Switch from SMP to TorchSeg

[adamjstewart]

Summary

We should consider switching from SMP to TorchSeg for our segmentation models.

Rationale

SMP is no longer maintained. TorchSeg is a fork providing important features we need:

• Support for modern versions of timm
• Support for ViT backends
• Fewer dependencies
• Better CI

Implementation

Should be a simple s/segmentation_models_pytorch/torchseg/g.

However, there are a few more things I would like to see in TorchSeg before committing to the switch:

• Dependabot used to manage all dependencies, especially timm
• Minimum version testing, especially timm

These are things I can set up myself when I have time (hopefully soon!)

Alternatives

I have not yet found a better alternative to SMP.

If SMP becomes maintained once again, we could contribute back all of the improvements in TorchSeg. But I find this unlikely, and I like the control we have over TorchSeg if we need to quickly add new features.

Additional information

@isaaccorley I'm opening this issue to make sure I don't forget to do this before the 0.6 release sometime in April.

[isaaccorley]

I can work on figuring out why dependabot and codecov wasn't working. I'll also try running some tests with different version of timm to see what happens.

[adamjstewart]

Seems like SMP is now maintained, we just need to start contributing PRs there instead of to TorchSeg.