File size: 13,196 Bytes
b5f33fd |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 |
"""
Code adapted from timm https://github.com/huggingface/pytorch-image-models
Modifications and additions for mivolo by / Copyright 2023, Irina Tolstykh, Maxim Kuprashevich
"""
import torch
import torch.nn as nn
from mivolo.model.cross_bottleneck_attn import CrossBottleneckAttn
from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
from timm.layers import trunc_normal_
from timm.models._builder import build_model_with_cfg
from timm.models._registry import register_model
from timm.models.volo import VOLO
__all__ = ["MiVOLOModel"] # model_registry will add each entrypoint fn to this
def _cfg(url="", **kwargs):
return {
"url": url,
"num_classes": 1000,
"input_size": (3, 224, 224),
"pool_size": None,
"crop_pct": 0.96,
"interpolation": "bicubic",
"fixed_input_size": True,
"mean": IMAGENET_DEFAULT_MEAN,
"std": IMAGENET_DEFAULT_STD,
"first_conv": None,
"classifier": ("head", "aux_head"),
**kwargs,
}
default_cfgs = {
"mivolo_d1_224": _cfg(
url="https://github.com/sail-sg/volo/releases/download/volo_1/d1_224_84.2.pth.tar", crop_pct=0.96
),
"mivolo_d1_384": _cfg(
url="https://github.com/sail-sg/volo/releases/download/volo_1/d1_384_85.2.pth.tar",
crop_pct=1.0,
input_size=(3, 384, 384),
),
"mivolo_d2_224": _cfg(
url="https://github.com/sail-sg/volo/releases/download/volo_1/d2_224_85.2.pth.tar", crop_pct=0.96
),
"mivolo_d2_384": _cfg(
url="https://github.com/sail-sg/volo/releases/download/volo_1/d2_384_86.0.pth.tar",
crop_pct=1.0,
input_size=(3, 384, 384),
),
"mivolo_d3_224": _cfg(
url="https://github.com/sail-sg/volo/releases/download/volo_1/d3_224_85.4.pth.tar", crop_pct=0.96
),
"mivolo_d3_448": _cfg(
url="https://github.com/sail-sg/volo/releases/download/volo_1/d3_448_86.3.pth.tar",
crop_pct=1.0,
input_size=(3, 448, 448),
),
"mivolo_d4_224": _cfg(
url="https://github.com/sail-sg/volo/releases/download/volo_1/d4_224_85.7.pth.tar", crop_pct=0.96
),
"mivolo_d4_448": _cfg(
url="https://github.com/sail-sg/volo/releases/download/volo_1/d4_448_86.79.pth.tar",
crop_pct=1.15,
input_size=(3, 448, 448),
),
"mivolo_d5_224": _cfg(
url="https://github.com/sail-sg/volo/releases/download/volo_1/d5_224_86.10.pth.tar", crop_pct=0.96
),
"mivolo_d5_448": _cfg(
url="https://github.com/sail-sg/volo/releases/download/volo_1/d5_448_87.0.pth.tar",
crop_pct=1.15,
input_size=(3, 448, 448),
),
"mivolo_d5_512": _cfg(
url="https://github.com/sail-sg/volo/releases/download/volo_1/d5_512_87.07.pth.tar",
crop_pct=1.15,
input_size=(3, 512, 512),
),
}
def get_output_size(input_shape, conv_layer):
padding = conv_layer.padding
dilation = conv_layer.dilation
kernel_size = conv_layer.kernel_size
stride = conv_layer.stride
output_size = [
((input_shape[i] + 2 * padding[i] - dilation[i] * (kernel_size[i] - 1) - 1) // stride[i]) + 1 for i in range(2)
]
return output_size
def get_output_size_module(input_size, stem):
output_size = input_size
for module in stem:
if isinstance(module, nn.Conv2d):
output_size = [
(
(output_size[i] + 2 * module.padding[i] - module.dilation[i] * (module.kernel_size[i] - 1) - 1)
// module.stride[i]
)
+ 1
for i in range(2)
]
return output_size
class PatchEmbed(nn.Module):
"""Image to Patch Embedding."""
