You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
This repository has been archived by the owner on Jun 8, 2023. It is now read-only.
I understand how this can be confusing: the convention we are using is to refer to the parameters of the direct convolution as i, k, s and p, and those of the transposed convolution (i.e., the one that we would apply to the output of the transposed convolution if we wanted to obtain a matrix of the same shape as the one the direct convolution was applied on) as i', k', s' and p'.
In other words, what Figure 4.2's captions is saying is that if we had a direct convolution from a matrix A to a matrix B, with matrix A being 5x5, using a 4x4 kernel, unit stride and 2x2 padding (i.e., i=5, k=4, s= 1 and p=2) and we applied a transposed convolution to B to obtain a matrix C of the same shape as A, the transposed convolution would have parameters i'=6, k'=4, s'= 1 and p'=1.
I have been trying to implement it in JAX, but being unable to match the dimensions with the formula is really putting me off. Should I apply the a-trous formula there or something else is wrong there? Some help would be really appreciated as I am stuck due to this.
Unfortunately we don't have the resources to provide help with your code nor to provide guidance on how to implement this, but please let me know if there is something that is still unclear about the Figure or the guide.
Sign up for freeto subscribe to this conversation on GitHub.
Already have an account?
Sign in.
I am attaching the two figures taken from the arXiV. In convolution:
it says the padding is 2x2
and in transposed, it also says the padding is 2x2. Shouldn't it be 1x1?
The text was updated successfully, but these errors were encountered: