MCPcopy Create free account
hub / github.com/HobbitLong/PyContrast / ResNet

Class ResNet

pycontrast/networks/resnet.py:129–223  ·  view source on GitHub ↗

Source from the content-addressed store, hash-verified

127
128
129class ResNet(nn.Module):
130
131 def __init__(self, block, layers, width=1, in_channel=3, zero_init_residual=False,
132 groups=1, width_per_group=64, replace_stride_with_dilation=None,
133 norm_layer=None):
134 super(ResNet, self).__init__()
135 if norm_layer is None:
136 norm_layer = nn.BatchNorm2d
137 self._norm_layer = norm_layer
138
139 self.inplanes = max(int(64 * width), 64)
140 self.base = int(64 * width)
141 self.dilation = 1
142 if replace_stride_with_dilation is None:
143 # each element in the tuple indicates if we should replace
144 # the 2x2 stride with a dilated convolution instead
145 replace_stride_with_dilation = [False, False, False]
146 if len(replace_stride_with_dilation) != 3:
147 raise ValueError("replace_stride_with_dilation should be None "
148 "or a 3-element tuple, got {}".format(replace_stride_with_dilation))
149 self.groups = groups
150 self.base_width = width_per_group
151 self.conv1 = nn.Conv2d(in_channel, self.inplanes, kernel_size=7, stride=2, padding=3,
152 bias=False)
153 self.bn1 = norm_layer(self.inplanes)
154 self.relu = nn.ReLU(inplace=True)
155 self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
156 self.layer1 = self._make_layer(block, self.base, layers[0])
157 self.layer2 = self._make_layer(block, self.base * 2, layers[1], stride=2,
158 dilate=replace_stride_with_dilation[0])
159 self.layer3 = self._make_layer(block, self.base * 4, layers[2], stride=2,
160 dilate=replace_stride_with_dilation[1])
161 self.layer4 = self._make_layer(block, self.base * 8, layers[3], stride=2,
162 dilate=replace_stride_with_dilation[2])
163 self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
164
165 # comment out fc layer for unsupervised learning, will have another head
166 # self.fc = nn.Linear(512 * block.expansion, num_classes)
167
168 for m in self.modules():
169 if isinstance(m, nn.Conv2d):
170 nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
171 elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
172 nn.init.constant_(m.weight, 1)
173 nn.init.constant_(m.bias, 0)
174
175 # Zero-initialize the last BN in each residual branch,
176 # so that the residual branch starts with zeros, and each residual block behaves like an identity.
177 # This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677
178 if zero_init_residual:
179 for m in self.modules():
180 if isinstance(m, Bottleneck):
181 nn.init.constant_(m.bn3.weight, 0)
182 elif isinstance(m, BasicBlock):
183 nn.init.constant_(m.bn2.weight, 0)
184
185 def _make_layer(self, block, planes, blocks, stride=1, dilate=False):
186 norm_layer = self._norm_layer

Callers 1

_resnetFunction · 0.70

Calls

no outgoing calls

Tested by

no test coverage detected