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.

If you have already forked this project before June 1st 2021, please fork again or pull recent changes. Major changes will be also announced on the challenge website. This template repository will have release tags for orientation.

Find an overview of repositories and websites.

Introduction

This is a Renku project. Renku is a platform for reproducible and collaborative data analysis. At its simplest a Renku project is a gitlab repository with added functionality. So you can use this project just as a gitlab repository if you wish. However, you may be surprised by what Renku has to offer and if you are curious the best place to start is the Renku documentation. 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 you fork the renku project and the underlying gitlab repository through the renkulab.io page.

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

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

When cloning this repository and you do not want to immediately download the git lfs-backed renku datasets, please use:

GIT_LFS_SKIP_SMUDGE=1 renku/git clone https://renkulab.io/projects/$YOURNAME/s2s-ai-challenge-$TEAMNAME.git

To be able to pull future changes from the template into your repository, add an upstream:

# in your fork locally
git remote add upstream https://renkulab.io/gitlab/aaron.spring/s2s-ai-challenge-template.git
git pull upstream master

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 navigate to the gitlab page by clicking on "View in gitlab" in the upper right corner. Under "Settings" - "General" - "Visibility" you can set your project private.

Now other people cannot steal your idea/code.

Please modify the README in your fork with your team's details and a description of your method.

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

The scorer follows the code shown in the verification notebook. The scorer's username on gitlab is s2saichallengescorer. You should add it to your project with Reporter permissions. Under "Members" - "Invite Members" - "GitLab member or Email address", add s2saichallengescorer. The scorer will only ever clone your repository and evaluate your submission. It will never make any changes to your code.

5. Add the s2s-ai-challenge topic to your repository

To add the project topic navigate to Settings -> General and then fill in the word s2s-ai-challenge in the Topics field near the top of the page. If you have multiple topics you can separate them by commas.

This allows your repository to be recognized as a participant of the competition. Without this project topic or if you have not added the scorer as a member of your project the automated scoring bot will not evaluate any of your submissions and none of your code or results will be considered for the competition.

Make Predictions

6. 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

7. Train your Machine Learning model

Get training data via

• climetlab
• renku datasets

Get corresponding observations/ground truth:

• climetlab
• IRIDL: temperature and accumulated precipitation

8. 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 ML_prediction_2020.nc, see example.

9. git commit training pipeline and netcdf submission

For later verification by the organizers, reproducibility and scoring of submissions, commit all code, input and output data. For the data files please use git lfs. If you are unfamiliar with git lfs a short introduction can be found here. This is very important because the organizers need to review and reliably reproduce your results. If your results cannot be reliably reproduced then you cannot win the competition - even if your submitted results had the highest score.

After committing, tag your submission and push your commit and tag. The automated scorer will evaulate any tag (regardless of which branch it is on) that starts with the word submission followed by any other combination of characters. In other words, any tags that satisfy the regex ^submission.* will be evaluated by the scorer. In addition, the scorer will only look for the results in a file named ML_prediction_2020.nc located in the submissions folder at the root of each competitor's repository.

Here is an example of a set of commands that would commit the results and add the scorer tag.

# run your training and create file ../submissions/ML_prediction_2020.nc
git lfs track "*.nc" # this will ensure that all *nc files are using lfs and needs to be done only once
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" # add this tag if this is to be evaluated by the scorer
git push --tags

Please note that only submitted/tagged commits will be considered for the competition. If you have code that produces better results after the competition ends and it has not been tagged or is tagged after the competition closed then this will not be considered.

10. 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

• in the s2s-ai-challenge wiki
• all different resources for this competition

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.