search

richzhang / colorization-pytorch

PyTorch reimplementation of Interactive Deep Colorization
https://richzhang.github.io/ideepcolor/
MIT License
597 stars 113 forks source link

Classification output in SIGGRAPHGenerator, #channel=529 ? #7

Closed azshue closed 5 years ago

azshue commented 5 years ago

Hi,

I'm reading the implementation detail of the SIGGRAPHGenerator, but had some trouble understanding the classification output of this network. Classification output in SIGGRAPHGenerator: https://github.com/richzhang/colorization-pytorch/blob/9fd9bd867bca53c861816298089c978617d7d5f5/models/networks.py#L315

Is 529 the number of quantized color Q? While I suppose this number to be 313 instead... Could anyone please explain this number?

Thanks,

richzhang commented 5 years ago

Not all (a,b) values are in-gamut. In the original Caffe implementation, I took out the 313 which were in-gamut. For the reimplementation, I just kept all 23*23=529 bins. It makes the encoding code a bit more straightforward. In practice, the other 529-313=216 bins are basically unused, so at test time they should output ~0% probability.

azshue commented 5 years ago

Not all (a,b) values are in-gamut. In the original Caffe implementation, I took out the 313 which were in-gamut. For the reimplementation, I just kept all 23*23=529 bins. It makes the encoding code a bit more straightforward. In practice, the other 529-313=216 bins are basically unused, so at test time they should output ~0% probability.

Thank you so much! This also explains the encoding code in PyTorch implementation.