Possible bug with padding

[georgepar]

Hey there,

I was going through the code and I noticed the following, which I found curious.

In Line 75, you pad the input tensor to a multiple of num_landmarks from the front:

x = F.pad(x, (0, 0, padding, 0), value = 0)

In Line 144 you trim the extra padding elements you inserted in the output tensor from the end.

out = out[:, :n]

Am I not getting something, or should we be removing the front elements of out?

out = out[:, out.size(1) - n:]

[lucidrains]

@georgepar Hi Giorgos! You indeed found a bug, thank you, and I am sorry if this caused any troubles with your research :(

[georgepar]

No trouble at all. Thanks for your awesome work :)