File size: 7,551 Bytes
18a82fb | 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 | # Model configurations separated from factory for better organization
MODEL_REGISTRY = {
# EfficientNet family
'efficientnet-b0': 'efficientnet_b0',
'efficientnet-b1': 'efficientnet_b1',
'efficientnet-b2': 'efficientnet_b2',
'efficientnet-b3': 'efficientnet_b3',
'efficientnet-b4': 'efficientnet_b4',
'efficientnet-b5': 'efficientnet_b5',
'efficientnet-b6': 'efficientnet_b6',
'efficientnet-b7': 'efficientnet_b7',
# ConvNeXt V2 family
'convnext-tiny': 'convnextv2_tiny.fcmae_ft_in22k_in1k',
'convnext-small': 'convnextv2_small.fcmae_ft_in22k_in1k',
'convnext-base': 'convnextv2_base.fcmae_ft_in22k_in1k',
'convnext-tiny-384': 'convnextv2_tiny.fcmae_ft_in22k_in1k_384',
# ResNet family
'resnet18': 'resnet18',
'resnet34': 'resnet34',
'resnet50': 'resnet50',
'resnet101': 'resnet101',
'resnet152': 'resnet152',
# ResNeXt family
'resnext50': 'resnext50_32x4d',
'resnext101': 'resnext101_32x8d',
# Vision Transformer family
'vit-tiny': 'vit_tiny_patch16_224',
'vit-small': 'vit_small_patch16_224',
'vit-base': 'vit_base_patch16_224',
# MobileNet family
'mobilenet-v3-small': 'mobilenetv3_small_100',
'mobilenet-v3-large': 'mobilenetv3_large_100',
}
# Training configurations for each model
TRAINING_CONFIGS = {
# EfficientNet configs
'efficientnet-b0': {
'input_size': 224,
'batch_size': 128,
'learning_rate': 1e-3,
'weight_decay': 1e-5,
'epochs': 30,
'warmup_epochs': 3,
'min_lr': 1e-6,
},
'efficientnet-b1': {
'input_size': 240,
'batch_size': 96,
'learning_rate': 1e-3,
'weight_decay': 1e-5,
'epochs': 30,
'warmup_epochs': 3,
'min_lr': 1e-6,
},
'efficientnet-b2': {
'input_size': 260,
'batch_size': 64,
'learning_rate': 1e-3,
'weight_decay': 1e-5,
'epochs': 30,
'warmup_epochs': 3,
'min_lr': 1e-6,
},
'efficientnet-b3': {
'input_size': 300,
'batch_size': 48,
'learning_rate': 1e-3,
'weight_decay': 1e-5,
'epochs': 30,
'warmup_epochs': 3,
'min_lr': 1e-6,
},
'efficientnet-b4': {
'input_size': 380,
'batch_size': 32,
'learning_rate': 1e-3,
'weight_decay': 1e-5,
'epochs': 30,
'warmup_epochs': 3,
'min_lr': 1e-6,
},
# ConvNeXt configs - these models benefit from different hyperparameters
'convnext-tiny': {
'input_size': 224,
'batch_size': 64,
'learning_rate': 5e-4,
'weight_decay': 0.05, # Higher weight decay for ConvNeXt
'epochs': 30,
'warmup_epochs': 5,
'min_lr': 1e-6,
'drop_path_rate': 0.1,
},
'convnext-base': {
'input_size': 224,
'batch_size': 32,
'learning_rate': 5e-4,
'weight_decay': 0.05,
'epochs': 30,
'warmup_epochs': 5,
'min_lr': 1e-6,
'drop_path_rate': 0.2,
},
'convnext-tiny-384': {
'input_size': 384,
'batch_size': 16, # Smaller batch due to larger image size
'learning_rate': 5e-4,
'weight_decay': 0.05,
'epochs': 30,
'warmup_epochs': 5,
'min_lr': 1e-6,
'drop_path_rate': 0.1,
},
# ResNet configs
'resnet50': {
'input_size': 224,
'batch_size': 64,
'learning_rate': 1e-3,
'weight_decay': 1e-4,
'epochs': 30,
'warmup_epochs': 3,
'min_lr': 1e-6,
},
'resnext50': {
'input_size': 224,
'batch_size': 48,
'learning_rate': 1e-3,
'weight_decay': 1e-4,
'epochs': 30,
