3D implementation ?

[romainVala]

Hello

very nice woks !
I do also struggle with spatial shift sensitivity of my 3D CNN and I would like to test your solution.
Is someone aware of a 3D implementation of those filter ? or can you guide us on the implementation ?

Many thanks

[richzhang]

I don't currently have an implementation planned. But yes it should be doable! You can follow my implementation. You'd modify this line to be across 3 dimensions, rather than 2

[romainVala]

thanks
I can give a try, but the links you post does not work for me ...

[richzhang]

I did a refactor -- here's the link! https://github.com/adobe/antialiased-cnns/blob/master/antialiased_cnns/blurpool.py#L40

[anirudh-chakravarthy]

Hi,

Thank you for your work!

I am also trying a 3D implementation. However, ReflectionPad3D is not available in PyTorch, while ReplicationPad3d and ConstantPad3D are. Hence, I was wondering if you evaluated the performance with the equivalent 2D layers in your implementation, and what performance did you observe?

This would give me an idea how important this is, and whether I can use another padding layer interchangeably, or focus on first writing my own 3D reflection padding layer. My guess from your comment in this issue is that it shouldn't matter very much?

Thanks for the help!