Determine minimum supported dependency versions

[adamjstewart]

Before releasing, we should determine the minimum supported version of each dependency. We should also consider a test with this version just to make sure it doesn't change.

[adamjstewart]

For testing, I think we can add a matrix field like:

matrix:
    dep-version: [oldest, latest]

Then, during action setup, if: ${{ matrix.dep-version == 'oldest' }}, we can run sed on requirements.txt to replace >= with ==.

We could also use an include: so that we don't have to run every test twice. Still haven't decided if we need to test this just for pytest or for all actions. Also not sure how to detect the minimum version without trying everything and seeing what breaks.

[adamjstewart]

pypa/pip#8085 would greatly help with this if it ever gets implemented.

[calebrob6]

We depend on the latest release candidate of pytorch-lightning (1.4.0) to fix an issue with mypy (the problem where mypy doesn't think a LightningModule's hparams property is indexable)

[adamjstewart]

I would like to differentiate between dependency versions required for torchgeo to function and versions required for things like mypy to pass. We have hundreds of comments across torchgeo that ignore mypy warnings due to missing typing. These will likely change frequently. So the actual version required by torchgeo should be the minimum where the pytest tests pass. For things like mypy/Sphinx, we will always use the latest release.

[adamjstewart]

Now that we're using requirements.txt and dependabot, I think it wouldn't be unreasonable to add a CI test that uses a separate requirements-min.txt with the minimum known working versions of our deps. We can then make sure the full range works as expected.

[adamjstewart]

Just discovered that matplotlib does the same thing: https://github.com/matplotlib/matplotlib/blob/main/requirements/testing/minver.txt