'warmup_epochs': 3,
'min_lr': 1e-6,
},
# Vision Transformer configs
'vit-tiny': {
'input_size': 224,
'batch_size': 128,
'learning_rate': 1e-4, # Lower LR for ViT
'weight_decay': 0.05,
'epochs': 30,
'warmup_epochs': 5,
'min_lr': 1e-6,
'drop_rate': 0.1,
},
}
# Augmentation configurations
AUGMENTATION_CONFIGS = {
'light': {
'rotation': 5,
'brightness': 0.1,
'contrast': 0.1,
'saturation': 0.1,
'hue': 0.02,
'perspective': 0.1,
'scale': (0.9, 1.0),
'h_flip_p': 0.5,
},
'medium': {
'rotation': 10,
'brightness': 0.2,
'contrast': 0.2,
'saturation': 0.2,
'hue': 0.05,
'perspective': 0.2,
'scale': (0.8, 1.0),
'h_flip_p': 0.5,
},
'heavy': {
'rotation': 15,
'brightness': 0.3,
'contrast': 0.3,
'saturation': 0.3,
'hue': 0.1,
'perspective': 0.3,
'scale': (0.7, 1.0),
'h_flip_p': 0.5,
'mixup_alpha': 0.2,
'cutmix_alpha': 1.0,
}
}
# Model-specific fine-tuning strategies
FINETUNE_CONFIGS = {
'efficientnet': {
'freeze_stages': ['stem', 'blocks.0', 'blocks.1'], # Freeze early layers
'unfreeze_epoch': 10, # Unfreeze all layers after this epoch
'discriminative_lr': True, # Use different LRs for different layers
'lr_mult': [0.1, 0.3, 0.5, 0.7, 1.0], # LR multipliers from bottom to top
},
'convnext': {
'freeze_stages': ['stem', 'stages.0'],
'unfreeze_epoch': 10,
'discriminative_lr': True,
'lr_mult': [0.1, 0.5, 1.0],
},
'resnet': {
'freeze_stages': ['conv1', 'bn1', 'layer1'],
'unfreeze_epoch': 10,
'discriminative_lr': False,
},
'vit': {
'freeze_stages': ['patch_embed', 'blocks.0', 'blocks.1'],
'unfreeze_epoch': 15,
'discriminative_lr': True,
'lr_mult': [0.1, 0.5, 1.0],
}
}
def get_model_family(model_name: str) -> str:
"""Get the family of a model (e.g., 'efficientnet', 'convnext')"""
if 'efficientnet' in model_name:
return 'efficientnet'
elif 'convnext' in model_name:
return 'convnext'
elif 'resnext' in model_name:
return 'resnet' # ResNeXt uses similar strategy as ResNet
elif 'resnet' in model_name:
return 'resnet'
elif 'vit' in model_name:
return 'vit'
elif 'mobilenet' in model_name:
return 'efficientnet' # Similar strategy
else:
return 'default'
def get_optimal_batch_size(model_name: str, gpu_memory_gb: int = 8) -> int:
"""
Get optimal batch size based on model and GPU memory
Args:
model_name: Name of the model
gpu_memory_gb: GPU memory in GB
Returns:
Recommended batch size
"""
base_config = TRAINING_CONFIGS.get(model_name, TRAINING_CONFIGS['resnet50'])
base_batch_size = base_config['batch_size']
# Adjust based on GPU memory (assuming base is for 8GB)
memory_multiplier = gpu_memory_gb / 8.0
# Some models scale better than others
if 'efficientnet' in model_name:
memory_multiplier *= 1.2 # EfficientNets are memory efficient
elif 'vit' in model_name:
memory_multiplier *= 0.8 # ViTs use more memory
adjusted_batch_size = int(base_batch_size * memory_multiplier)
# Round to nearest power of 2 or multiple of 8
return max(8, min(256, (adjusted_batch_size // 8) * 8))
# Default config for unknown models
DEFAULT_CONFIG = {
'input_size': 224,
'batch_size': 32,
'learning_rate': 1e-3,
'weight_decay': 1e-4,
'epochs': 30,
'warmup_epochs': 3,
'min_lr': 1e-6,
} |