Questions about the performance of representing

[wongchieh]

Hi openclimatefix !

It is awsome to see that you collect so many advanced AI models for weather prediction and provide the training fracture.

I have some questions about using this in a correct way.

First is how good is it to represent the model, to my understanding what you do here is to re-write the code based on the published paper, which is not officially? Therefore, how good is the accuracy of model here compared with the paper, for example, is the gencast implemented here as good as what the google does? So as for other models 'FGN, Aurora, Fengwu...etc'

Second is how to use this. It seems like you put different model together and we can use different part of the models very flesible. For example, we add use Lora or diffusion to the model together. It sounds quiet useful, because we can focuse more on our project instead of the model coding. Could you give me some pipeline on traing and forecast the model ? or could I make it working after I read the 'test' part ?

Best,
cheih

[jacobbieker]

Hi,

A lot of these models are more reference implementations based off of the papers, we haven't necessarily trained all, or even most of them. GenCast has been trained, although the implementation here is not as good as the official Google one. Most of the others are just implementations, so they would need to be trained before they could be evaluated.

It should be a fairly flexible architecture that can swap in different components. Right now, you can look at the tests to see how some of it works, and should go off of the tests to see what is functional. The documentation is a bit out of date unfortanately.

Thanksm
Jacob

[wongchieh]

Dear Jacob,

Thanks for you reply.

I hope one day you could provide details on training all these. And most interesting is how good is the performance compared to the official implementation, this determines whether the implement is good.

Best,
Chieh.