CNN 卷积层输出计算

$$output\_size=\lceil \frac{input\_size+2\times padding-kernel\_size}{stride}\rceil +1$$

class Classifier(nn.Module):
    def __init__(self):
        super(Classifier, self).__init__()
        # The arguments for commonly used modules:
        # torch.nn.Conv2d(in_channels, out_channels, kernel_size, stride=1, padding=0)
        # torch.nn.MaxPool2d(kernel_size, stride=kernel_size, padding=0)

        # input: [batch_size, 3, 128, 128]
        # Conv2d/MaxPool2d compute: [(input_size + 2*padding - kernel_size)/stride]+1  []中内容向上取整
        self.cnn_layers = nn.Sequential(
            nn.Conv2d(3, 64, kernel_size=3, stride=1, padding=1), # (128+2-3)/1+1=128
            nn.BatchNorm2d(64),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=2, stride=2, padding=0), # (128+0-2)/2+1=64

            nn.Conv2d(64, 128, kernel_size=3, stride=1, padding=1), # (64+2-3)/1+1=64
            nn.BatchNorm2d(128),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=2, stride=2, padding=0), # (64+0-2)/2+1=32

            nn.Conv2d(128, 256, kernel_size=3, stride=1, padding=1), # (32+2-3)/1+1=32
            nn.BatchNorm2d(256),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=4, stride=4, padding=0), # (32+0-4)/4+1=8
        )
        self.fc_layers = nn.Sequential(
            nn.Linear(256 * 8 * 8, 256),
            nn.ReLU(),
            nn.Linear(256, 256),
            nn.ReLU(),
            nn.Linear(256, 11)
        )

    def forward(self, x):
        # input (x): [batch_size, 3, 128, 128]
        # output: [batch_size, 11]

        # Extract features by convolutional layers.
        x = self.cnn_layers(x)

        # The extracted feature map must be flatten before going to fully-connected layers.
        x = x.flatten(1)

        # The features are transformed by fully-connected layers to obtain the final logits.
        x = self.fc_layers(x)
        return x

打印出每层维度信息

net = Classifier().cnn_layers
X = torch.rand(size=(1, 3, 128, 128), dtype=torch.float32)
for layer in net:
  X = layer(X)
  print(layer.__class__.__name__, 'output shape: \t', X.shape)

-----------------------------------------------------------------
Conv2d output shape: 	 torch.Size([1, 64, 128, 128])
BatchNorm2d output shape: 	 torch.Size([1, 64, 128, 128])
ReLU output shape: 	 torch.Size([1, 64, 128, 128])
MaxPool2d output shape: 	 torch.Size([1, 64, 64, 64])
Conv2d output shape: 	 torch.Size([1, 128, 64, 64])
BatchNorm2d output shape: 	 torch.Size([1, 128, 64, 64])
ReLU output shape: 	 torch.Size([1, 128, 64, 64])
MaxPool2d output shape: 	 torch.Size([1, 128, 32, 32])
Conv2d output shape: 	 torch.Size([1, 256, 32, 32])
BatchNorm2d output shape: 	 torch.Size([1, 256, 32, 32])
ReLU output shape: 	 torch.Size([1, 256, 32, 32])
MaxPool2d output shape: 	 torch.Size([1, 256, 8, 8])

Conclusion

1. $3\times 3$的卷积核(stride=1，padding=1)不改变图片大小
2. flatten()函数作用