Padding strategy for nn.ConvTranspose2d

vision
1

In tensorflow, there are padding strategy like ‘SAME’, ‘VALID’.
However, in pytorch, we should specify the padding for each dimension. Does anyone know how to to different padding strategy for nn.ConvTranspose2d

2

The output shape of the ConvTranspose2d layer is explained in the docs.
You would have to calculate the appropriate padding for your use case.

2 Likes
3

hi @ptrblck,

here’s what I want to do:
I have a tensor with shape (128, 3, 32, 32) and I want to apply deconv on it and the output size to be (128, 64, 32, 32)
with kernel size=4, stride=1
If I calculate the padding size with the formula here
I will get 2*padding = 3

then how should I define the padding number?

4

Are you sure you need a ConvTranspose layer for this?
It seems, that your width and height should not change in the output.
You could just use a Conv2d layer with e.g. kernel_size=3 and padding=1.

EDIT: Using kernel_size=4 is a bit tricky. Since the filter size is even, you will end up with an output size of 31 or 33.

2 Likes