AutoNN.CNN.cnnBlocks

cnnBlocks.SkipLayer

This is basically a residual block, which is the core component of AutoNN's CNN architecture. Bunch of residual blocks are stacked together along with Pooling layers to give the full CNN architecture.

Source Code

class SkipLayer(nn.Module):
    def __init__(self,in_channels,featureMaps1,featureMaps2,
        kernel=(3,3),stride=(1,1),padding=1):
        super(SkipLayer,self).__init__()

        self.skiplayers=nn.Sequential(
            nn.Conv2d(in_channels,featureMaps1,kernel,stride,padding=padding),
            nn.BatchNorm2d(featureMaps1),
            nn.ReLU(),

            nn.Conv2d(featureMaps1,featureMaps2,kernel,stride,padding=padding),
            nn.BatchNorm2d(featureMaps2)
            )
        self.skip_connection = nn.Conv2d(in_channels,featureMaps2,kernel_size=(1,1),
                                        stride=stride)
        self.relu=nn.ReLU()

    def forward(self,x):
        x0 = x.clone()
        x = self.skiplayers(x)
        x0 = self.skip_connection(x0)
        x+=x0
        return self.relu(x)

cnnBlocks.Pooling

Pooling layers consists of both MaxPool and AvgPool, which can be selected at the time of creating an instance of the Pooling class by providing the pool_type.

Source Code

class Pooling(nn.Module):
    def __init__(self,pool_type='maxpool'):
        super(Pooling,self).__init__()

        """
        Args:
            pool_type: 'maxpool' | nn.MaxPool2d
                       'avgpool' | nn.AvgPool2d
        """
        if pool_type.lower() =='maxpool':
            self.pool = nn.MaxPool2d(kernel_size=(2,2),stride=(2,2))
        else:
            self.pool = nn.AvgPool2d(kernel_size=(2,2),stride=(2,2))

    def forward(self,x):
        return self.pool(x)