Checkerboard artifacts in reconstructed image

[parthnatekar]

Hi, I'm training an mselpips model on a custom dataset.

I'm getting checkerboard artifacts in my reconstructed image, shown below. This is a part of the background and should hence be completely zero

Why does this happen? Do I need to change my hyperparameters?

Thanks

[fab-jul]

Hi, not sure what's happening, but note that in the paper, we report that without a GAN loss (which I think is what the mselpips model is), we also see gridding. See FigA2 in the appendix here: https://arxiv.org/pdf/2006.09965.pdf

[parthnatekar]

Hi, thanks for the reply.

A follow up on this - in archs.py, the final layer activation for the decoder is 'linear', which is what Keras defaults to. However, the MSE and LPIPS losses are designed for 0-1 float outputs (or 0-255 ints). Is this intended?