Support darts time series models

[sotte]

darts is a great library for time series prediction. It would be great if bentoml supports the darts library.

Note: darts wraps a bunch of existing (time series) libraries such as statsmodels, pytorch, etc.

[aarnphm]

I wonder whats the advantages of this comparing to statsmodels? Looks to me its a more powerful statsmodel. I would like to learn more about your usecase with darts.

[sotte]

@aarnphm thanks for the quick response.

darts allows you to use a bunch of different time series libraries, it wraps them with a nice unifying interface and adds some convenient evaluation functions on top.

So from the data scientist point of view, you can use statsmodels, facebook's prophet, pytorch-forecasting, sklearn, or darts internal models such as nbeats (based on pytorch) in the same fashion. You can basically think in terms of problem (time series) independent of the underlying library.

I'm very interested in this feature and could potentially help out.

[aarnphm]

Hi @sotte, Sorry for the late response. We are currently working on some internal designs changes and will get back asap on how you can contribute.

[hrzn]

An integration of Darts models would be awesome. On top of what has been said I also would like to add that a few key features set Darts apart from libraries like statsmodels is that it supports training models on multiple (potentially a large number) of time series, the time series can potentially each be multidimensional, and some of the models (incl. all deep learning models) can offer probabilistic forecasting in a fairly advanced way (e.g. fitting different kinds of probability distributions).

[ssheng]

We will leave this issue open for the community to upvote for darts support. We also welcome community contribution for adding new framework support. cc: @hrzn

For those who are looking for a workaround, BentoML supports use of custom models through custom models.