Convolution pad "same" option

1

Hi,

In theano or tensorflow, they have 2 different option for convolution operation : ‘valid’, ‘same’. I want to produce same size convolution output when I am using even number filter.

However, with current convolution layer, there is no way I can produce same output dimension if the kernel size is even number.

For example : [ 1 2 3 ] convolved with kernel [a b], if I use no pad, I will get 2 output. If I use pad 1, I will get 4 output. I can’t get 3 output.

Is there any solution for this ?

Thanks :slight_smile:

2

with the current convolution operator there is no way to do this.

3

Note that you can add an extra F.pad before the convolution to adjust the size to make it have the same size.

4

I see. For now I will use it. Thanks

IMHO it would be nice if convolution operator give those choices so we can avoid if-else for even number kernel size.