About PixContrast loss

[ZMHH-H]

Hi, thanks for sharing your code!
i'm confused about PixContrastLoss in your implementation. Since msked_select() will return a 1D tensor and all feature maps have batch dimension. I think your implementation add all positive pairs and negative pairs in one minibatch together, which is not consistent with the paper. Is that correct?

[dev-sungman]

Thanks for comment, I'll check it later.