ContextPrior
ContextPrior copied to clipboard
How to downsample GT?
softmax function needs LongTensor for the GT, if you downsample GT with pool or Interpolation, how to get LongTensor, then how to guarantee the downsampled GT is mainingful, which means the value of per pixel represents the real class in the corresponding downsampled images.In my opinion, downsampling GT makes no sense. Hope to hear your understanding
I use nearest neighbor to downsample GT. This workaround can successfully back propagate. However, I don't think that single left-top pixel can represent 8x8 patch in original size GT.