Suggestions: Efficient Partial Convolution about pytorch-inpainting-with-partial-conv

[naoto0804]

@yu45020
Thank you so much for your all the detailed suggestions!
Especially, I was surprised by the speed of masked_fill_.
I'll also try cyclical learning rate and in-place batchnorm later.

For your personal project:
Actually, I have come up with a bit similar idea because the research group that I belong to is also familiar with such kind of topics (e.g., the large-scale comic dataset).
But you implemented it in a much more sophisticated way! I'm pleased to see such a great project!

[jaredstorm]

@yu45020, mind if I use this Partial Convolution implementation as well? I'm hoping to extend NVIDIA's research into a tool for conceptual design.

[yu45020]

@hiltonjp Feel free to use it. I also have a slightly faster implementation: in the encoder part, if we are sure all the holes have the same positions across channels, then we can update the mask by computing only one channel and expand it.

[jaredstorm]

@yu45020 Wonderful, thank you!