GraphCast: Fine-tuning (Impact: moderate, effort: high)

[mnabian]

No description provided.

[oublalkhalid]

Hello @mnabian,

Thank you for your excellent implementation using modulus! I was wondering if it's possible to inherit the pretrained checkpoint from the main JAX model. I'm exploring alternatives to leverage the pre-trained model for finetuning in this context.

Have there been any attempts to convert JAX model weights for use in the torch modulus implementation? Thanks in advance!

Best,
Koublal

[Flionay]

Looking forward to further progress on the GraphCast model, or demonstration of training results using Modulus Vs DeepMind jax.