Is it possible to do conv_transpose2d for more than one channel with the same kernel?
fold the channels into the mini-batch dimension, perform conv_transpose2d, and unfold the channels back.
For example:
Input: [N, C, H, W]
Make it: [N * C, 1, H, W]
do conv_transpose2d
Then make it: [N, C, OH, OW]
Thanks a lot