Animations of Convolution and Deconvolution#

The concepts convolution, deconvolution (=transposed convolution), strides and padding have been introduced in the previous section. Below, these concepts are demonstrated. The animations are from [DV16]. In the demos only a single channel is at the input and only a single feature map is calculated. In a convolution- or deconvolution-layer typically many feature maps are calculated from many channels at the input.

Convolution#

padding = 0, stride = 1

padding = 1, stride = 1

padding = 0, stride = 2

padding = 1, stride = 2

Deconvolution#

padding = 0, stride = 1, transposed

padding = 1, stride = 1, transposed

padding = 0, stride = 2, transposed

padding = 1, stride = 2, transposed