tensorflow version bump

[tmchartrand]

Was there a reason the tensorflow requirement was bumped from 2.5 to 2.17? The commit message calls it a rollback but that's actually a much stricter requirement which doesn't seem necessary, and causing issues for me.

[tmchartrand]

(@alam-shahul ?)

[alam-shahul]

Hmm, looking at my environment for testing this, I actually have TensorFlow 2.14.0. So I might have meant to roll it back to that.

If you install it in a fresh environment and then downgrade to 2.14.0, does it work for you? I remember I had to play around with the versions of TensorFlow to get something that worked for my setup.

[tmchartrand]

I'm currently using TF 2.9 and it works just fine - 2.14 wouldn't be a downgrade, so the looser requirement should be used unless there's a reason to make it stricter.

[alam-shahul]

Well, I suspect that it depends on the environment you're using... it seems like recent versions of TensorFlow are quite a mess in terms of dependencies (see tensorflow/tensorflow#62075 (comment) for example). So many of the minor versions between 2.9 and 2.14 may also be broken, which is another reason to bump the version.

[tmchartrand]

the typical way of excluding problematic versions would be to explicitly exclude the ones with issues, eg add !=2.10 to the specifier, not create an unneccessary minimum that excludes versions known to work.

[alam-shahul]

That makes sense, unfortunately I am not that invested that I want to diagnose exactly which versions of TensorFlow + other dependencies are compatible with this package.

Actually, something I recently read says that one should have as few pinned version dependencies as possible in the pyproject.toml. So perhaps we should just remove the versioning altogether and let people figure out their own TensorFlow installation issues as they arise.

[willtownes]

@alam-shahul unless this breaks your installation I would favor Tom's suggestion to go with the most permissive version of tensorflow possible. I don't want to remove the versioning altogether because there is a big difference between tensorflow 2.5 and earlier versions (maybe we could get away with 2.0 but if memory serves even that didn't work when I was developing the package. We definitely don't want tensorflow 1.x).