def __init__(
self, img_size=224, stem_conv=False, stem_stride=1, patch_size=8, in_chans=3, hidden_dim=64, embed_dim=384
):
super().__init__()
assert patch_size in [4, 8, 16]
assert in_chans in [3, 6]
self.with_persons_model = in_chans == 6
self.use_cross_attn = True
if stem_conv:
if not self.with_persons_model:
self.conv = self.create_stem(stem_stride, in_chans, hidden_dim)
else:
self.conv = True # just to match interface
# split
self.conv1 = self.create_stem(stem_stride, 3, hidden_dim)
self.conv2 = self.create_stem(stem_stride, 3, hidden_dim)
else:
self.conv = None
if self.with_persons_model:
self.proj1 = nn.Conv2d(
hidden_dim, embed_dim, kernel_size=patch_size // stem_stride, stride=patch_size // stem_stride
)
self.proj2 = nn.Conv2d(
hidden_dim, embed_dim, kernel_size=patch_size // stem_stride, stride=patch_size // stem_stride
)
stem_out_shape = get_output_size_module((img_size, img_size), self.conv1)
self.proj_output_size = get_output_size(stem_out_shape, self.proj1)
self.map = CrossBottleneckAttn(embed_dim, dim_out=embed_dim, num_heads=1, feat_size=self.proj_output_size)
else:
self.proj = nn.Conv2d(
hidden_dim, embed_dim, kernel_size=patch_size // stem_stride, stride=patch_size // stem_stride
)
self.patch_dim = img_size // patch_size
self.num_patches = self.patch_dim**2
def create_stem(self, stem_stride, in_chans, hidden_dim):
return nn.Sequential(
nn.Conv2d(in_chans, hidden_dim, kernel_size=7, stride=stem_stride, padding=3, bias=False), # 112x112
nn.BatchNorm2d(hidden_dim),
nn.ReLU(inplace=True),
nn.Conv2d(hidden_dim, hidden_dim, kernel_size=3, stride=1, padding=1, bias=False), # 112x112
nn.BatchNorm2d(hidden_dim),
nn.ReLU(inplace=True),
nn.Conv2d(hidden_dim, hidden_dim, kernel_size=3, stride=1, padding=1, bias=False), # 112x112
nn.BatchNorm2d(hidden_dim),
nn.ReLU(inplace=True),
)
def forward(self, x):
if self.conv is not None:
if self.with_persons_model:
x1 = x[:, :3]
x2 = x[:, 3:]
x1 = self.conv1(x1)
x1 = self.proj1(x1)
x2 = self.conv2(x2)
x2 = self.proj2(x2)
x = torch.cat([x1, x2], dim=1)
x = self.map(x)
else:
x = self.conv(x)
x = self.proj(x) # B, C, H, W
return x
class MiVOLOModel(VOLO):
"""
Vision Outlooker, the main class of our model
"""
def __init__(
self,
layers,
img_size=224,
in_chans=3,
num_classes=1000,
global_pool="token",
patch_size=8,
stem_hidden_dim=64,
embed_dims=None,
num_heads=None,
downsamples=(True, False, False, False),
outlook_attention=(True, False, False, False),
mlp_ratio=3.0,
qkv_bias=False,
drop_rate=0.0,
attn_drop_rate=0.0,
drop_path_rate=0.0,
norm_layer=nn.LayerNorm,
post_layers=("ca", "ca"),
use_aux_head=True,
use_mix_token=False,
pooling_scale=2,
):
super().__init__(
layers,
img_size,
in_chans,
num_classes,
global_pool,
patch_size,
stem_hidden_dim,
embed_dims,
num_heads,
downsamples,
outlook_attention,
mlp_ratio,
qkv_bias,
drop_rate,
attn_drop_rate,
drop_path_rate,
norm_layer,
post_layers,
use_aux_head,
use_mix_token,
pooling_scale,
)
self.patch_embed = PatchEmbed(
stem_conv=True,
stem_stride=2,
patch_size=patch_size,
in_chans=in_chans,
hidden_dim=stem_hidden_dim,
embed_dim=embed_dims[0],
)
trunc_normal_(self.pos_embed, std=0.02)
self.apply(self._init_weights)
def forward_features(self, x):
x = self.patch_embed(x).permute(0, 2, 3, 1) # B,C,H,W-> B,H,W,C
# step2: tokens learning in the two stages
x = self.forward_tokens(x)
# step3: post network, apply class attention or not
if self.post_network is not None:
x = self.forward_cls(x)
x = self.norm(x)
return x
def forward_head(self, x, pre_logits: bool = False, targets=None, epoch=None):
if self.global_pool == "avg":
out = x.mean(dim=1)
elif self.global_pool == "token":
out = x[:, 0]
else:
out = x
if pre_logits:
return out
features = out
fds_enabled = hasattr(self, "_fds_forward")
if fds_enabled:
features = self._fds_forward(features, targets, epoch)
out = self.head(features)
if self.aux_head is not None:
# generate classes in all feature tokens, see token labeling
aux = self.aux_head(x[:, 1:])
out = out + 0.5 * aux.max(1)[0]
return (out, features) if (fds_enabled and self.training) else out
def forward(self, x, targets=None, epoch=None):
"""simplified forward (without mix token training)"""
x = self.forward_features(x)
x = self.forward_head(x, targets=targets, epoch=epoch)
return x
def _create_mivolo(variant, pretrained=False, **kwargs):
if kwargs.get("features_only", None):
raise RuntimeError("features_only not implemented for Vision Transformer models.")
