Dependencies: torch 1.9+ required

[adamjstewart]

#927 uses torch.linalg.vector_norm, which requires torch 1.9+. This wasn't caught in CI because the tests weren't run after #918 was merged.

P.S. We should require all tests to pass before a PR can be merged. Let me know if you need help updating the branch protection rules.

[qubvel]

Thanks!

It looks like dims=None is not supported even in torch 1.9, we should probably revert to if/else statement in the loss

[adamjstewart]

Looks like 1.9.0 also has issues. Do these make sense to anyone? I'll keep bumping the min version until it works, but let me know if these make sense.

[adamjstewart]

Possibly related to #937

[qubvel]

Torch 1.9 + dims fix looks fine, loss tests passed for minimal version
https://github.com/qubvel-org/segmentation_models.pytorch/actions/runs/11212688281/job/31164191825?pr=937

[qubvel]

We can just merge #937

[adamjstewart]

Do you want to make all tests required now? This will force even older PRs to be rerun before merging.

[qubvel]

Can you provide a bit more details? I was actually thinking about reducing the number of platforms in PRs, or maybe conditioning it on some label, but not sure how this can be implemented 🙂
e.g. run tests for all Python on ubuntu, but macos and windows only for the latest python, then run everything on main.

[adamjstewart]

Can you provide a bit more details?

Basically, make main a protected branch that can only be merged to when a certain set of CI tests pass.

run tests for all Python on ubuntu, but macos and windows only for the latest python, then run everything on main.

You can either include or exclude certain combinations until you get the desired result. But I wouldn't worry too much about the additional CI time, it's free.