|
| 1 | +from collections import OrderedDict |
| 2 | +from typing import Callable, Union |
| 3 | + |
| 4 | +from torch import nn |
| 5 | + |
| 6 | +from .layers import ConvBnAct, get_act |
| 7 | +from .model_constructor import ModelCfg, ModelConstructor |
| 8 | + |
| 9 | +__all__ = [ |
| 10 | + "XResBlock", |
| 11 | + "ModelConstructor", |
| 12 | + "XResNet34", |
| 13 | + "XResNet50", |
| 14 | +] |
| 15 | + |
| 16 | + |
| 17 | +# TModelCfg = TypeVar("TModelCfg", bound="ModelCfg") |
| 18 | + |
| 19 | + |
| 20 | +class XResBlock(nn.Module): |
| 21 | + """Universal XResnet block. Basic block if expansion is 1, otherwise is Bottleneck.""" |
| 22 | + |
| 23 | + def __init__( |
| 24 | + self, |
| 25 | + expansion: int, |
| 26 | + in_channels: int, |
| 27 | + mid_channels: int, |
| 28 | + stride: int = 1, |
| 29 | + conv_layer=ConvBnAct, |
| 30 | + act_fn: type[nn.Module] = nn.ReLU, |
| 31 | + zero_bn: bool = True, |
| 32 | + bn_1st: bool = True, |
| 33 | + groups: int = 1, |
| 34 | + dw: bool = False, |
| 35 | + div_groups: Union[None, int] = None, |
| 36 | + pool: Union[Callable[[], nn.Module], None] = None, |
| 37 | + se: Union[nn.Module, None] = None, |
| 38 | + sa: Union[nn.Module, None] = None, |
| 39 | + ): |
| 40 | + super().__init__() |
| 41 | + # pool defined at ModelConstructor. |
| 42 | + out_channels, in_channels = mid_channels * expansion, in_channels * expansion |
| 43 | + if div_groups is not None: # check if groups != 1 and div_groups |
| 44 | + groups = int(mid_channels / div_groups) |
| 45 | + if expansion == 1: |
| 46 | + layers = [ |
| 47 | + ( |
| 48 | + "conv_0", |
| 49 | + conv_layer( |
| 50 | + in_channels, |
| 51 | + mid_channels, |
| 52 | + 3, |
| 53 | + stride=stride, # type: ignore |
| 54 | + act_fn=act_fn, |
| 55 | + bn_1st=bn_1st, |
| 56 | + groups=in_channels if dw else groups, |
| 57 | + ), |
| 58 | + ), |
| 59 | + ( |
| 60 | + "conv_1", |
| 61 | + conv_layer( |
| 62 | + mid_channels, |
| 63 | + out_channels, |
| 64 | + 3, |
| 65 | + zero_bn=zero_bn, |
| 66 | + act_fn=False, |
| 67 | + bn_1st=bn_1st, |
| 68 | + groups=mid_channels if dw else groups, |
| 69 | + ), |
| 70 | + ), |
| 71 | + ] |
| 72 | + else: |
| 73 | + layers = [ |
| 74 | + ( |
| 75 | + "conv_0", |
| 76 | + conv_layer( |
| 77 | + in_channels, |
| 78 | + mid_channels, |
| 79 | + 1, |
| 80 | + act_fn=act_fn, |
| 81 | + bn_1st=bn_1st, |
| 82 | + ), |
| 83 | + ), |
| 84 | + ( |
| 85 | + "conv_1", |
| 86 | + conv_layer( |
| 87 | + mid_channels, |
| 88 | + mid_channels, |
| 89 | + 3, |
| 90 | + stride=stride, |
| 91 | + act_fn=act_fn, |
| 92 | + bn_1st=bn_1st, |
| 93 | + groups=mid_channels if dw else groups, |
| 94 | + ), |
| 95 | + ), |
| 96 | + ( |
| 97 | + "conv_2", |
| 98 | + conv_layer( |
| 99 | + mid_channels, |
| 100 | + out_channels, |
| 101 | + 1, |
| 102 | + zero_bn=zero_bn, |
| 103 | + act_fn=False, |
