ecmwf / anemoi-training

Apache License 2.0
17 stars 17 forks source link

Feature: MSE metrics inside/outside regional domain for stretched grid models (KNMI) #130

Open jswijnands opened 1 week ago

jswijnands commented 1 week ago

Description of functionality Four new metrics have been added for stretched grid models. The first two metrics compute the weighted MSE (i) only inside the regional domain and (ii) only outside the regional domain. The former will allow for comparison of LAM and stretched grid models in the future. Further, two additional metrics have been added that compute the contributions of these components to the overall weighted MSE. This enables monitoring which of the two is the main driver of the training process, as it indicates whether we are mainly optimising the loss inside or outside the regional domain.

Configuration options Both features can be turned on or off in the config options by specifying the following stretched grid metrics (sg_metrics) in the diagnostics config:

Outputs Plots of these metrics are generated in MLFlow, specified as ‘wmse_inside_lam_epoch’, ‘wmse_outside_lam_epoch’, ‘wmse_inside_lam_contribution_epoch’, and ‘wmse_outside_lam_contribution_epoch’.

MLFlow plots w-MSE splitting_default_area_weighting

Further comments

FussyDuck commented 1 week ago

CLA assistant check
Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
3 out of 4 committers have signed the CLA.

:white_check_mark: jswijnands
:white_check_mark: havardhhaugen
:white_check_mark: Sooph-knmi
:x: Sophie Buurman


Sophie Buurman seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
You have signed the CLA already but the status is still pending? Let us recheck it.