This question should be quite generic but I faced the problem in the case of the TiF-GAN generator so I am going to use it as an example. (Link to paper)

If you check the penultimate page in the paper you can find the architecture design of the generator.

The generator has a dense layer and then a reshape layer converting the hidden layer feature map to a dimensionality of 8x4x512 (given that $s = 64$)

Then what follows is a transpose convolution operation with a kernel size of 12x3 with $512$ filters and a stride of $2$ in all dimensions. The output of this layer should be then 16x8x512.

After fiddling with some coding I found out that the authors also used the setting padding=same in their tensor flow code.

So, my question is: How and what do you pad when you perform such a transpose convolution to get those output dimensions?

Without any padding I would assume that you should get an output of 26x9x1534 assuming that each output dimension is equal to dim = kernel_dim + strides * (input_dim - 1)



You must log in to answer this question.