| 104 | + bn_1st=bn_1st, |
| 105 | + ), |
| 106 | + ), # noqa E501 |
| 107 | + ] |
| 108 | + if se: |
| 109 | + layers.append(("se", se(out_channels))) |
| 110 | + if sa: |
| 111 | + layers.append(("sa", sa(out_channels))) |
| 112 | + self.convs = nn.Sequential(OrderedDict(layers)) |
| 113 | + if stride != 1 or in_channels != out_channels: |
| 114 | + id_layers = [] |
| 115 | + if ( |
| 116 | + stride != 1 and pool is not None |
| 117 | + ): # if pool - reduce by pool else stride 2 art id_conv |
| 118 | + id_layers.append(("pool", pool())) |
| 119 | + if in_channels != out_channels or (stride != 1 and pool is None): |
| 120 | + id_layers += [ |
| 121 | + ( |
| 122 | + "id_conv", |
| 123 | + conv_layer( |
| 124 | + in_channels, |
| 125 | + out_channels, |
| 126 | + 1, |
| 127 | + stride=1 if pool else stride, |
| 128 | + act_fn=False, |
| 129 | + ), |
| 130 | + ) |
| 131 | + ] |
| 132 | + self.id_conv = nn.Sequential(OrderedDict(id_layers)) |
| 133 | + else: |
| 134 | + self.id_conv = None |
| 135 | + self.act_fn = get_act(act_fn) |
| 136 | + |
| 137 | + def forward(self, x): |
| 138 | + identity = self.id_conv(x) if self.id_conv is not None else x |
| 139 | + return self.act_fn(self.convs(x) + identity) |
| 140 | + |
| 141 | + |
| 142 | +class YaResBlock(nn.Module): |
| 143 | + """YaResBlock. Reduce by pool instead of stride 2""" |
| 144 | + |
| 145 | + def __init__( |
| 146 | + self, |
| 147 | + expansion: int, |
| 148 | + in_channels: int, |
| 149 | + mid_channels: int, |
| 150 | + stride: int = 1, |
| 151 | + conv_layer=ConvBnAct, |
| 152 | + act_fn: type[nn.Module] = nn.ReLU, |
| 153 | + zero_bn: bool = True, |
| 154 | + bn_1st: bool = True, |
| 155 | + groups: int = 1, |
| 156 | + dw: bool = False, |
| 157 | + div_groups: Union[None, int] = None, |
| 158 | + pool: Union[Callable[[], nn.Module], None] = None, |
| 159 | + se: Union[type[nn.Module], None] = None, |
| 160 | + sa: Union[type[nn.Module], None] = None, |
| 161 | + ): |
| 162 | + super().__init__() |
| 163 | + # pool defined at ModelConstructor. |
| 164 | + out_channels, in_channels = mid_channels * expansion, in_channels * expansion |
| 165 | + if div_groups is not None: # check if groups != 1 and div_groups |
| 166 | + groups = int(mid_channels / div_groups) |
| 167 | + |
| 168 | + if stride != 1: |
| 169 | + if pool is None: |
| 170 | + self.reduce = conv_layer(in_channels, in_channels, 1, stride=2) |
| 171 | + # warnings.warn("pool not passed") # need to warn? |
| 172 | + else: |
| 173 | + self.reduce = pool() |
| 174 | + else: |
| 175 | + self.reduce = None |
| 176 | + if expansion == 1: |
| 177 | + layers = [ |
| 178 | + ( |
| 179 | + "conv_0", |
| 180 | + conv_layer( |
| 181 | + in_channels, |
| 182 | + mid_channels, |
| 183 | + 3, |
| 184 | + stride=1, |
| 185 | + act_fn=act_fn, |
| 186 | + bn_1st=bn_1st, |
| 187 | + groups=in_channels if dw else groups, |
