File size: 13,693 Bytes
9c6594c |
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 |
"""Pytorch impl of Aligned Xception 41, 65, 71
This is a correct, from scratch impl of Aligned Xception (Deeplab) models compatible with TF weights at
https://github.com/tensorflow/models/blob/master/research/deeplab/g3doc/model_zoo.md
Hacked together by / Copyright 2020 Ross Wightman
"""
from functools import partial
import torch
import torch.nn as nn
from timm.data import IMAGENET_INCEPTION_MEAN, IMAGENET_INCEPTION_STD
from .helpers import build_model_with_cfg, checkpoint_seq
from .layers import ClassifierHead, ConvNormAct, create_conv2d, get_norm_act_layer
from .layers.helpers import to_3tuple
from .registry import register_model
__all__ = ['XceptionAligned']
def _cfg(url='', **kwargs):
return {
'url': url,
'num_classes': 1000, 'input_size': (3, 299, 299), 'pool_size': (10, 10),
'crop_pct': 0.903, 'interpolation': 'bicubic',
'mean': IMAGENET_INCEPTION_MEAN, 'std': IMAGENET_INCEPTION_STD,
'first_conv': 'stem.0.conv', 'classifier': 'head.fc',
**kwargs
}
default_cfgs = dict(
xception41=_cfg(
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/tf_xception_41-e6439c97.pth'),
xception65=_cfg(
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-tpu-weights/xception65_ra3-1447db8d.pth',
crop_pct=0.94,
),
xception71=_cfg(
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/tf_xception_71-8eec7df1.pth'),
xception41p=_cfg(
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-tpu-weights/xception41p_ra3-33195bc8.pth',
crop_pct=0.94,
),
xception65p=_cfg(
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-tpu-weights/xception65p_ra3-3c6114e4.pth',
crop_pct=0.94,
),
)
class SeparableConv2d(nn.Module):
def __init__(
self, in_chs, out_chs, kernel_size=3, stride=1, dilation=1, padding='',
act_layer=nn.ReLU, norm_layer=nn.BatchNorm2d):
super(SeparableConv2d, self).__init__()
self.kernel_size = kernel_size
self.dilation = dilation
# depthwise convolution
self.conv_dw = create_conv2d(
in_chs, in_chs, kernel_size, stride=stride,
padding=padding, dilation=dilation, depthwise=True)
self.bn_dw = norm_layer(in_chs)
self.act_dw = act_layer(inplace=True) if act_layer is not None else nn.Identity()
# pointwise convolution
self.conv_pw = create_conv2d(in_chs, out_chs, kernel_size=1)
self.bn_pw = norm_layer(out_chs)
self.act_pw = act_layer(inplace=True) if act_layer is not None else nn.Identity()
def forward(self, x):
x = self.conv_dw(x)
x = self.bn_dw(x)
x = self.act_dw(x)
x = self.conv_pw(x)
x = self.bn_pw(x)
x = self.act_pw(x)
return x
class PreSeparableConv2d(nn.Module):
def __init__(
self, in_chs, out_chs, kernel_size=3, stride=1, dilation=1, padding='',
act_layer=nn.ReLU, norm_layer=nn.BatchNorm2d, first_act=True):
super(PreSeparableConv2d, self).__init__()
norm_act_layer = get_norm_act_layer(norm_layer, act_layer=act_layer)
self.kernel_size = kernel_size
self.dilation = dilation
self.norm = norm_act_layer(in_chs, inplace=True) if first_act else nn.Identity()
# depthwise convolution
self.conv_dw = create_conv2d(
in_chs, in_chs, kernel_size, stride=stride,
padding=padding, dilation=dilation, depthwise=True)
# pointwise convolution
self.conv_pw = create_conv2d(in_chs, out_chs, kernel_size=1)
def forward(self, x):
x = self.norm(x)
x = self.conv_dw(x)
x = self.conv_pw(x)
return x
class XceptionModule(nn.Module):
def __init__(
self, in_chs, out_chs, stride=1, dilation=1, pad_type='',
start_with_relu=True, no_skip=False, act_layer=nn.ReLU, norm_layer=None):
super(XceptionModule, self).__init__()
out_chs = to_3tuple(out_chs)
self.in_channels = in_chs
self.out_channels = out_chs[-1]
self.no_skip = no_skip
if not no_skip and (self.out_channels != self.in_channels or stride != 1):
self.shortcut = ConvNormAct(
in_chs, self.out_channels, 1, stride=stride, norm_layer=norm_layer, apply_act=False)
