semantic-pyramid-pytorch
semantic-pyramid-pytorch copied to clipboard
mask in reconstruction loss
I think your reconstruction loss is missing mask, which should be applied after subtracting two features as in (3) of the paper.
https://github.com/rosinality/semantic-pyramid-pytorch/blob/6d4796cbb2abdbe9afaafb8dae75b60a3b619d2d/train.py#L146
Oh, thank you! Fixed it in d3777bf.