Resnet50 pytorch复现
Resnet50 pytorch复现
之前复现过Resnet18，今天复现与Resnet网络结构稍有不同的Resnet50
Resnet50的基本结构是1x1卷积->3x3卷积->1x1卷积。而每一组卷积是这样的结构：卷积->BN->RELU组合而成。
如果所示，上面左边的为Resnet18,34的残差结构，右边的是Resnet50 101 152的残差结构。从图中可以看出Resnet50 的 1x1的卷积->3x3卷积->1x1卷积结构中，第一个1x1的卷积是进行降维操作，再做3x3的卷积提取特征，再做1x1的卷积进行升维。与Unet的网络不同，这里的残差结构是add操作，即两组同样的shape的权值，对应位置的值进行相加，而Unet中concatenate的操作，是对应宽高的权值，进行通道叠加。
以下代码是基本结构的代码
import torch.nn as nn import torch.onnx class Block(nn.Module): def __init__(self, in_channels, channels, stride, downsample=None): super(Block, self).__init__() # 1x1的卷积降维操作 self.conv1 = nn.Conv2d(in_channels=in_channels, out_channels=channels, kernel_size=(1, 1), bias=False) self.bn1 = nn.BatchNorm2d(channels) # 3x3的卷积提取特征操作 self.conv2 = nn.Conv2d(in_channels=channels, out_channels=channels, kernel_size=(3, 3), stride=stride, padding=1, bias=False) self.bn2 = nn.BatchNorm2d(channels) # 1x1的卷积升维操作 self.conv3 = nn.Conv2d(in_channels=channels, out_channels=channels * 4, kernel_size=(1, 1), bias=False) self.bn3 = nn.BatchNorm2d(channels * 4) self.relu = nn.ReLU(inplace=True) self.downsample = downsample # 4组卷积层的头一层网络会做一次降采样 if self.downsample is not None: self.dconv = nn.Conv2d(in_channels, channels * 4, stride=stride, kernel_size=(1, 1), bias=False) self.dbn = nn.BatchNorm2d(channels * 4) def forward(self, x): identity = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) out = self.relu(out) out = self.conv3(out) out = self.bn3(out) if self.downsample is not None: identity = self.dconv(identity) identity = self.dbn(identity) out += identity out = self.relu(out) return out以下是主体结构
class Resnet50(nn.Module): def __init__(self, num_classes): super(Resnet50, self).__init__() self.conv1 = nn.Conv2d(3, 64, kernel_size=(7, 7), stride=2, padding=3, bias=False) self.bn1 = nn.BatchNorm2d(64) self.relu1 = nn.ReLU(inplace=True) self.maxpool1 = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) # 对应第1组网络层，3*Resnet的基本结构 self.conv64_1 = Block(64, 64, stride=1, downsample=True) self.conv64_2 = Block(256, 64, stride=1) self.conv64_3 = Block(256, 64, stride=1) # 对应第2组网络层，4*Resnet的基本结构 self.conv128_1 = Block(256, 128, stride=2, downsample=True) self.conv128_2 = Block(128 * 4, 128, stride=1) self.conv128_3 = Block(128 * 4, 128, stride=1) self.conv128_4 = Block(128 * 4, 128, stride=1) # 对应第3组网络层，6*Resnet的基本结构 self.conv256_1 = Block(512, 256, stride=2, downsample=True) self.conv256_2 = Block(256 * 4, 256, stride=1) self.conv256_3 = Block(256 * 4, 256, stride=1) self.conv256_4 = Block(256 * 4, 256, stride=1) self.conv256_5 = Block(256 * 4, 256, stride=1) self.conv256_6 = Block(256 * 4, 256, stride=1) # 对应第4组网络层，3*Resnet的基本结构 self.conv512_1 = Block(1024, 512, stride=2, downsample=True) self.conv512_2 = Block(512 * 4, 512, stride=1) self.conv512_3 = Block(512 * 4, 512, stride=1) self.avgpool = nn.AdaptiveAvgPool2d((1,1)) self.fc = nn.Linear(2048, num_classes) def forward(self, x): x = self.conv1(x) x = self.bn1(x) x = self.relu1(x) x = self.maxpool1(x) x = self.conv64_1(x) x = self.conv64_2(x) x = self.conv64_3(x) x = self.conv128_1(x) x = self.conv128_2(x) x = self.conv128_3(x) x = self.conv128_4(x) x = self.conv256_1(x) x = self.conv256_2(x) x = self.conv256_3(x) x = self.conv256_4(x) x = self.conv256_5(x) x = self.conv256_6(x) x = self.conv512_1(x) x = self.conv512_2(x) x = self.conv512_3(x) x = self.avgpool(x) x = torch.flatten(x, 1) x = self.fc(x) return x if __name__ == '__main__': model = Resnet50(1000) from torchsummary import summary print(summary(model, (3, 224, 224))) torch.onnx.export(model, torch.randn(1,3,224,224),r"Y:\code\python\202203\Resnet50.onnx")
下图是Resnet50的网络结构图
