search

google-deepmind / graphcast

Apache License 2.0
4.37k stars 538 forks source link

How diffs_stddev_by_level.nc is computed? #8

Closed omshinde closed 9 months ago

omshinde commented 9 months ago

Hi! Thanks for all the efforts in making this repository open.

I wish to know how the normalization values for scales i.e. diffs_stddev_by_level.nc (provided here - https://console.cloud.google.com/storage/browser/dm_graphcast) is computed. Is it the difference of standard deviation across time-steps?

I must have overlooked it but I cannot find reference for computing this stat in the paper. Thank you for your time.

alvarosg commented 9 months ago

Hi @omshinde, thanks for your message.

This is what we mention in he paper here:

image

The statistics are approximated by sampling a short 2-step sequence from the dataset (the two steps $X^t$ and $X^{t+1}$, separated by 6h), and computing the difference between the two steps. Then you do this for many two step trajectories (ideally the whole dataset, but in our case we did it only for 8192 sequences), and compute the standard deviation of those differences across all of the trajectories, and all of the latitude and longitude points of those trajectories.

omshinde commented 9 months ago

Thank you so much @alvarosg ..Also. apologies for overlooking this from the paper. I am closing this issue as I don't have any further questions as of now.