Spaces:
Running
on
Zero
Running
on
Zero
File size: 12,009 Bytes
a7dedf9 |
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 |
from timm import create_model
from torch import nn, Tensor
from typing import Optional
from functools import partial
from ..utils import _get_activation, _get_norm_layer, ConvUpsample, ConvDownsample
from ..utils import LightConvUpsample, LightConvDownsample, LighterConvUpsample, LighterConvDownsample
from ..utils import ConvRefine, LightConvRefine, LighterConvRefine
regular_models = [
"resnet18", "resnet34", "resnet50", "resnet101", "resnet152",
"convnext_nano", "convnext_tiny", "convnext_small", "convnext_base",
"mobilenetv4_conv_large",
]
heavy_models = [
"convnext_large", "convnext_xlarge", "convnext_xxlarge",
]
light_models = [
"mobilenetv1_100", "mobilenetv1_125",
"mobilenetv2_100", "mobilenetv2_140",
"mobilenetv3_large_100",
"mobilenetv4_conv_medium",
]
lighter_models = [
"mobilenetv2_050",
"mobilenetv3_small_050", "mobilenetv3_small_075", "mobilenetv3_small_100",
"mobilenetv4_conv_small_050", "mobilenetv4_conv_small"
]
supported_models = regular_models + heavy_models + light_models + lighter_models
refiner_in_channels = {
# ResNet
"resnet18": 512,
"resnet34": 512,
"resnet50": 2048,
"resnet101": 2048,
"resnet152": 2048,
# ConvNeXt
"convnext_nano": 640,
"convnext_tiny": 768,
"convnext_small": 768,
"convnext_base": 1024,
"convnext_large": 1536,
"convnext_xlarge": 2048,
"convnext_xxlarge": 3072,
# MobileNet V1
"mobilenetv1_100": 1024,
"mobilenetv1_125": 1280,
# MobileNet V2
"mobilenetv2_050": 160,
"mobilenetv2_100": 320,
"mobilenetv2_140": 448,
# MobileNet V3
"mobilenetv3_small_050": 288,
"mobilenetv3_small_075": 432,
"mobilenetv3_small_100": 576,
"mobilenetv3_large_100": 960,
# MobileNet V4
"mobilenetv4_conv_small_050": 480,
"mobilenetv4_conv_small": 960,
"mobilenetv4_conv_medium": 960,
"mobilenetv4_conv_large": 960,
}
refiner_out_channels = {
# ResNet
"resnet18": 512,
"resnet34": 512,
"resnet50": 2048,
"resnet101": 2048,
"resnet152": 2048,
# ConvNeXt
"convnext_nano": 640,
"convnext_tiny": 768,
"convnext_small": 768,
"convnext_base": 1024,
"convnext_large": 1536,
"convnext_xlarge": 2048,
"convnext_xxlarge": 3072,
# MobileNet V1
"mobilenetv1_100": 512,
"mobilenetv1_125": 640,
# MobileNet V2
"mobilenetv2_050": 160,
"mobilenetv2_100": 320,
"mobilenetv2_140": 448,
# MobileNet V3
"mobilenetv3_small_050": 288,
"mobilenetv3_small_075": 432,
"mobilenetv3_small_100": 576,
"mobilenetv3_large_100": 480,
# MobileNet V4
"mobilenetv4_conv_small_050": 480,
"mobilenetv4_conv_small": 960,
"mobilenetv4_conv_medium": 960,
"mobilenetv4_conv_large": 960,
}
groups = {
# ResNet
"resnet18": 1,
"resnet34": 1,
"resnet50": refiner_in_channels["resnet50"] // 512,
"resnet101": refiner_in_channels["resnet101"] // 512,
"resnet152": refiner_in_channels["resnet152"] // 512,
# ConvNeXt
"convnext_nano": 8,
"convnext_tiny": 8,
"convnext_small": 8,
"convnext_base": 8,
"convnext_large": refiner_in_channels["convnext_large"] // 512,
"convnext_xlarge": refiner_in_channels["convnext_xlarge"] // 512,
"convnext_xxlarge": refiner_in_channels["convnext_xxlarge"] // 512,
# MobileNet V1
"mobilenetv1_100": None,
"mobilenetv1_125": None,
# MobileNet V2
"mobilenetv2_050": None,
"mobilenetv2_100": None,
"mobilenetv2_140": None,
# MobileNet V3
"mobilenetv3_small_050": None,
"mobilenetv3_small_075": None,
"mobilenetv3_small_100": None,
"mobilenetv3_large_100": None,
# MobileNet V4
"mobilenetv4_conv_small_050": None,
"mobilenetv4_conv_small": None,
"mobilenetv4_conv_medium": None,
"mobilenetv4_conv_large": 1,
}
class TIMMModel(nn.Module):
def __init__(
self,
model_name: str,
block_size: Optional[int] = None,
norm: str = "none",
act: str = "none"
) -> None:
super().__init__()
assert model_name in supported_models, f"Backbone {model_name} not supported. Supported models are {supported_models}"
assert block_size is None or block_size in [8, 16, 32], f"Block size should be one of [8, 16, 32], but got {block_size}."
