Increase channels for custom residual blocks

1

I’m trying to insert residual connection in my small 3 conv-layer.

Here’s what it looks like -

class ResClassifier(nn.Module):
    def __init__(self, num_classes):
        super(ResClassifier( self).__init__()
        
        
        self.block1 = self.conv_block(c_in=1, c_out=16, dropout=0.1, kernel_size=5, stride=1, padding=2)
        self.block2 = self.conv_block(c_in=16, c_out=32, dropout=0.1, kernel_size=3, stride=1, padding=1)
        self.block3 = self.conv_block(c_in=32, c_out=64, dropout=0.1, kernel_size=3, stride=1, padding=1)
        
        self.lastcnn = nn.Conv2d(in_channels=64, out_channels=num_classes, kernel_size=28, stride=1, padding=0)
       
        
    def forward(self, x):
             
        x = self.block1(x)
        
        x = self.block2(x)
    
        x = self.block3(x)
        
        x = self.lastcnn(x)
        
        return x
    
    
    def conv_block(self, c_in, c_out, dropout, **kwargs):
        seq_block = nn.Sequential(
            nn.Conv2d(in_channels=c_in, out_channels=c_out, **kwargs),
            nn.BatchNorm2d(num_features=c_out),
            nn.ReLU(),
            nn.Dropout2d(p=dropout)
        )
        
        return seq_block

Let’s say I want to add skip connection from block-1 to block-3. To do that, I can simply do -

   def forward(self, x):
        
        residual_1 = x

        x = self.block1(x)        

        x = self.block2(x)

        x += residual_1

        x = self.block3(x)

        x = self.lastcnn(x)
        
        return x

But the issue is, output of block-3 has 64 channels while the input (residual_1) has only 1.

Because of this mismatch, I can’t add the residual connection. How do I increase the number of channels in residual_1 before I can add it to the output of block-3.

2

You could use a downsample layer as seen in the original ResNet implementation, which could just use a 1x1 conv with the desired number of output channels to match the activation.

1 Like