| 188 | + ), |
| 189 | + ), |
| 190 | + ( |
| 191 | + "conv_1", |
| 192 | + conv_layer( |
| 193 | + mid_channels, |
| 194 | + out_channels, |
| 195 | + 3, |
| 196 | + zero_bn=zero_bn, |
| 197 | + act_fn=False, |
| 198 | + bn_1st=bn_1st, |
| 199 | + groups=mid_channels if dw else groups, |
| 200 | + ), |
| 201 | + ), |
| 202 | + ] |
| 203 | + else: |
| 204 | + layers = [ |
| 205 | + ( |
| 206 | + "conv_0", |
| 207 | + conv_layer( |
| 208 | + in_channels, |
| 209 | + mid_channels, |
| 210 | + 1, |
| 211 | + act_fn=act_fn, |
| 212 | + bn_1st=bn_1st, |
| 213 | + ), |
| 214 | + ), |
| 215 | + ( |
| 216 | + "conv_1", |
| 217 | + conv_layer( |
| 218 | + mid_channels, |
| 219 | + mid_channels, |
| 220 | + 3, |
| 221 | + stride=1, |
| 222 | + act_fn=act_fn, |
| 223 | + bn_1st=bn_1st, |
| 224 | + groups=mid_channels if dw else groups, |
| 225 | + ), |
| 226 | + ), |
| 227 | + ( |
| 228 | + "conv_2", |
| 229 | + conv_layer( |
| 230 | + mid_channels, |
| 231 | + out_channels, |
| 232 | + 1, |
| 233 | + zero_bn=zero_bn, |
| 234 | + act_fn=False, |
| 235 | + bn_1st=bn_1st, |
| 236 | + ), |
| 237 | + ), # noqa E501 |
| 238 | + ] |
| 239 | + if se: |
| 240 | + layers.append(("se", se(out_channels))) # type: ignore |
| 241 | + if sa: |
| 242 | + layers.append(("sa", sa(out_channels))) # type: ignore |
| 243 | + self.convs = nn.Sequential(OrderedDict(layers)) |
| 244 | + if in_channels != out_channels: |
| 245 | + self.id_conv = conv_layer( |
| 246 | + in_channels, |
| 247 | + out_channels, |
| 248 | + 1, |
| 249 | + stride=1, |
| 250 | + act_fn=False, |
| 251 | + ) |
| 252 | + else: |
| 253 | + self.id_conv = None |
| 254 | + self.merge = get_act(act_fn) |
| 255 | + |
| 256 | + def forward(self, x): |
| 257 | + if self.reduce: |
| 258 | + x = self.reduce(x) |
| 259 | + identity = self.id_conv(x) if self.id_conv is not None else x |
| 260 | + return self.merge(self.convs(x) + identity) |
| 261 | + |
| 262 | + |
| 263 | +def make_stem(cfg: ModelCfg) -> nn.Sequential: # type: ignore |
| 264 | + """Create xResnet stem -> 3 conv 3*3 instead 1 conv 7*7""" |
| 265 | + len_stem = len(cfg.stem_sizes) |
| 266 | + stem: list[tuple[str, nn.Module]] = [ |
| 267 | + ( |
| 268 | + f"conv_{i}", |
| 269 | + cfg.conv_layer( |
| 270 | + cfg.stem_sizes[i - 1] if i else cfg.in_chans, # type: ignore |
| 271 | + cfg.stem_sizes[i], |
| 272 | + stride=2 if i == cfg.stem_stride_on else 1, |
| 273 | + bn_layer=(not cfg.stem_bn_end) if i == (len_stem - 1) else True, |
| 274 | + act_fn=cfg.act_fn, |
| 275 | + bn_1st=cfg.bn_1st, |
| 276 | + ), |
| 277 | + ) |
| 278 | + for i in range(len_stem) |
| 279 | + ] |
| 280 | + if cfg.stem_pool: |
| 281 | + stem.append(("stem_pool", cfg.stem_pool())) |
| 282 | + if cfg.stem_bn_end: |
| 283 | + stem.append(("norm", cfg.norm(cfg.stem_sizes[-1]))) # type: ignore |
| 284 | + return nn.Sequential(OrderedDict(stem)) |
