About the transposed convolution

[ilan-theodoro]

Hi everyone! This project is awesome, thank you all for the initiative!

I was looking at your VQVAE implementation, and I noticed that you adopt transposed convolution as your upsample method.

https://github.com/Project-MONAI/GenerativeModels/blob/1b34e60f80555e02036915224dd1783409095d3c/generative/networks/nets/vqvae.py#L257C21-L257C40

As far as I know, transposed convolutions have the undesired effect of possibly introducing the checkerboard effect. It is still used in the original VQVAE implementation, but the authors of VQGAN chose to adopt interpolation as their upsampling method.

Did you only implement the original VQVAE or did you evaluate and choose to use transposed convolution? If the former, a flag to select simple interpolation would be helpful :)

[marksgraham]

I recall @danieltudosiu encountering checkerboard artefacts and experimenting with ways to mitigate them - do you remember experimenting with interpolation for upscaling, dan?