Why chomp in Temporal Convolutional Network?

Question

In the code that accompanies the paper describing the TCN network, the activations of the temporal convolutions are "chomped", or sliced at the end by the number of zero-padding that was added. This ensures that the size of the input and the hidden layers match. Why is it that only one end of the activations are sliced? Why not have the activations sliced evenly from both ends? Is this what causes the convolution to be causal?

Code: https://github.com/locuslab/TCN/blob/master/TCN/tcn.py

Paper: https://arxiv.org/pdf/1803.01271.pdf

Answer 1

This is because the padding is done in the Convolution layer thus padding left and right. The Chomp module simply removes excess right padding from Convolutional output.

If you left pad only before the convolution, you won't need to chomp.