search

aditya-grover / climate-learn

Source code for ClimateLearn
MIT License
310 stars 49 forks source link

Question Regarding Bilinear Interpolation for Downscaling before DL models #121

Closed Escape142 closed 6 months ago

Escape142 commented 7 months ago

Hello,

I've noticed the use of bilinear interpolation for downscaling climate data in loaders.py here. Given the complexity of climate data, I'm curious about this choice.

Could you share the rationale behind using bilinear interpolation?

Thank you in advance.

tung-nd commented 7 months ago

We only perform bilinear interpolation before feeding it to the DL model. This is a common practice. The idea is that the DL will learn to fix the error of bilinear interpolation

Escape142 commented 7 months ago

Thank you for your reply!

Can you clarify why did you use exactly 16 shards here?

tung-nd commented 7 months ago

so each shard is a collection of data we load to memory at once. If you increase the shard size you will have to read the data from disk less but each read will take longer, and vice versa. We found 16 to provide a good balance

Escape142 commented 7 months ago

Could you please specify the sources for the land-sea mask and orography constant fields necessary to replicate the results presented in the paper?

tung-nd commented 7 months ago

Can you check this https://huggingface.co/datasets/jasonjewik/climate-learn/tree/main/weatherbench/era5/5.625deg

Escape142 commented 7 months ago

Thank you!

According to the paper, ClimaX was supposed to be incorporated into ClimateLearn. However, it appears to be missing at the moment. Is my understanding correct?

tung-nd commented 7 months ago

It's not been pushed to the main branch. You can checkout the tung_exps branch which has the climax implementation