| 285 | + |
| 286 | + |
| 287 | +def make_layer(cfg: ModelCfg, layer_num: int) -> nn.Sequential: # type: ignore |
| 288 | + """Create layer (stage)""" |
| 289 | + # if no pool on stem - stride = 2 for first layer block in body |
| 290 | + stride = 1 if cfg.stem_pool and layer_num == 0 else 2 |
| 291 | + num_blocks = cfg.layers[layer_num] |
| 292 | + block_chs = [cfg.stem_sizes[-1] // cfg.expansion] + cfg.block_sizes |
| 293 | + return nn.Sequential( |
| 294 | + OrderedDict( |
| 295 | + [ |
| 296 | + ( |
| 297 | + f"bl_{block_num}", |
| 298 | + cfg.block( |
| 299 | + cfg.expansion, # type: ignore |
| 300 | + block_chs[layer_num] |
| 301 | + if block_num == 0 |
| 302 | + else block_chs[layer_num + 1], |
| 303 | + block_chs[layer_num + 1], |
| 304 | + stride if block_num == 0 else 1, |
| 305 | + sa=cfg.sa |
| 306 | + if (block_num == num_blocks - 1) and layer_num == 0 |
| 307 | + else None, |
| 308 | + conv_layer=cfg.conv_layer, |
| 309 | + act_fn=cfg.act_fn, |
| 310 | + pool=cfg.pool, |
| 311 | + zero_bn=cfg.zero_bn, |
| 312 | + bn_1st=cfg.bn_1st, |
| 313 | + groups=cfg.groups, |
| 314 | + div_groups=cfg.div_groups, |
| 315 | + dw=cfg.dw, |
| 316 | + se=cfg.se, |
| 317 | + ), |
| 318 | + ) |
| 319 | + for block_num in range(num_blocks) |
| 320 | + ] |
| 321 | + ) |
| 322 | + ) |
| 323 | + |
| 324 | + |
| 325 | +def make_body(cfg: ModelCfg) -> nn.Sequential: # type: ignore |
| 326 | + """Create model body.""" |
| 327 | + return nn.Sequential( |
| 328 | + OrderedDict( |
| 329 | + [ |
| 330 | + (f"l_{layer_num}", cfg.make_layer(cfg, layer_num)) # type: ignore |
| 331 | + for layer_num in range(len(cfg.layers)) |
| 332 | + ] |
| 333 | + ) |
| 334 | + ) |
| 335 | + |
| 336 | + |
| 337 | +class XResNet(ModelConstructor): |
| 338 | + """Base Xresnet constructor.""" |
| 339 | + make_stem: Callable[[ModelCfg], Union[nn.Module, nn.Sequential]] = make_stem |
| 340 | + make_layer: Callable[[ModelCfg, int], Union[nn.Module, nn.Sequential]] = make_layer |
| 341 | + make_body: Callable[[ModelCfg], Union[nn.Module, nn.Sequential]] = make_body |
| 342 | + block: type[nn.Module] = XResBlock |
| 343 | + |
| 344 | + |
| 345 | +class XResNet34(XResNet): |
| 346 | + layers: list[int] = [3, 4, 6, 3] |
| 347 | + |
| 348 | + |
| 349 | +class XResNet50(XResNet34): |
| 350 | + expansion: int = 4 |
| 351 | + |
| 352 | + |
| 353 | +class YaResNet(XResNet): |
| 354 | + """Base Yaresnet constructor. |
| 355 | + YaResBlock, Mish activation, custom stem. |
| 356 | + """ |
| 357 | + block: type[nn.Module] = YaResBlock |
| 358 | + stem_sizes: list[int] = [3, 32, 64, 64] |
| 359 | + act_fn: type[nn.Module] = nn.Mish |
| 360 | + |
| 361 | + |
| 362 | +class YaResNet34(YaResNet): |
| 363 | + layers: list[int] = [3, 4, 6, 3] |
| 364 | + |
| 365 | + |
| 366 | +class YaResNet50(YaResNet34): |
| 367 | + expansion: int = 4 |
0 commit comments