Make setup lighter about darts HOT 10 CLOSED

Yes, I think we can remove them. Although it would be nice to replace AutoARIMA with ARIMA.

I'll remove AutoARIMA from deps and log a warning that it's not available while keeping ARIMA. Would that it be ok for you?

edit: clarification

from darts.

@plorenzatto Thanks a lot for contributing, I think your proposed changes could be very useful! I think making torch and fbprophet optional like you suggested makes sense. But I'm a bit unsure when it comes to statsmodels, since in addition to being used in 3 models, it is also used in some functions from our statistics module.

I believe these parts should work without any compromises in the core version:

• TimeSeries class
• backtesting module (without TorchForecastingModel support)
• at least some models from the models module
• preprocessing module
• utils module (except the torch part)
• metrics module

Pretty much everything except some of the more fancy forecasting models for now I believe. What are your thoughts?

from darts.

Hi @plorenzatto,

thank you for bringing this up - this is a very good point that's been also in our backlog for a while already.

Is it something you would like to maybe contribute yourself/are you okay with waiting until we deliver this on our side?

from darts.

I'm planning to work on this and send a PR when it's ready.

I have some questions about this:

• Which functionalities/modules must absolutely work with the default installation?
• How did you originally thought of solving this issue?

from darts.

@TheMP I'm working on #177 on this. Could you give me some input on the points that I mentioned in the previous message?

from darts.

I agree with the general idea @pennfranc.

Indeed statsmodels is a dependency of metrics via statistics so better to keep it in the main requirements.

I have some doubts:

• darts/tests/test_dataset.py depends on ..models.torch_forecasting_model._TimeSeriesSequentialDataset not sure if it's meant to be torch specific or not.
• backtesting.explore_models uses AutoARIMA and Prophet by default. Can we remove them?

from darts.

Good points.

• This is definitely specific to our TorchForecastingModel since it tests a subclass of ours of a torch class (torch.utils.data.Dataset). The naming of this test case could probably be made to better reflect this fact.
• Yes, I think we can remove them. Although it would be nice to replace AutoARIMA with ARIMA.

from darts.

Hi - Just wondering if there is an estimate when this change might be in place. Sorry I'm not helping with the work but would like to use the u8darts library in production but previously mentioned dependencies create too large of a Docker image.

from darts.

@dboyleitrs We are evaluating our options to solve this issue. Hopefully we should be able to release this within a couple of weeks or so.

from darts.

This is implemented in develop, and will be part of the next release:

• pip install u8darts installs core only
• pip install u8darts[all] installs everything
• pip install u8darts[torch] installs core + torch models
• pip install u8darts[fbprophet] installs core + Prophet
• pip install u8darts[pmdarima] installs core + AutoARIMA

from darts.