It turns out that the output shape of a transpose convolution is ambiguous. There are at least two ways it could work out. Here’s a thread on which mentor Raymond explains how this arises. You need to read most of the thread to get to Raymond’s post.
If I’m missing your point, please let me know. 