Spaces:
Build error
Build error
File size: 12,056 Bytes
30a0ec5 |
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 |
"""
****************** COPYRIGHT AND CONFIDENTIALITY INFORMATION ******************
Copyright (c) 2018 [Thomson Licensing]
All Rights Reserved
This program contains proprietary information which is a trade secret/business \
secret of [Thomson Licensing] and is protected, even if unpublished, under \
applicable Copyright laws (including French droit d'auteur) and/or may be \
subject to one or more patent(s).
Recipient is to retain this program in confidence and is not permitted to use \
or make copies thereof other than as permitted in a written agreement with \
[Thomson Licensing] unless otherwise expressly allowed by applicable laws or \
by [Thomson Licensing] under express agreement.
Thomson Licensing is a company of the group TECHNICOLOR
*******************************************************************************
This scripts permits one to reproduce training and experiments of:
Engilberge, M., Chevallier, L., Pérez, P., & Cord, M. (2018, April).
Finding beans in burgers: Deep semantic-visual embedding with localization.
In Proceedings of CVPR (pp. 3984-3993)
Author: Martin Engilberge
"""
import torch
import torch.nn as nn
import torchvision.models as models
##########################################################
# translated from torch version: #
# https://github.com/durandtibo/weldon.resnet.pytorch #
##########################################################
"""
****************** COPYRIGHT AND CONFIDENTIALITY INFORMATION ******************
Copyright (c) 2018 [Thomson Licensing]
All Rights Reserved
This program contains proprietary information which is a trade secret/business \
secret of [Thomson Licensing] and is protected, even if unpublished, under \
applicable Copyright laws (including French droit d'auteur) and/or may be \
subject to one or more patent(s).
Recipient is to retain this program in confidence and is not permitted to use \
or make copies thereof other than as permitted in a written agreement with \
[Thomson Licensing] unless otherwise expressly allowed by applicable laws or \
by [Thomson Licensing] under express agreement.
Thomson Licensing is a company of the group TECHNICOLOR
*******************************************************************************
This scripts permits one to reproduce training and experiments of:
Engilberge, M., Chevallier, L., Pérez, P., & Cord, M. (2018, April).
Finding beans in burgers: Deep semantic-visual embedding with localization.
In Proceedings of CVPR (pp. 3984-3993)
Author: Martin Engilberge
"""
import torch
import torch.nn as nn
import torchvision.models as models
##########################################################
# translated from torch version: #
# https://github.com/durandtibo/weldon.resnet.pytorch #
##########################################################
class WeldonPooling(nn.Module): #
# Pytorch implementation of WELDON pooling
def __init__(self, nMax=1, nMin=None):
super(WeldonPooling, self).__init__()
self.nMax = nMax
if(nMin is None):
self.nMin = nMax
else:
self.nMin = nMin
self.input = torch.Tensor()
self.output = torch.Tensor()
self.indicesMax = torch.Tensor()
self.indicesMin = torch.Tensor()
def forward(self, input):
self.batchSize = 0
self.numChannels = 0
self.h = 0
self.w = 0
if input.dim() == 4:
self.batchSize = input.size(0)
self.numChannels = input.size(1)
self.h = input.size(2)
self.w = input.size(3)
elif input.dim() == 3:
self.batchSize = 1
self.numChannels = input.size(0)
self.h = input.size(1)
self.w = input.size(2)
else:
print('error in WeldonPooling:forward - incorrect input size')
self.input = input
nMax = self.nMax
if nMax <= 0:
nMax = 0
elif nMax < 1:
nMax = torch.clamp(torch.floor(nMax * self.h * self.w), min=1)
nMin = self.nMin
if nMin <= 0:
nMin = 0
elif nMin < 1:
nMin = torch.clamp(torch.floor(nMin * self.h * self.w), min=1)
x = input.view(self.batchSize, self.numChannels, self.h * self.w)
# sort scores by decreasing order
scoreSorted, indices = torch.sort(x, x.dim() - 1, True)
# compute top max
self.indicesMax = indices[:, :, 0:nMax]
self.output = torch.sum(scoreSorted[:, :, 0:nMax], dim=2, keepdim=True)
self.output = self.output.div(nMax)
# compute top min
if nMin > 0:
self.indicesMin = indices[
:, :, self.h * self.w - nMin:self.h * self.w]
yMin = torch.sum(
scoreSorted[:, :, self.h * self.w - nMin:self.h * self.w], 2, keepdim=True).div(nMin)
self.output = torch.add(self.output, yMin)
if input.dim() == 4:
self.output = self.output.view(
self.batchSize, self.numChannels, 1, 1)
elif input.dim() == 3:
self.output = self.output.view(self.numChannels, 1, 1)
return self.output
def backward(self, grad_output, _indices_grad=None):
nMax = self.nMax
if nMax <= 0:
nMax = 0
elif nMax < 1:
nMax = torch.clamp(torch.floor(nMax * self.h * self.w), min=1)
nMin = self.nMin
if nMin <= 0:
nMin = 0
elif nMin < 1:
nMin = torch.clamp(torch.floor(nMin * self.h * self.w), min=1)
yMax = grad_output.clone().view(self.batchSize, self.numChannels,
1).expand(self.batchSize, self.numChannels, nMax)
z = torch.zeros(self.batchSize, self.numChannels,
self.h * self.w).type_as(self.input)
z = z.scatter_(2, self.indicesMax, yMax).div(nMax)
if nMin > 0:
yMin = grad_output.clone().view(self.batchSize, self.numChannels, 1).div(
nMin).expand(self.batchSize, self.numChannels, nMin)
self.gradInput = z.scatter_(2, self.indicesMin, yMin).view(
self.batchSize, self.numChannels, self.h, self.w)
else:
self.gradInput = z.view(
self.batchSize, self.numChannels, self.h, self.w)
if self.input.dim() == 3:
self.gradInput = self.gradInput.view(
self.numChannels, self.h, self.w)
return self.gradInput
class ResNet_weldon(nn.Module):
def __init__(self, args, pretrained=True, weldon_pretrained_path=None):
super(ResNet_weldon, self).__init__()
resnet = models.resnet152(pretrained=pretrained)
self.base_layer = nn.Sequential(*list(resnet.children())[:-2])
self.spaConv = nn.Conv2d(2048, 2400, 1,)
# add spatial aggregation layer
self.wldPool = WeldonPooling(15)
# Linear layer for imagenet classification
self.fc = nn.Linear(2400, 1000)
# Loading pretrained weights of resnet weldon on imagenet classification
if pretrained:
try:
state_di = torch.load(
weldon_pretrained_path, map_location=lambda storage, loc: storage)['state_dict']
self.load_state_dict(state_di)
except Exception:
print("Error when loading pretrained resnet weldon")
def forward(self, x):
x = self.base_layer(x)
x = self.spaConv(x)
x = self.wldPool(x)
x = x.view(x.size(0), -1)
x = self.fc(x)
return x
class DynamicPooling(nn.Module): #
# Pytorch implementation of WELDON pooling
def __init__(self, nMax=1, nMin=None):
super(DynamicPooling, self).__init__()
self.nMax = nMax
if(nMin is None):
self.nMin = nMax
else:
self.nMin = nMin
self.input = torch.Tensor()
self.output = torch.Tensor()
self.indicesMax = torch.Tensor()
self.indicesMin = torch.Tensor()
self.conv2d = nn.Conv2d(in_channels=2400, out_channels=2400, kernel_size=3, groups=2400)
self.avgpool = nn.AdaptiveAvgPool2d(output_size=1)
self.act = nn.ReLU()
def fore_back_layer(self, x):
x_fore = self.conv2d(x)
x_back = self.conv2d(x)
x_fore = self.avgpool(x_fore)
x_back = self.avgpool(x_back)
x_fore = self.act(x_fore)
x_back = self.act(x_back)
return x_fore, x_back
def forward(self, input):
self.batchSize = 0
self.numChannels = 0
self.h = 0
self.w = 0
if input.dim() == 4:
self.batchSize = input.size(0)
self.numChannels = input.size(1)
self.h = input.size(2)
self.w = input.size(3)
elif input.dim() == 3:
self.batchSize = 1
self.numChannels = input.size(0)
self.h = input.size(1)
self.w = input.size(2)
else:
print('error in WeldonPooling:forward - incorrect input size')
self.input = input
nMax = self.nMax
if nMax <= 0:
nMax = 0
elif nMax < 1:
nMax = torch.clamp(torch.floor(nMax * self.h * self.w), min=1)
nMin = self.nMin
if nMin <= 0:
nMin = 0
elif nMin < 1:
nMin = torch.clamp(torch.floor(nMin * self.h * self.w), min=1)
# calculate the foreground coefficient
weight_fore, weight_back = self.fore_back_layer(input)
x = input.view(self.batchSize, self.numChannels, self.h * self.w)
# sort scores by decreasing order
scoreSorted, indices = torch.sort(x, x.dim() - 1, True)
# compute top max
self.indicesMax = indices[:, :, 0:nMax] # torch.Size([40, 2400, 15])
self.output = weight_fore.squeeze(dim=-1) * torch.sum(scoreSorted[:, :, 0:nMax], dim=2, keepdim=True)
self.output = self.output.div(nMax)
# compute top min
if nMin > 0:
self.indicesMin = indices[
:, :, self.h * self.w - nMin:self.h * self.w]
yMin = weight_back.squeeze(dim=-1) * torch.sum(
scoreSorted[:, :, self.h * self.w - nMin:self.h * self.w], 2, keepdim=True).div(nMin)
self.output = torch.add(self.output, yMin)
if input.dim() == 4:
self.output = self.output.view(
self.batchSize, self.numChannels, 1, 1)
elif input.dim() == 3:
self.output = self.output.view(self.numChannels, 1, 1)
return self.output
def backward(self, grad_output, _indices_grad=None):
nMax = self.nMax
if nMax <= 0:
nMax = 0
elif nMax < 1:
nMax = torch.clamp(torch.floor(nMax * self.h * self.w), min=1)
nMin = self.nMin
if nMin <= 0:
nMin = 0
elif nMin < 1:
nMin = torch.clamp(torch.floor(nMin * self.h * self.w), min=1)
yMax = grad_output.clone().view(self.batchSize, self.numChannels,
1).expand(self.batchSize, self.numChannels, nMax)
z = torch.zeros(self.batchSize, self.numChannels,
self.h * self.w).type_as(self.input)
z = z.scatter_(2, self.indicesMax, yMax).div(nMax)
if nMin > 0:
yMin = grad_output.clone().view(self.batchSize, self.numChannels, 1).div(
nMin).expand(self.batchSize, self.numChannels, nMin)
self.gradInput = z.scatter_(2, self.indicesMin, yMin).view(
self.batchSize, self.numChannels, self.h, self.w)
else:
self.gradInput = z.view(
self.batchSize, self.numChannels, self.h, self.w)
if self.input.dim() == 3:
self.gradInput = self.gradInput.view(
self.numChannels, self.h, self.w)
return self.gradInput
|