self.model_name = model_name
self.encoder = create_model(model_name, pretrained=True, features_only=True, out_indices=[-1])
self.encoder_channels = self.encoder.feature_info.channels()[-1]
self.encoder_reduction = self.encoder.feature_info.reduction()[-1]
self.block_size = block_size if block_size is not None else self.encoder_reduction
if model_name in lighter_models:
upsample_block = LighterConvUpsample
downsample_block = LighterConvDownsample
decoder_block = LighterConvRefine
elif model_name in light_models:
upsample_block = LightConvUpsample
downsample_block = LightConvDownsample
decoder_block = LightConvRefine
else:
upsample_block = partial(ConvUpsample, groups=groups[model_name])
downsample_block = partial(ConvDownsample, groups=groups[model_name])
decoder_block = partial(ConvRefine, groups=groups[model_name])
if norm == "bn":
norm_layer = nn.BatchNorm2d
elif norm == "ln":
norm_layer = nn.LayerNorm
else:
norm_layer = _get_norm_layer(self.encoder)
if act == "relu":
activation = nn.ReLU(inplace=True)
elif act == "gelu":
activation = nn.GELU()
else:
activation = _get_activation(self.encoder)
if self.block_size > self.encoder_reduction:
if self.block_size > self.encoder_reduction * 2:
assert self.block_size == self.encoder_reduction * 4, f"Block size {self.block_size} is not supported for model {self.model_name}. Supported block sizes are {self.encoder_reduction}, {self.encoder_reduction * 2}, and {self.encoder_reduction * 4}."
self.refiner = nn.Sequential(
downsample_block(
in_channels=self.encoder_channels,
out_channels=refiner_in_channels[self.model_name],
norm_layer=norm_layer,
activation=activation,
),
downsample_block(
in_channels=refiner_in_channels[self.model_name],
out_channels=refiner_out_channels[self.model_name],
norm_layer=norm_layer,
activation=activation,
)
)
else:
assert self.block_size == self.encoder_reduction * 2, f"Block size {self.block_size} is not supported for model {self.model_name}. Supported block sizes are {self.encoder_reduction}, {self.encoder_reduction * 2}, and {self.encoder_reduction * 4}."
self.refiner = downsample_block(
in_channels=self.encoder_channels,
out_channels=refiner_out_channels[self.model_name],
norm_layer=norm_layer,
activation=activation,
)
self.refiner_channels = refiner_out_channels[self.model_name]
elif self.block_size < self.encoder_reduction:
if self.block_size < self.encoder_reduction // 2:
assert self.block_size == self.encoder_reduction // 4, f"Block size {self.block_size} is not supported for model {self.model_name}. Supported block sizes are {self.encoder_reduction}, {self.encoder_reduction // 2}, and {self.encoder_reduction // 4}."
self.refiner = nn.Sequential(
upsample_block(
in_channels=self.encoder_channels,
out_channels=refiner_in_channels[self.model_name],
norm_layer=norm_layer,
activation=activation,
),
upsample_block(
in_channels=refiner_in_channels[self.model_name],
out_channels=refiner_out_channels[self.model_name],
norm_layer=norm_layer,
activation=activation,
)
)
else:
assert self.block_size == self.encoder_reduction // 2, f"Block size {self.block_size} is not supported for model {self.model_name}. Supported block sizes are {self.encoder_reduction}, {self.encoder_reduction // 2}, and {self.encoder_reduction // 4}."
self.refiner = upsample_block(
in_channels=self.encoder_channels,
out_channels=refiner_out_channels[self.model_name],
norm_layer=norm_layer,
activation=activation,
)
self.refiner_channels = refiner_out_channels[self.model_name]
else:
self.refiner = nn.Identity()
self.refiner_channels = self.encoder_channels
self.refiner_reduction = self.block_size
if self.refiner_channels <= 256:
self.decoder = nn.Identity()
self.decoder_channels = self.refiner_channels
elif self.refiner_channels <= 512:
self.decoder = decoder_block(
in_channels=self.refiner_channels,
out_channels=self.refiner_channels // 2,
norm_layer=norm_layer,
activation=activation,
)
self.decoder_channels = self.refiner_channels // 2
elif self.refiner_channels <= 1024:
self.decoder = nn.Sequential(
decoder_block(
in_channels=self.refiner_channels,
out_channels=self.refiner_channels // 2,
norm_layer=norm_layer,
activation=activation,
),
decoder_block(
in_channels=self.refiner_channels // 2,
out_channels=self.refiner_channels // 4,
norm_layer=norm_layer,
activation=activation,
),
)
self.decoder_channels = self.refiner_channels // 4
else:
self.decoder = nn.Sequential(
decoder_block(
in_channels=self.refiner_channels,
out_channels=self.refiner_channels // 2,
norm_layer=norm_layer,
activation=activation,
),
decoder_block(
in_channels=self.refiner_channels // 2,
out_channels=self.refiner_channels // 4,
norm_layer=norm_layer,
activation=activation,
),
decoder_block(
in_channels=self.refiner_channels // 4,
out_channels=self.refiner_channels // 8,
norm_layer=norm_layer,
activation=activation,
),
)
self.decoder_channels = self.refiner_channels // 8
self.decoder_reduction = self.refiner_reduction
def encode(self, x: Tensor) -> Tensor:
return self.encoder(x)[0]
def refine(self, x: Tensor) -> Tensor:
return self.refiner(x)
def decode(self, x: Tensor) -> Tensor:
return self.decoder(x)
def forward(self, x: Tensor) -> Tensor:
x = self.encode(x)
x = self.refine(x)
x = self.decode(x)
return x
def _timm_model(model_name: str, block_size: Optional[int] = None, norm: str = "none", act: str = "none") -> TIMMModel:
return TIMMModel(model_name, block_size=block_size, norm=norm, act=act)
|