Tensforflow to pytorch migration

[Antonio Politi]

Hello,
I realized that tensforflow discontinued the support on Windows Native for tensorflow > 2.11 (https://www.tensorflow.org/install/pip). This makes stardist on the medium to long term inaccessible to Windows users. Please note that most of the microscopist/biologist that would like to use stardist are using Windows.

From the user side I see following options:

1. Use lower version on tensorflow (not great as it get you stucked on a version)
2. Use Linux for training and prediction. Good option, but there is a certain reluctance for windows users to use Linux, so it is harder to deploy in a facility context.
3. Use WSL2. No experience with this, but it sounds horrible :-( with a lot of overhead. Have you tried it? Is it stable?
4. Use a different framework. With pytorch the support on Windows work well. For instance cellpose works like a charm. I have seen a few non-official implementation of stardist using pytorch, I never tried them (e.g. https://github.com/ASHISRAVINDRAN/stardist_pytorch, https://github.com/hthierno/pytorch-stardist).

Do you have any opinion, suggestions, plans?

Antonio