I’m attempting to give me data more temporal understanding. To do that, I want to pad only the left side with each successive layer so that I’ll maintain the same shape, but progressively “roll up” information from the older / earlier elements. Is this possible with PyTorch?
You could leave padding=0
in the conv layers and use F.pad
in the forward method to pad one-sided.
1 Like
You can also use a larger padding, and then throw away the last couple of outputs to maintain the size.