Contributing to tsai

[dnth]

I am fairly new to fastai and tsai. However I am inspired to start contributing to tsai because I think it really helps make time series problem more accessible. tsai has also helped me tremendously in my time series regression problem.

Is learning about nv_dev the first step to start contributing? Or would you point me to other resources?
Is there any other aspect of tsai that anyone can help to improve?

Thanks for the hard work in maintaining this awesome lib.

[oguiza]

That’s great @dnth!
Thanks for offering to contribute to tsai.
Learning nbdev would be a good start. Something I used to learn more about it is to create a dummy repo, clone it to my laptop, make changes and commit them. It’s easier to practice when you know you can’t break anything.
I’d say there are 4 requirements where some expertise is required:

• Git & GitHub
• nbdev
• fastai
• tsai: it’d be good to understand the tutorial notebooks.

There are multiple ways in which tsai can be improved/ developed. I always have more ideas than time, and I’m you’ll bring your as well. So if you are interested, there’s lot of areas where you can contribute.

[dnth]

Sign me up!

[oguiza]

Hi @dnth,
I'd like to know more about what areas of tsai you are interested in developing. In my view, there are many :)
Here's a list of potential areas I'd like to integrate into the library at some point in the future:

• tasks:
  • forecasting
  • imputation
  • anomaly detection
  • segmentation
  • clustering
  • self-supervised
  • semi-supervised
• type of time series:
  • missing values
  • irregular
  • with covariates (categorical and continuous metadata)
• types of models:
  • transformers
  • multimodal (time series + categorical + continuous)
  • tabular models
  • new state-of-the-art models: CNNs, Transformers, Hybrid, ...
• Interpretability:
  • Integration of Captum
• MLOps:
  • Docker
  • TorchServe
• maintenance:
  • fix bugs
  • improve testing
  • improve documentation
  • adapt the library to fastai v3 when available
  • adapt tsai to package upgrades (Pytorch, ...)

As you can see, there're a lot of things that could be done. That's why I'm more than happy to accept contributors to tsai.

Are you interested in any of these in particular?

[dnth]

Wow that is a big list! I like to learn and contribute more in types of models. Recently, im quite fascinated with using reinforcement learning with time series. Any chance that makes into the list?

[oguiza]

Could you please share an example (maybe a paper) of this (RL applied to TS)?

[williamsdoug]

Hi @dnth. Here's a couple startup issues I encountered when submitting my first pull request to tsai:

• After you clone the tsai repo, be sure to execute the command nbdev_install_git_hooks. This needs to be done only once, but must to be performed before any new commits.
• Before executing nbdev_build_docs all package dependencies must be available. In my case since I use Conda, I execute conda activate <your_env_name>. At build time nbdev seems to ignore the kernel associated with your jupyter notebooks, hence the need to explicitly specify environment.

[dnth]

@williamsdoug thanks for the heads up! Those two points could potentially save me hours :)

@oguiza i realized RL applied to TS mostly are in the case of algorithmic trading time series.. Probably out of TSAI scope..

[oguiza]

Yes, Reinforcement Learning is out of the scope of this library for now.
I agree it's a very interesting area but as you said, there's still a very long list of pending items that I believe may add substantial value to practitioners working with time-series data.