target_strategy=None,
target_feature_dim=None,
single_target_output_dim=None,
**kwargs,
):
"""Create model based on architecture specification."""
if single_target_output_dim is not None:
target_feature_dim = single_target_output_dim
if architecture == "ResNet-18":
return ModelFactory._create_resnet18(
num_classes,
pretrained_weights,
fine_tuning_strategy,
dropout,
biomarker_config,
single_target_strategy,
target_feature_dim,
)
if architecture == "ResNet-34":
return ModelFactory._create_resnet34(
num_classes,
pretrained_weights,
fine_tuning_strategy,
dropout,
biomarker_config,
single_target_strategy,
target_feature_dim,
)
if architecture == "DenseNet-121":
return ModelFactory._create_densenet121(
num_classes,
pretrained_weights,
fine_tuning_strategy,
dropout,
biomarker_config,
single_target_strategy,
target_feature_dim,
)
if architecture == "EfficientNet-B0":
return ModelFactory._create_efficientnet_b0(
num_classes,
pretrained_weights,
fine_tuning_strategy,
dropout,
biomarker_config,
single_target_strategy,
target_feature_dim,
)
if architecture == "ViT-Small (DINOv2)":
return ModelFactory._create_dinov2_vit(
architecture,
num_classes,
fine_tuning_strategy,
dropout,
biomarker_config,
single_target_strategy,
target_feature_dim,
)
if architecture == "Swin Transformer-Base":
return ModelFactory._create_swin_base(
num_classes,
pretrained_weights,
fine_tuning_strategy,
dropout,
biomarker_config,
single_target_strategy,
target_feature_dim,
)
if architecture == "ResNet-50 (RadImageNet)":
return ModelFactory._create_resnet50_radimgnet(
num_classes,
fine_tuning_strategy,
dropout,
biomarker_config,
single_target_strategy,
target_feature_dim,
)
raise ValueError(
f"Unsupported architecture: {architecture}. "
f"Supported: {list(ModelFactory.SUPPORTED_ARCHITECTURES)}"
)
@staticmethod
def _create_multitask_head(feature_dim, dropout, biomarker_config, head_type="flexible", single_target_strategy=None, target_feature_dim=None):
if biomarker_config is not None:
if head_type == "linear_probe":
from .flexible_multitask_head import LinearProbeMultiTaskHead
return LinearProbeMultiTaskHead(
feature_dim,
biomarker_config,
dropout=dropout,
single_target_strategy=single_target_strategy,
target_feature_dim=target_feature_dim,
)
from .flexible_multitask_head import FlexibleMultiTaskHead
return FlexibleMultiTaskHead(
feature_dim,
biomarker_config,
dropout=dropout,
single_target_strategy=single_target_strategy,
target_feature_dim=target_feature_dim,
)
return MultiTaskHead(feature_dim, dropout=dropout)
@staticmethod
def _freeze_for_linear_probe(model, head_module):
for param in model.parameters():
param.requires_grad = False
for param in head_module.parameters():
param.requires_grad = True
@staticmethod
def _create_resnet18(num_classes, pretrained_weights, fine_tuning_strategy, dropout, biomarker_config, single_target_strategy=None, target_feature_dim=None):
head_type = "linear_probe" if fine_tuning_strategy == "linear_probe" else "flexible"
if pretrained_weights == "ImageNet":
model = models.resnet18(weights=ResNet18_Weights.IMAGENET1K_V1)
else:
model = models.resnet18(weights=None)
model.conv1 = nn.Conv2d(1, 64, kernel_size=7, stride=2, padding=3, bias=False)
feature_dim = model.fc.in_features
model.fc = ModelFactory._create_multitask_head(
feature_dim, dropout, biomarker_config, head_type, single_target_strategy, target_feature_dim
)
if fine_tuning_strategy == "linear_probe":
ModelFactory._freeze_for_linear_probe(model, model.fc)
return model
@staticmethod
def _create_resnet34(num_classes, pretrained_weights, fine_tuning_strategy, dropout, biomarker_config, single_target_strategy=None, target_feature_dim=None):
head_type = "linear_probe" if fine_tuning_strategy == "linear_probe" else "flexible"
if pretrained_weights == "ImageNet":
model = models.resnet34(weights=ResNet34_Weights.IMAGENET1K_V1)
else:
model = models.resnet34(weights=None)
model.conv1 = nn.Conv2d(1, 64, kernel_size=7, stride=2, padding=3, bias=False)
feature_dim = model.fc.in_features
model.fc = ModelFactory._create_multitask_head(
feature_dim, dropout, biomarker_config, head_type, single_target_strategy, target_feature_dim
)
if fine_tuning_strategy == "linear_probe":
ModelFactory._freeze_for_linear_probe(model, model.fc)
return model
@staticmethod
def _create_densenet121(num_classes, pretrained_weights, fine_tuning_strategy, dropout, biomarker_config=None, single_target_strategy=None, target_feature_dim=None):
head_type = "linear_probe" if fine_tuning_strategy == "linear_probe" else "flexible"
if pretrained_weights == "ImageNet":
model = models.densenet121(weights=DenseNet121_Weights.IMAGENET1K_V1)
else:
model = models.densenet121(weights=None)
feature_dim = model.classifier.in_features
model.classifier = ModelFactory._create_multitask_head(
feature_dim,
dropout,
biomarker_config,
head_type=head_type,
