3_neural_networks_tutorial中nn.Linear的输入特征数量 about pytorch-handbook

[onlyonewater]

它这是大小为1655的池化层连接到全连接啊，全连接的大小为120，所以是（1655,120）吧？

[onlyonewater]

16 * 5 * 5

[gandad]

明白了，其实这里的5*5是图像featuremap的大小，碰巧与kernel size相同

        # Max pooling over a (2, 2) window
        # conv1(x)=>32->(32-5)/1+1=28; 28 downsampling to=>14
        x = F.max_pool2d(F.relu(self.conv1(x)), (2, 2)) 
        # If the size is a square you can only specify a single number
        # conv2(x)=>14->(14-5)/1+1=10; 10 downsampling to=>5
        x = F.max_pool2d(F.relu(self.conv2(x)), 2)
        x = x.view(-1, self.num_flat_features(x))