File size: 8,289 Bytes
1f2e262 | 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 | """
Model Factory - Stage 6
Factory functions for creating timm models with custom classifier heads.
"""
from typing import Optional
import torch
import torch.nn as nn
import timm
# Supported model architectures
SUPPORTED_MODELS = {
'efficientnet_b0': 'efficientnet_b0',
'resnet50': 'resnet50',
'convnext_tiny': 'convnext_tiny',
}
def create_model(config: dict) -> nn.Module:
"""
Create a timm model with custom classifier head for brain tumor classification.
Loads pretrained ImageNet weights and modifies the classifier head to output
4 classes (glioma, meningioma, notumor, pituitary). Applies dropout and
prepares the model for training.
Supported models:
- efficientnet_b0: ~5M params, best accuracy per VRAM
- resnet50: ~25M params, standard baseline
- convnext_tiny: ~28M params, modern architecture
Args:
config: Configuration dictionary with structure:
{
'model': {
'name': str, # Model architecture name
'pretrained': bool, # Whether to use pretrained weights
'drop_rate': float # Dropout rate for classifier head
},
'data': {
'num_classes': int # Number of output classes
}
}
Returns:
Model ready for training with modified classifier head
Raises:
ValueError: If model name is not supported
KeyError: If required config keys are missing
Example:
>>> config = {
... 'model': {'name': 'efficientnet_b0', 'pretrained': True, 'drop_rate': 0.2},
... 'data': {'num_classes': 4}
... }
>>> model = create_model(config)
>>> model = model.to('cuda')
>>> # Optionally convert to channels_last
>>> model = model.to(memory_format=torch.channels_last)
"""
# Extract config parameters
try:
model_name = config['model']['name']
pretrained = config['model']['pretrained']
drop_rate = config['model'].get('drop_rate', 0.0) # Optional, default to 0.0
num_classes = config['data']['num_classes']
except KeyError as e:
raise KeyError(f"Missing required config key: {e}")
# Validate model name
if model_name not in SUPPORTED_MODELS:
raise ValueError(
f"Unsupported model '{model_name}'. "
f"Supported models: {list(SUPPORTED_MODELS.keys())}"
)
# Map to timm model name
timm_model_name = SUPPORTED_MODELS[model_name]
# Create model with timm
# For some architectures (like ConvNeXt), we need to create with num_classes directly
# For others, we can use num_classes=0 and add a custom head
# Special handling for ConvNeXt which doesn't work well with num_classes=0
if 'convnext' in model_name.lower():
# Create with final num_classes directly
model = timm.create_model(
timm_model_name,
pretrained=pretrained,
num_classes=num_classes,
drop_rate=drop_rate,
drop_path_rate=config['model'].get('drop_path_rate', 0.0)
)
else:
# For other models, create without classifier and add custom head
model = timm.create_model(
timm_model_name,
pretrained=pretrained,
num_classes=0, # Remove the classifier head
drop_rate=drop_rate
)
# Get the number of features from the backbone
num_features = model.num_features
# Create custom classifier head
classifier = nn.Sequential(
nn.Dropout(p=drop_rate),
nn.Linear(num_features, num_classes)
)
# Attach classifier to model
# Different architectures use different attribute names for the classifier
if hasattr(model, 'fc'):
# ResNet uses 'fc'
model.fc = classifier
elif hasattr(model, 'classifier'):
# EfficientNet and others use 'classifier'
model.classifier = classifier
elif hasattr(model, 'head'):
# ConvNeXt and ViT use 'head'
model.head = classifier
else:
# Fallback: try to add as 'head'
model.head = classifier
return model
def get_model_info(model: nn.Module) -> dict:
"""
Get information about a model.
Args:
model: PyTorch model
Returns:
Dictionary with model information:
- total_params: Total number of parameters
- trainable_params: Number of trainable parameters
- model_size_mb: Approximate model size in MB
"""
total_params = sum(p.numel() for p in model.parameters())
trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
# Approximate size in MB (assuming float32)
model_size_mb = total_params * 4 / (1024 ** 2)
return {
'total_params': total_params,
'trainable_params': trainable_params,
'model_size_mb': round(model_size_mb, 2)
}
def freeze_backbone(model: nn.Module) -> nn.Module:
"""
Freeze all parameters except the classifier head.
Used for Stage 1 training (head-only).
Args:
model: Model to freeze
Returns:
Model with frozen backbone
"""
# Freeze all parameters
for param in model.parameters():
param.requires_grad = False
# Unfreeze classifier head
# Check different possible attribute names
if hasattr(model, 'fc') and model.fc is not None:
for param in model.fc.parameters():
param.requires_grad = True
elif hasattr(model, 'classifier') and model.classifier is not None:
for param in model.classifier.parameters():
param.requires_grad = True
elif hasattr(model, 'head') and model.head is not None:
for param in model.head.parameters():
param.requires_grad = True
return model
def unfreeze_all(model: nn.Module) -> nn.Module:
"""
Unfreeze all model parameters.
Used for Stage 2 training (full fine-tuning).
Args:
model: Model to unfreeze
Returns:
Model with all parameters trainable
"""
for param in model.parameters():
param.requires_grad = True
return model
def convert_to_channels_last(model: nn.Module) -> nn.Module:
"""
Convert model to channels_last memory format for better performance.
Should be called after moving model to GPU.
Args:
model: Model to convert
Returns:
Model in channels_last format
Example:
>>> model = create_model(config)
>>> model = model.to('cuda')
>>> model = convert_to_channels_last(model)
"""
return model.to(memory_format=torch.channels_last)
def prepare_model_for_training(
config: dict,
device: str = 'cuda',
stage: int = 1
) -> nn.Module:
"""
Create and prepare a model for training.
Convenience function that creates the model, moves it to device,
optionally converts to channels_last, and freezes the backbone for stage 1.
Args:
config: Model configuration
device: Device to move model to ('cuda' or 'cpu')
stage: Training stage (1 = head only, 2 = full fine-tune)
Returns:
Model ready for training
Example:
>>> # Stage 1: head-only training
>>> model = prepare_model_for_training(config, device='cuda', stage=1)
>>>
>>> # Stage 2: full fine-tuning
>>> model = prepare_model_for_training(config, device='cuda', stage=2)
"""
# Create model
model = create_model(config)
# Move to device
model = model.to(device)
# Convert to channels_last if configured
if config.get('training', {}).get('channels_last', False):
model = convert_to_channels_last(model)
# Freeze backbone for stage 1
if stage == 1:
model = freeze_backbone(model)
else:
model = unfreeze_all(model)
return model
|