Timeline for dropping Python 3.6 support

[weiji14]

Continuing on from discussion in #326 and #409 (comment). Python 3.6 has reached end of life (EOL) as of 23 Dec 2021. This issue is to discuss and track when torchgeo should drop Python 3.6 support.

Is torchgeo still supporting Python 3.6?

For the time being, yes. I would prefer to stick to Python's EOL timeline, but there are many systems that still use Python 3.6. I strongly disagree with numpy's aggressive deprecation timeline, as Google Colab still runs Python 3.7 by default. I think we'll likely compromise and stick to whatever versions of Python that PyTorch supports (currently 3.6–3.9 in the latest release, soon to be 3.7–3.9 in the next release). Unless I can convince @calebrob6 to drop 3.6 support since PyTorch has already dropped it in master.

Originally posted by @adamjstewart in #409 (comment)

[adamjstewart]

PyTorch 1.11 should be coming pretty soon: pytorch/pytorch#72267

In general, we can always backport things like dataclasses and typing-extensions if we need to. There are rarely newer language features that we need that can't be backported.

[calebrob6]

I only have the slightest preference for not dropping 3.6 (it just seemed silly to do some extra work to make it slightly harder to use the library 😄). PyTorch 1.11 seems like a great reason to drop it.

[adamjstewart]

Based on https://pypistats.org/packages/torchgeo very few of our users are using Python 3.6.