else:
self.shortcut = None
separable_act_layer = None if start_with_relu else act_layer
self.stack = nn.Sequential()
for i in range(3):
if start_with_relu:
self.stack.add_module(f'act{i + 1}', act_layer(inplace=i > 0))
self.stack.add_module(f'conv{i + 1}', SeparableConv2d(
in_chs, out_chs[i], 3, stride=stride if i == 2 else 1, dilation=dilation, padding=pad_type,
act_layer=separable_act_layer, norm_layer=norm_layer))
in_chs = out_chs[i]
def forward(self, x):
skip = x
x = self.stack(x)
if self.shortcut is not None:
skip = self.shortcut(skip)
if not self.no_skip:
x = x + skip
return x
class PreXceptionModule(nn.Module):
def __init__(
self, in_chs, out_chs, stride=1, dilation=1, pad_type='',
no_skip=False, act_layer=nn.ReLU, norm_layer=None):
super(PreXceptionModule, self).__init__()
out_chs = to_3tuple(out_chs)
self.in_channels = in_chs
self.out_channels = out_chs[-1]
self.no_skip = no_skip
if not no_skip and (self.out_channels != self.in_channels or stride != 1):
self.shortcut = create_conv2d(in_chs, self.out_channels, 1, stride=stride)
else:
self.shortcut = nn.Identity()
self.norm = get_norm_act_layer(norm_layer, act_layer=act_layer)(in_chs, inplace=True)
self.stack = nn.Sequential()
for i in range(3):
self.stack.add_module(f'conv{i + 1}', PreSeparableConv2d(
in_chs, out_chs[i], 3, stride=stride if i == 2 else 1, dilation=dilation, padding=pad_type,
act_layer=act_layer, norm_layer=norm_layer, first_act=i > 0))
in_chs = out_chs[i]
def forward(self, x):
x = self.norm(x)
skip = x
x = self.stack(x)
if not self.no_skip:
x = x + self.shortcut(skip)
return x
class XceptionAligned(nn.Module):
"""Modified Aligned Xception
"""
def __init__(
self, block_cfg, num_classes=1000, in_chans=3, output_stride=32, preact=False,
act_layer=nn.ReLU, norm_layer=nn.BatchNorm2d, drop_rate=0., global_pool='avg'):
super(XceptionAligned, self).__init__()
assert output_stride in (8, 16, 32)
self.num_classes = num_classes
self.drop_rate = drop_rate
self.grad_checkpointing = False
layer_args = dict(act_layer=act_layer, norm_layer=norm_layer)
self.stem = nn.Sequential(*[
ConvNormAct(in_chans, 32, kernel_size=3, stride=2, **layer_args),
create_conv2d(32, 64, kernel_size=3, stride=1) if preact else
ConvNormAct(32, 64, kernel_size=3, stride=1, **layer_args)
])
curr_dilation = 1
curr_stride = 2
self.feature_info = []
self.blocks = nn.Sequential()
module_fn = PreXceptionModule if preact else XceptionModule
for i, b in enumerate(block_cfg):
b['dilation'] = curr_dilation
if b['stride'] > 1:
name = f'blocks.{i}.stack.conv2' if preact else f'blocks.{i}.stack.act3'
self.feature_info += [dict(num_chs=to_3tuple(b['out_chs'])[-2], reduction=curr_stride, module=name)]
next_stride = curr_stride * b['stride']
if next_stride > output_stride:
curr_dilation *= b['stride']
b['stride'] = 1
else:
curr_stride = next_stride
self.blocks.add_module(str(i), module_fn(**b, **layer_args))
self.num_features = self.blocks[-1].out_channels
self.feature_info += [dict(
num_chs=self.num_features, reduction=curr_stride, module='blocks.' + str(len(self.blocks) - 1))]
self.act = act_layer(inplace=True) if preact else nn.Identity()
self.head = ClassifierHead(
in_chs=self.num_features, num_classes=num_classes, pool_type=global_pool, drop_rate=drop_rate)
@torch.jit.ignore
def group_matcher(self, coarse=False):
return dict(
stem=r'^stem',
blocks=r'^blocks\.(\d+)',
)
@torch.jit.ignore
def set_grad_checkpointing(self, enable=True):
self.grad_checkpointing = enable
@torch.jit.ignore
def get_classifier(self):
return self.head.fc
def reset_classifier(self, num_classes, global_pool='avg'):
self.head = ClassifierHead(self.num_features, num_classes, pool_type=global_pool, drop_rate=self.drop_rate)
def forward_features(self, x):
x = self.stem(x)
