pytroch resnet构建过程理解

?

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

class ResNet(nn.Module):   def __init__(self, block, layers, num_classes=1000):     self.inplanes = 64     super(ResNet, self).__init__()     self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3,                  bias=False)     self.bn1 = nn.BatchNorm2d(64)     self.relu = nn.ReLU(inplace=True)     self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=0, ceil_mode=True) # change   第一次pooling     self.layer1 = self._make_layer(block, 64, layers[0])     self.layer2 = self._make_layer(block, 128, layers[1], stride=2)     self.layer3 = self._make_layer(block, 256, layers[2], stride=2)     self.layer4 = self._make_layer(block, 512, layers[3], stride=2)     # it is slightly better whereas slower to set stride = 1     # self.layer4 = self._make_layer(block, 512, layers[3], stride=1)     self.avgpool = nn.AvgPool2d(7)     self.fc = nn.Linear(512 * block.expansion, num_classes)       for m in self.modules():       if isinstance(m, nn.Conv2d):         n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels         m.weight.data.normal_(0, math.sqrt(2. / n))       elif isinstance(m, nn.BatchNorm2d):         m.weight.data.fill_(1)         m.bias.data.zero_()     def  _make_layer(self, block, planes, blocks, stride=1):     downsample = None     if stride != 1 or self.inplanes != planes * block.expansion:       downsample = nn.Sequential(         nn.Conv2d(self.inplanes, planes * block.expansion,               kernel_size=1, stride=stride, bias=False),         nn.BatchNorm2d(planes * block.expansion),       )       layers = []     layers.append(block(self.inplanes, planes, stride, downsample))     self.inplanes = planes * block.expansion     for i in range(1, blocks):       layers.append(block(self.inplanes, planes))       return nn.Sequential(*layers)     def forward(self, x):     x = self.conv1(x)     x = self.bn1(x)     x = self.relu(x)     x = self.maxpool(x)                       x = self.layer1(x)     x = self.layer2(x)     x = self.layer3(x)     x = self.layer4(x)       x = self.avgpool(x)     x = x.view(x.size(0), -1)     x = self.fc(x)       return x