why change mask before output computation?

[Joe Mathai]

hi,
Thanks for the implementation. I want to know why you have decided to do this

pytorch-inpainting-with-partial-conv/net.py

Line 85 in 76ba1ee

 mask_sum = output_mask.masked_fill_(no_update_holes, 1.0) 

in your implementation when the original implementation details don't mention it.
http://masc.cs.gmu.edu/wiki/partialconv

[Naoto Inoue]

Thank you for visiting this repository!
This is to assign pseudo value where output_mask == 0. Actually, the value 1.0 can be any value a except 0.0. Without that, zero division error will raise at the pixel where mask_sum == 0 in

pytorch-inpainting-with-partial-conv/net.py

Line 87 in 76ba1ee

 output_pre = (output - output_bias) / mask_sum + output_bias 

After that, we replace a with 0.0 again to keep correct value in

pytorch-inpainting-with-partial-conv/net.py

Line 91 in 76ba1ee

 new_mask = new_mask.masked_fill_(no_update_holes, 0.0) 

[Joe Mathai]

Thanks for the clarification.