File size: 15,370 Bytes
793ef19 d80dc9a 793ef19 d80dc9a 793ef19 d80dc9a 3f1b078 793ef19 3f1b078 793ef19 3f1b078 793ef19 d80dc9a 793ef19 d80dc9a 793ef19 d80dc9a 793ef19 d80dc9a 793ef19 d80dc9a 793ef19 d80dc9a 793ef19 d80dc9a 793ef19 d80dc9a 793ef19 d80dc9a 793ef19 d80dc9a |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 |
"""PyTorch ResNet
This started as a copy of https://github.com/pytorch/vision 'resnet.py' (BSD-3-Clause) with
additional dropout and dynamic global avg/max pool.
ResNeXt, SE-ResNeXt, SENet, and MXNet Gluon stem/downsample variants, tiered stems added by Ross Wightman
Copyright 2019, Ross Wightman
"""
import math
from functools import partial
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
from timm.layers import DropBlock2d, DropPath, AvgPool2dSame, BlurPool2d, GroupNorm, create_attn, get_attn, \
get_act_layer, get_norm_layer, create_classifier, LayerNorm2d
from ._builder import build_model_with_cfg
from ._registry import register_model, model_entrypoint
def get_padding(kernel_size, stride, dilation=1):
padding = ((stride - 1) + dilation * (kernel_size - 1)) // 2
return padding
class softball(nn.Module):
def __init__(self, radius2=None, inplace=True):
super(softball, self).__init__()
self.radius2 = radius2 if radius2 is not None else None
def forward(self, x):
if self.radius2 is None:
self.radius2 = x.size()[1]
norm = torch.sqrt(1 + (x*x).sum(1, keepdim=True) / self.radius2)
return x / norm
class hardball(nn.Module):
def __init__(self, radius2=None):
super(hardball, self).__init__()
self.radius = np.sqrt(radius2) if radius2 is not None else None
def forward(self, x):
norm = torch.sqrt((x*x).sum(1, keepdim=True))
if self.radius is None:
self.radius = np.sqrt(x.size()[1])
return torch.where(norm > self.radius, self.radius * x / norm, x)
class ConvBN(nn.Module):
def __init__(self, conv, bn):
super(ConvBN, self).__init__()
self.conv = conv
self.bn = bn
self.fused_weight = None
self.fused_bias = None
def forward(self, x):
if self.training:
x = self.conv(x)
x = self.bn(x)
else:
if self.fused_weight is not None and self.fused_bias is not None:
x = F.conv2d(x, self.fused_weight, self.fused_bias,
self.conv.stride, self.conv.padding,
self.conv.dilation, self.conv.groups)
else:
x = self.conv(x)
x = self.bn(x)
return x
def fuse_bn(self):
if self.training:
raise RuntimeError("Call fuse_bn only in eval mode")
# Calculate the fused weight and bias
w = self.conv.weight
mean = self.bn.running_mean
var = torch.sqrt(self.bn.running_var + self.bn.eps)
gamma = self.bn.weight
beta = self.bn.bias
self.fused_weight = w * (gamma / var).reshape(-1, 1, 1, 1)
self.fused_bias = beta - (gamma * mean / var)
class QLBlock(nn.Module): # quasilinear hyperbolic system
expansion = 1
def __init__(
self,
inplanes,
planes,
stride=1,
downsample=None,
cardinality=1,
base_width=64,
reduce_first=1,
dilation=1,
first_dilation=None,
act_layer=nn.ReLU,
norm_layer=nn.BatchNorm2d,
):
super(QLBlock, self).__init__()
self.k = 8 if inplanes <= 128 else 4 if inplanes <= 256 else 2
width = inplanes * self.k
outplanes = inplanes if downsample is None else inplanes * 2
first_dilation = first_dilation or dilation
self.conv1 = ConvBN(
nn.Conv2d(inplanes, width*2, kernel_size=1, stride=1,
dilation=first_dilation, groups=1, bias=False),
norm_layer(width*2))
# self.conv2 = nn.Conv2d(1, self.k, kernel_size=3, stride=stride,
# padding=1, dilation=first_dilation, groups=1, bias=False)
self.conv2 = nn.Conv2d(width, width, kernel_size=3, stride=stride,
padding=1, dilation=first_dilation, groups=width, bias=False)
self.bn2 = norm_layer(width)
self.conv3 = ConvBN(
nn.Conv2d(width, outplanes, kernel_size=1, groups=1, bias=False),
norm_layer(outplanes))
self.skip = ConvBN(
nn.Conv2d(inplanes, outplanes, kernel_size=1, stride=stride,
dilation=first_dilation, groups=1, bias=False),
norm_layer(outplanes)) if downsample is not None else nn.Identity()
self.act3 = hardball(radius2=outplanes) # if downsample is not None else None
def zero_init_last(self):
if getattr(self.conv3.bn, 'weight', None) is not None:
nn.init.zeros_(self.conv3.bn.weight)
def conv_forward(self, x):
conv = self.conv2
C = x.size(1) // self.k
kernel = conv.weight.repeat(C, 1, 1, 1)
bias = conv.bias.repeat(C) if conv.bias is not None else None
return F.conv2d(x, kernel, bias, conv.stride,
conv.padding, conv.dilation, x.size(1))
def forward(self, x):
x0 = self.skip(x)
x = self.conv1(x)
x = x[:, ::2, :, :] * x[:, 1::2, :, :]
x = self.conv2(x)
x = self.bn2(x)
x = self.conv3(x)
x += x0
if self.act3 is not None:
x = self.act3(x)
return x
def make_blocks(
block_fn,
channels,
block_repeats,
inplanes,
reduce_first=1,
output_stride=32,
down_kernel_size=1,
avg_down=False,
**kwargs,
):
stages = []
feature_info = []
net_num_blocks = sum(block_repeats)
net_block_idx = 0
net_stride = 4
dilation = prev_dilation = 1
for stage_idx, (planes, num_blocks) in enumerate(zip(channels, block_repeats)):
stage_name = f'layer{stage_idx + 1}' # never liked this name, but weight compat requires it
stride = 1 if stage_idx == 0 else 2
if net_stride >= output_stride:
dilation *= stride
stride = 1
else:
net_stride *= stride
downsample = None
if stride != 1 or inplanes != planes * block_fn.expansion:
downsample = True
block_kwargs = dict(reduce_first=reduce_first, dilation=dilation, **kwargs)
blocks = []
for block_idx in range(num_blocks):
downsample = downsample if block_idx == 0 else None
stride = stride if block_idx == 0 else 1
blocks.append(block_fn(
inplanes, planes, stride, downsample, first_dilation=prev_dilation,
**block_kwargs))
prev_dilation = dilation
inplanes = planes * block_fn.expansion
net_block_idx += 1
stages.append((stage_name, nn.Sequential(*blocks)))
feature_info.append(dict(num_chs=inplanes, reduction=net_stride, module=stage_name))
return stages, feature_info
class QLNet(nn.Module):
# based on timm code for ResNet / ResNeXt / SE-ResNeXt / SE-Net
def __init__(
self,
block=QLBlock, # new block
layers=[3,4,12,3], # [3,4,6,3] as in resnet50
num_classes=1000,
in_chans=3,
output_stride=32,
global_pool='avg',
cardinality=1,
base_width=64,
stem_width=32,
stem_type='', # 'deep' for resnet50d
replace_stem_pool=False,
block_reduce_first=1,
down_kernel_size=1,
avg_down=False,
act_layer=nn.ReLU,
norm_layer=nn.BatchNorm2d,
zero_init_last=True,
block_args=None,
):
"""
Args:
block (nn.Module): class for the residual block. Options are BasicBlock, Bottleneck.
layers (List[int]) : number of layers in each block
num_classes (int): number of classification classes (default 1000)
in_chans (int): number of input (color) channels. (default 3)
output_stride (int): output stride of the network, 32, 16, or 8. (default 32)
global_pool (str): Global pooling type. One of 'avg', 'max', 'avgmax', 'catavgmax' (default 'avg')
cardinality (int): number of convolution groups for 3x3 conv in Bottleneck. (default 1)
base_width (int): bottleneck channels factor. `planes * base_width / 64 * cardinality` (default 64)
stem_width (int): number of channels in stem convolutions (default 64)
stem_type (str): The type of stem (default ''):
* '', default - a single 7x7 conv with a width of stem_width
* 'deep' - three 3x3 convolution layers of widths stem_width, stem_width, stem_width * 2
* 'deep_tiered' - three 3x3 conv layers of widths stem_width//4 * 3, stem_width, stem_width * 2
replace_stem_pool (bool): replace stem max-pooling layer with a 3x3 stride-2 convolution
block_reduce_first (int): Reduction factor for first convolution output width of residual blocks,
1 for all archs except senets, where 2 (default 1)
down_kernel_size (int): kernel size of residual block downsample path,
1x1 for most, 3x3 for senets (default: 1)
avg_down (bool): use avg pooling for projection skip connection between stages/downsample (default False)
act_layer (str, nn.Module): activation layer
norm_layer (str, nn.Module): normalization layer
zero_init_last (bool): zero-init the last weight in residual path (usually last BN affine weight)
block_args (dict): Extra kwargs to pass through to block module
"""
super(QLNet, self).__init__()
block_args = block_args or dict()
assert output_stride in (8, 16, 32)
self.num_classes = num_classes
self.grad_checkpointing = False
act_layer = get_act_layer(act_layer)
norm_layer = get_norm_layer(norm_layer)
# Stem
deep_stem = 'deep' in stem_type
inplanes = stem_width * 2 if deep_stem else 64
if deep_stem:
stem_chs = (stem_width, stem_width)
if 'tiered' in stem_type:
stem_chs = (3 * (stem_width // 4), stem_width)
self.conv1 = nn.Sequential(*[
nn.Conv2d(in_chans, stem_chs[0], 3, stride=2, padding=1, bias=False),
norm_layer(stem_chs[0]),
act_layer(inplace=True),
nn.Conv2d(stem_chs[0], stem_chs[1], 3, stride=1, padding=1, bias=False),
norm_layer(stem_chs[1]),
act_layer(inplace=True),
nn.Conv2d(stem_chs[1], inplanes, 3, stride=1, padding=1, bias=False)])
else:
self.conv1 = nn.Conv2d(in_chans, inplanes, kernel_size=7, stride=2, padding=3, bias=False)
self.bn1 = norm_layer(inplanes)
# self.act1 = act_layer(inplace=True)
self.feature_info = [dict(num_chs=inplanes, reduction=2, module='act1')]