if self.grad_checkpointing and not torch.jit.is_scripting():
x = checkpoint_seq(self.blocks, x)
else:
x = self.blocks(x)
x = self.act(x)
return x
def forward_head(self, x, pre_logits: bool = False):
return self.head(x, pre_logits=pre_logits)
def forward(self, x):
x = self.forward_features(x)
x = self.forward_head(x)
return x
def _xception(variant, pretrained=False, **kwargs):
return build_model_with_cfg(
XceptionAligned, variant, pretrained,
feature_cfg=dict(flatten_sequential=True, feature_cls='hook'),
**kwargs)
@register_model
def xception41(pretrained=False, **kwargs):
""" Modified Aligned Xception-41
"""
block_cfg = [
# entry flow
dict(in_chs=64, out_chs=128, stride=2),
dict(in_chs=128, out_chs=256, stride=2),
dict(in_chs=256, out_chs=728, stride=2),
# middle flow
*([dict(in_chs=728, out_chs=728, stride=1)] * 8),
# exit flow
dict(in_chs=728, out_chs=(728, 1024, 1024), stride=2),
dict(in_chs=1024, out_chs=(1536, 1536, 2048), stride=1, no_skip=True, start_with_relu=False),
]
model_args = dict(block_cfg=block_cfg, norm_layer=partial(nn.BatchNorm2d, eps=.001, momentum=.1), **kwargs)
return _xception('xception41', pretrained=pretrained, **model_args)
@register_model
def xception65(pretrained=False, **kwargs):
""" Modified Aligned Xception-65
"""
block_cfg = [
# entry flow
dict(in_chs=64, out_chs=128, stride=2),
dict(in_chs=128, out_chs=256, stride=2),
dict(in_chs=256, out_chs=728, stride=2),
# middle flow
*([dict(in_chs=728, out_chs=728, stride=1)] * 16),
# exit flow
dict(in_chs=728, out_chs=(728, 1024, 1024), stride=2),
dict(in_chs=1024, out_chs=(1536, 1536, 2048), stride=1, no_skip=True, start_with_relu=False),
]
model_args = dict(block_cfg=block_cfg, norm_layer=partial(nn.BatchNorm2d, eps=.001, momentum=.1), **kwargs)
return _xception('xception65', pretrained=pretrained, **model_args)
@register_model
def xception71(pretrained=False, **kwargs):
""" Modified Aligned Xception-71
"""
block_cfg = [
# entry flow
dict(in_chs=64, out_chs=128, stride=2),
dict(in_chs=128, out_chs=256, stride=1),
dict(in_chs=256, out_chs=256, stride=2),
dict(in_chs=256, out_chs=728, stride=1),
dict(in_chs=728, out_chs=728, stride=2),
# middle flow
*([dict(in_chs=728, out_chs=728, stride=1)] * 16),
# exit flow
dict(in_chs=728, out_chs=(728, 1024, 1024), stride=2),
dict(in_chs=1024, out_chs=(1536, 1536, 2048), stride=1, no_skip=True, start_with_relu=False),
]
model_args = dict(block_cfg=block_cfg, norm_layer=partial(nn.BatchNorm2d, eps=.001, momentum=.1), **kwargs)
return _xception('xception71', pretrained=pretrained, **model_args)
@register_model
def xception41p(pretrained=False, **kwargs):
""" Modified Aligned Xception-41 w/ Pre-Act
"""
block_cfg = [
# entry flow
dict(in_chs=64, out_chs=128, stride=2),
dict(in_chs=128, out_chs=256, stride=2),
dict(in_chs=256, out_chs=728, stride=2),
# middle flow
*([dict(in_chs=728, out_chs=728, stride=1)] * 8),
# exit flow
dict(in_chs=728, out_chs=(728, 1024, 1024), stride=2),
dict(in_chs=1024, out_chs=(1536, 1536, 2048), no_skip=True, stride=1),
]
model_args = dict(block_cfg=block_cfg, preact=True, norm_layer=nn.BatchNorm2d, **kwargs)
return _xception('xception41p', pretrained=pretrained, **model_args)
@register_model
def xception65p(pretrained=False, **kwargs):
""" Modified Aligned Xception-65 w/ Pre-Act
"""
block_cfg = [
# entry flow
dict(in_chs=64, out_chs=128, stride=2),
dict(in_chs=128, out_chs=256, stride=2),
dict(in_chs=256, out_chs=728, stride=2),
# middle flow
*([dict(in_chs=728, out_chs=728, stride=1)] * 16),
# exit flow
dict(in_chs=728, out_chs=(728, 1024, 1024), stride=2),
dict(in_chs=1024, out_chs=(1536, 1536, 2048), stride=1, no_skip=True),
]
model_args = dict(
block_cfg=block_cfg, preact=True, norm_layer=partial(nn.BatchNorm2d, eps=.001, momentum=.1), **kwargs)
return _xception('xception65p', pretrained=pretrained, **model_args)
|