single_target_strategy=single_target_strategy,
target_feature_dim=target_feature_dim,
)
if fine_tuning_strategy == "linear_probe":
ModelFactory._freeze_for_linear_probe(model, model.classifier)
return model
@staticmethod
def _create_efficientnet_b0(num_classes, pretrained_weights, fine_tuning_strategy, dropout, biomarker_config=None, single_target_strategy=None, target_feature_dim=None):
head_type = "linear_probe" if fine_tuning_strategy == "linear_probe" else "flexible"
if pretrained_weights == "ImageNet":
model = models.efficientnet_b0(weights=EfficientNet_B0_Weights.IMAGENET1K_V1)
else:
model = models.efficientnet_b0(weights=None)
feature_dim = model.classifier[1].in_features
model.classifier = ModelFactory._create_multitask_head(
feature_dim,
dropout,
biomarker_config,
head_type=head_type,
single_target_strategy=single_target_strategy,
target_feature_dim=target_feature_dim,
)
if fine_tuning_strategy == "linear_probe":
ModelFactory._freeze_for_linear_probe(model, model.classifier)
return model
@staticmethod
def _create_dinov2_vit(architecture, num_classes, fine_tuning_strategy, dropout, biomarker_config, single_target_strategy=None, target_feature_dim=None):
head_type = "linear_probe" if fine_tuning_strategy == "linear_probe" else "flexible"
model = timm.create_model("vit_small_patch14_dinov2", pretrained=True, num_classes=0, img_size=256)
feature_dim = model.num_features
model.head = ModelFactory._create_multitask_head(
feature_dim,
dropout,
biomarker_config,
head_type=head_type,
single_target_strategy=single_target_strategy,
target_feature_dim=target_feature_dim,
)
if fine_tuning_strategy == "linear_probe":
ModelFactory._freeze_for_linear_probe(model, model.head)
return model
@staticmethod
def _create_swin_base(num_classes, pretrained_weights, fine_tuning_strategy, dropout, biomarker_config=None, single_target_strategy=None, target_feature_dim=None):
head_type = "linear_probe" if fine_tuning_strategy == "linear_probe" else "flexible"
if pretrained_weights == "ImageNet-22K":
model = models.swin_b(weights=Swin_B_Weights.IMAGENET1K_V1)
else:
model = models.swin_b(weights=None)
model.features[0][0] = nn.Conv2d(1, 128, kernel_size=4, stride=4)
feature_dim = model.head.in_features
model.head = ModelFactory._create_multitask_head(
feature_dim,
dropout,
biomarker_config,
head_type=head_type,
single_target_strategy=single_target_strategy,
target_feature_dim=target_feature_dim,
)
if fine_tuning_strategy == "linear_probe":
ModelFactory._freeze_for_linear_probe(model, model.head)
return model
@staticmethod
def _create_resnet50_radimgnet(num_classes, fine_tuning_strategy, dropout, biomarker_config, single_target_strategy=None, target_feature_dim=None):
head_type = "linear_probe" if fine_tuning_strategy == "linear_probe" else "flexible"
ckpt_path = (
Path(__file__).resolve().parents[1]
/ "radimagenet_ckpt"
/ "resnet50"
/ "ResNet50_RadImageNet.pt"
)
try:
if ckpt_path.exists():
model = models.resnet50(weights=None)
checkpoint = torch.load(str(ckpt_path), map_location="cpu")
state_dict = checkpoint.get("model", checkpoint.get("state_dict", checkpoint))
model_state_dict = model.state_dict()
filtered_state_dict = {}
for key, value in state_dict.items():
if key.startswith("fc.") or key.startswith("classifier."):
continue
mapped_key = key
if key.startswith("backbone."):
mapped_key = key[9:]
if mapped_key.startswith("4."):
mapped_key = "layer1." + mapped_key[2:]
elif mapped_key.startswith("5."):
mapped_key = "layer2." + mapped_key[2:]
elif mapped_key.startswith("6."):
mapped_key = "layer3." + mapped_key[2:]
elif mapped_key.startswith("7."):
mapped_key = "layer4." + mapped_key[2:]
elif mapped_key.startswith("0."):
mapped_key = "conv1." + mapped_key[2:]
elif mapped_key.startswith("1."):
mapped_key = "bn1." + mapped_key[2:]
elif key.startswith("features."):
mapped_key = key[9:]
if mapped_key in model_state_dict:
filtered_state_dict[mapped_key] = value
model.load_state_dict(filtered_state_dict, strict=False)
else:
model = models.resnet50(weights=ResNet50_Weights.IMAGENET1K_V1)
except Exception:
model = models.resnet50(weights=ResNet50_Weights.IMAGENET1K_V1)
feature_dim = model.fc.in_features
model.fc = ModelFactory._create_multitask_head(
feature_dim,
dropout,
biomarker_config,
head_type=head_type,
single_target_strategy=single_target_strategy,
target_feature_dim=target_feature_dim,
)
if fine_tuning_strategy == "linear_probe":
ModelFactory._freeze_for_linear_probe(model, model.fc)
return model
def get_model_memory_requirement(architecture):
"""Get expected GPU memory requirement for supported public architectures."""
memory_map = {
"ResNet-18": "4-6GB",
"ResNet-34": "6-8GB",
"DenseNet-121": "8-10GB",
"EfficientNet-B0": "6-8GB",
"ViT-Small (DINOv2)": "8-10GB",
"Swin Transformer-Base": "12-16GB",
"ResNet-50 (RadImageNet)": "6-8GB",
}
return memory_map.get(architecture, "Unknown")
|