Earthquake detection use case tutorial

[DarthReca]

This is the draft tutorial for QuakeSet. I have provided two different pipelines:

• Full training using TorchGeo
• RandomForest with TorchGeo model embeddings

I am open to suggestions before I finish adding the explanations.

[calebrob6]

Thanks @DarthReca!!

Quick link to the notebook

[calebrob6]

Notebook ran perfectly for me! The issue with the failing test is because the test is trying to run the notebook end-to-end on a small test-runner VM without a GPU (which will take forever), then timing out.

@adamjstewart -- do you recommend basically skipping all the logic for test purposes?

[adamjstewart]

I don't skip any logic, I just make it faster. Specifically, use smaller dataset subsets, or train for a single step instead of tens of epochs. We use nbmake for notebook testing, and nbmake specifically added support for variable mocking just for TorchGeo: https://github.com/treebeardtech/nbmake?tab=readme-ov-file#mock-out-variables-to-simplify-testing

We use this in many of our other tests, see the Trainer tutorial for an example, search for fast_dev_run. You have to edit the .ipynb directly using something like vim though, I don't know how to do this in Jupyter itself.

P.S. I realize that this is annoying and makes contributing tutorials prohibitively difficult, but in my experience, any code or documentation that isn't actively tested becomes broken in a matter of months, not years.

[DarthReca]

Thanks for the suggestions. If I see it correctly from the logs, the system fails at inference rather than training.

A subset will probably work to pass the tests. I can go for a customizable islice. If you think it is fine.

[adamjstewart]

Another CI limitation is that these runners have very limited storage space, so cutting down the area used for inference could help a lot there too. Also starting with a pretrained model allows you to train on a much smaller dataset much more quickly.

[DarthReca]

Thanks for all the info