Skip to content
Snippets Groups Projects
Forked from Aaron Spring / s2s-ai-challenge-template
60 commits behind the upstream repository.
Aaron Spring's avatar
19fc6764

S2S AI Challenge Template

This is a template repository with running examples how to join and contribute to the s2s-ai-challenge.

You were likely referred here from the public website.

The competition starts in June, so examples are still work in progress and joining the competition not possible until then, but you can already look around. If you fork this project before June, please rebase or fork again in June.

Find an overview of repositories and websites

Introduction

This is a Renku project - basically a git repository with some bells and whistles. You'll find we have already created some useful things like data and notebooks directories and a Dockerfile.

Join the challenge

1. The simplest way to join the S2S AI Challenge is forking this renku project.

Ensure that you do not fork the gitlab repository, but the renku project.

Fork this template renku project from https://renkulab.io/projects/aaron.spring/s2s-ai-challenge-template/settings.

Name your fork s2s-ai-challenge-$TEAMNAME.

Your fork will inherit the tags from the template repo. The tag s2s-ai-challenge is needed for the scorer bot to find your repo.

2. Fill our registration form.

Registrations are not required before October 31st 2021, but highly appreciated for the flow of information.

3. Make the project private

Now check out the gitlab repository by clicking on "View in gitlab". Under "Settings" - "General" - "Visibility" you can set your project private.

Now other people cannot steal your idea/code.

Now please modify the README in your fork with team details and a description of your method.

Please use different branches if you try out different methods. The scorer finds branches from all branches.

4. Add the scorer user to your repo with Reporter permissions

The scorer is not yet ready, but will follow this verification notebook.

Todo: How to add scorer to repo

Make Predictions

5. Start jupyter on renku or locally

The simplest way to contribute is right from the Renku platform - just click on the Environments tab in your renku project and start a new session. This will start an interactive environment right in your browser.

If the docker image fails initially, please re-build docker or touch the enviroment.yml file.

To work with the project anywhere outside the Renku platform, click the Settings tab where you will find the renku project URLs - use renku clone to clone the project on whichever machine you want. Install renku first with pipx, and then renku clone https://renkulab.io/gitlab/$YOURNAME/s2s-ai-challenge-$GROUPNAME.git

6. Train your Machine Learning model

Get training data via

Get corresponding observations/ground truth:

7. Let the Machine Learning model perform subseasonal 2020 predictions

and save them as netcdf files. The submissions have to placed in the submissions folder with filename submission_your_choice.nc, see example.

8. git commit training pipeline and netcdf submission

For later verification of the organizers and reviewers, reproducibility and scoring of submissions, the training notebook/pipeline and submission file ML_prediction_2020.nc must be committed with git lfs:

# run your training and create file ../submissions/ML_prediction_2020.nc
git lfs track "*.nc" # do once, already done in template
git add ../submissions/ML_prediction_2020.nc
git commit -m "commit submission for my_method_name" # whatever message you want
git tag "submission-my_method_name-0.0.1" # if this is to be checked by scorer, only the last submitted==tagged version will be considered
git push --tags

9. RPSS scoring by scorer bot

The scorer will fetch your tagged submissions, score them with RPSS against recalibrated ECMWF real-time forecasts. Your score will be added to the private leaderboard, which will be made public in early November 2021.

The scorer is not active for the competition yet.

More information

Changing interactive environment dependencies

Initially we install a very minimal set of packages to keep the images small. However, you can add python and conda packages in requirements.txt and environment.yml to your heart's content. If you need more fine-grained control over your environment, please see the documentation.