Add AutoregressionTask

[keves1]

This PR adds the following:

• Autoregression Trainer (see Time Series Support #2382)
• LSTM Sequence-to-Sequence (Seq2Seq) model for time series forecasting
• Air Quality time series data set from the UCI Machine Learning Repository

Together these additions enable training air quality forecasting models as well as other autoregression tasks as additional datasets and models are added.

[keves1]

@adamjstewart See below, this test passes when I run it locally but it is failing here. I don't understand this error message, do you know how to resolve this?

FAILED tests/trainers/test_autoregression.py::TestAutoregressionTask::test_trainer[True-air_quality] - ValueError: Does not validate against any of the Union subtypes
Subtypes: [<class 'NoneType'>, <class 'torchgeo.trainers.base.BaseTask'>]
Errors:
  - Expected a <class 'NoneType'>
  - Unexpected keyword arguments: `num_outputs`
Given value type: <class 'jsonargparse._namespace.Namespace'>
Given value: Namespace(class_path='torchgeo.trainers.AutoregressionTask', init_args=Namespace(model='lstm_seq2seq', input_size=3, input_size_decoder=1, hidden_size=1, output_size=1, target_indices=[2], encoder_indices=[2, 12, 13], decoder_indices=[2], timesteps_ahead=1, num_layers=1, loss='mse', lr=0.001, patience=10, teacher_force_prob=None))

[adamjstewart]

Looks like a great start! For datasets, I'm wondering if we also want to consider spatiotemporal datasets like the one you found and whether we should add GeoDatasets. For models, I'm wondering if we should also add a GNN since that's commonly used for non-gridded data. Some kind of model that combines both and can make spatiotemporal predictions would be cool. But none of those are required for a first pass. Let's first clean this up and get it merged.

I don't know why the tests are failing, but it's only the minimum version tests, so worst case scenario we can bump the minimum supported versions. We're already doing that when we drop Python 3.10 support anyway, so I would ignore them for now and focus on everything else.

[adamjstewart]

Confirmed the license from https://archive.ics.uci.edu/dataset/360/air+quality

[adamjstewart]

Why is this needed? Is there a problem with the base class implementation?

[keves1]

I don't think we want to be applying a Kornia augmentation (self.aug) to this data since it is not an image, so this was my way of avoiding that. I think your suggestion for a no-op data augmentation would make this override unnecessary.

[adamjstewart]

Should we add some kind of no-op data augmentation to override the base class?

[keves1]

That would be great, is there something like that in Kornia? I didn't see anything when I looked through the docs. Or is there another way to do it?

[adamjstewart]

No not really. We would have to write our own, or use one that won't apply like normalization.

[adamjstewart]

Instead of documenting and passing all of these arguments one at a time, maybe we could just have an extra **kwargs argument that gets passed to the LSTM class?

[keves1]

That sounds great, I'll work on that.

[adamjstewart]

Suggested change

	loss: str = 'mse',
	loss: Literal['mse', 'mae'] = 'mse',

[adamjstewart]

Can we add plotting here?

[adamjstewart]

Do we somewhere revert normalization before calculating metrics? Are we normalizing both inputs and outputs or only inputs?

[adamjstewart]

I wonder if we should be adding torch.no_grad to more of our tests to speed them up...