Thank you for your amazing code! It's been extremely helpful for some dissertation projects.

I've recommended your codebase to quite a few academic/industry researchers. Their biggest adoption hurdle seems to be getting PyTorch & Tensorflow to play nicely together with CUDA.

Do you think it's possible to remove the Tensorflow dependency altogether? (i.e. a pure-Pytorch approach?) Or do you recommend against re-implementing fused_bias_act / upfirdn2d in Pytorch?

Please let me know what you think - thanks!