Can padding be 0.5?

vision
1

I would like to use a convolution for a input image, but the kernel size is defined by [1,2], and I have to make the output feature map same size with the input image. I’m wondering if I can use a padding as (0,0.5)?

Thanks.

2

No. You can either pad by 1 and later throw away a column or use the separate torch.nn.functional.pad function.

Best regards

Thomas

1 Like
3

Thanks.

What do you mean by using the separate torch.nn.pad function? Can you please provide more details?

4

I fixed the link, sorry. :slight_smile:

5

You can use pad to match the dimension you want.

nn.Conv2d(...padding=1) # H*W
nn.Conv2d(...padding=2) # (H+2)*(W+2)

If you want (H+1)*(W+1) in this case, you can produce H*W and pad by (left=0,right=1,top=0,bottom=1) to get (H+1)*(W+1).

6

@tom @m75

Thanks for your answers.