openhackathons-org / end-to-end-ai-for-science Goto Github PK

The current FourCastNet training uses modified hyperparameters from the paper "FourCastNet: A Global Data-driven High-resolution Weather Model using Adaptive Fourier Neural Operators", the paper can be accessed here

The goal is to tune hyperparameters for Modified FourCastNet in Modulus to achieve convergence in fewer timesteps. The current model is trained on 6 months of data and is used to predict a 1-week long prediction for validation.

the following link: https://docs.nvidia.com/deeplearning/modulus/text/features/configuration.html is broken, it is found in the Modulus bootcamp container, in the projectile motion notebook in step 5.

Hi,

I'm assisting a user who is trying to work through this material on Perlmutter at NERSC. I was wondering if anyone has notes on Perlmutter specific directions they could share for getting setup on the machine.

Thank you!

Erik

The current bootcamp materials make sure of a Single GPU for training the models.

The goal of this feature request is to have an additional section that allows the participants of this material to train models at scale in both multi-GPU and multi-node environments.

A good starting point would be to refer the documentation here

For files which are hosted on google drive, a new version of gdown is needed to maintain compatibility. Currently users can get around this by manually upgrading gdown with pip, but ideally the container should have the latest version.

The Omniverse get-kit.sh script is not present, how can I build docker image?