# Stem pooling. The name 'maxpool' remains for weight compatibility.
if replace_stem_pool:
self.maxpool = nn.Sequential(*filter(None, [
nn.Conv2d(inplanes, inplanes, 3, stride=2, padding=1, bias=False),
norm_layer(inplanes),
act_layer(inplace=True)
]))
else:
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
# Feature Blocks
channels = [64, 128, 256, 512]
stage_modules, stage_feature_info = make_blocks(
block,
channels,
layers,
inplanes,
cardinality=cardinality,
base_width=base_width,
output_stride=output_stride,
reduce_first=block_reduce_first,
avg_down=avg_down,
down_kernel_size=down_kernel_size,
act_layer=act_layer,
norm_layer=norm_layer,
**block_args,
)
for stage in stage_modules:
self.add_module(*stage) # layer1, layer2, etc
self.feature_info.extend(stage_feature_info)
# self.act = hardball(radius2=512)
# self.act = nn.Hardtanh(max_val=5, min_val=-5, inplace=True)
# self.act = nn.ReLU(inplace=True)
# Head (Pooling and Classifier)
self.num_features = 512 * block.expansion
self.global_pool, self.fc = create_classifier(self.num_features, self.num_classes, pool_type=global_pool)
self.init_weights(zero_init_last=zero_init_last)
@staticmethod
def from_pretrained(model_name: str, load_weights=True, **kwargs) -> 'ResNet':
entry_fn = model_entrypoint(model_name, 'resnet')
return entry_fn(pretrained=not load_weights, **kwargs)
@torch.jit.ignore
def init_weights(self, zero_init_last=True):
for n, m in self.named_modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='linear') # 'linear' for non-relu activations
# nn.init.xavier_normal_(m.weight)
if zero_init_last:
for m in self.modules():
if hasattr(m, 'zero_init_last'):
m.zero_init_last()
@torch.jit.ignore
def group_matcher(self, coarse=False):
matcher = dict(stem=r'^conv1|bn1|maxpool', blocks=r'^layer(\d+)' if coarse else r'^layer(\d+)\.(\d+)')
return matcher
@torch.jit.ignore
def set_grad_checkpointing(self, enable=True):
self.grad_checkpointing = enable
@torch.jit.ignore
def get_classifier(self, name_only=False):
return 'fc' if name_only else self.fc
def reset_classifier(self, num_classes, global_pool='avg'):
self.num_classes = num_classes
self.global_pool, self.fc = create_classifier(self.num_features, 99, # self.num_classes,
pool_type=global_pool)
def forward_features(self, x):
x = self.conv1(x)
x = self.bn1(x)
# x = self.act1(x)
x = self.maxpool(x)
if self.grad_checkpointing and not torch.jit.is_scripting():
x = checkpoint_seq([self.layer1, self.layer2, self.layer3, self.layer4], x, flatten=True)
else:
x = self.layer1(x)
x = self.layer2(x)
x = self.layer3(x)
x = self.layer4(x)
return x
def forward_head(self, x, pre_logits: bool = False):
x = self.global_pool(x)
return x if pre_logits else self.fc(x)
def forward(self, x):
x = self.forward_features(x)
# x = self.act(x)
x = self.forward_head(x)
return x
def _create_qlnet(variant, pretrained=False, **kwargs):
return build_model_with_cfg(QLNet, variant, pretrained, **kwargs)
@register_model
def qlnet22(pretrained=False, **kwargs):
"""Constructs a QLNet22 model.
"""
model_args = dict(block=QLBlock, layers=[3, 4, 12, 3], **kwargs)
return _create_qlnet('qlnet22', pretrained, **dict(model_args, **kwargs))
@register_model
def qlnet26(pretrained=False, **kwargs):
"""Constructs a QLNet26 model.
"""
model_args = dict(block=QLBlock, layers=[3, 8, 12, 3], **kwargs)
return _create_qlnet('qlnet26', pretrained, **dict(model_args, **kwargs))
|