return build_model_with_cfg(MiVOLOModel, variant, pretrained, **kwargs)
@register_model
def mivolo_d1_224(pretrained=False, **kwargs):
model_args = dict(layers=(4, 4, 8, 2), embed_dims=(192, 384, 384, 384), num_heads=(6, 12, 12, 12), **kwargs)
model = _create_mivolo("mivolo_d1_224", pretrained=pretrained, **model_args)
return model
@register_model
def mivolo_d1_384(pretrained=False, **kwargs):
model_args = dict(layers=(4, 4, 8, 2), embed_dims=(192, 384, 384, 384), num_heads=(6, 12, 12, 12), **kwargs)
model = _create_mivolo("mivolo_d1_384", pretrained=pretrained, **model_args)
return model
@register_model
def mivolo_d2_224(pretrained=False, **kwargs):
model_args = dict(layers=(6, 4, 10, 4), embed_dims=(256, 512, 512, 512), num_heads=(8, 16, 16, 16), **kwargs)
model = _create_mivolo("mivolo_d2_224", pretrained=pretrained, **model_args)
return model
@register_model
def mivolo_d2_384(pretrained=False, **kwargs):
model_args = dict(layers=(6, 4, 10, 4), embed_dims=(256, 512, 512, 512), num_heads=(8, 16, 16, 16), **kwargs)
model = _create_mivolo("mivolo_d2_384", pretrained=pretrained, **model_args)
return model
@register_model
def mivolo_d3_224(pretrained=False, **kwargs):
model_args = dict(layers=(8, 8, 16, 4), embed_dims=(256, 512, 512, 512), num_heads=(8, 16, 16, 16), **kwargs)
model = _create_mivolo("mivolo_d3_224", pretrained=pretrained, **model_args)
return model
@register_model
def mivolo_d3_448(pretrained=False, **kwargs):
model_args = dict(layers=(8, 8, 16, 4), embed_dims=(256, 512, 512, 512), num_heads=(8, 16, 16, 16), **kwargs)
model = _create_mivolo("mivolo_d3_448", pretrained=pretrained, **model_args)
return model
@register_model
def mivolo_d4_224(pretrained=False, **kwargs):
model_args = dict(layers=(8, 8, 16, 4), embed_dims=(384, 768, 768, 768), num_heads=(12, 16, 16, 16), **kwargs)
model = _create_mivolo("mivolo_d4_224", pretrained=pretrained, **model_args)
return model
@register_model
def mivolo_d4_448(pretrained=False, **kwargs):
"""VOLO-D4 model, Params: 193M"""
model_args = dict(layers=(8, 8, 16, 4), embed_dims=(384, 768, 768, 768), num_heads=(12, 16, 16, 16), **kwargs)
model = _create_mivolo("mivolo_d4_448", pretrained=pretrained, **model_args)
return model
@register_model
def mivolo_d5_224(pretrained=False, **kwargs):
model_args = dict(
layers=(12, 12, 20, 4),
embed_dims=(384, 768, 768, 768),
num_heads=(12, 16, 16, 16),
mlp_ratio=4,
stem_hidden_dim=128,
**kwargs
)
model = _create_mivolo("mivolo_d5_224", pretrained=pretrained, **model_args)
return model
@register_model
def mivolo_d5_448(pretrained=False, **kwargs):
model_args = dict(
layers=(12, 12, 20, 4),
embed_dims=(384, 768, 768, 768),
num_heads=(12, 16, 16, 16),
mlp_ratio=4,
stem_hidden_dim=128,
**kwargs
)
model = _create_mivolo("mivolo_d5_448", pretrained=pretrained, **model_args)
return model
@register_model
def mivolo_d5_512(pretrained=False, **kwargs):
model_args = dict(
layers=(12, 12, 20, 4),
embed_dims=(384, 768, 768, 768),
num_heads=(12, 16, 16, 16),
mlp_ratio=4,
stem_hidden_dim=128,
**kwargs
)
model = _create_mivolo("mivolo_d5_512", pretrained=pretrained, **model_args)
return model
|