Closed mountain closed 3 years ago
And the default demo model is for a 3-day prediction of t850, and can reach a performance of 1.98 which is at same level of Weyn et al. 2020 (iterative). Meanwhile we did not introduce complex coordinate transformation and more data from other datasource.
Sorry, there is a bug in the evaluation routines, I drawback the above claim. Sorry for that.
This is great. Let me know if you get results once the evaluation is fixed. We can put it on the WeatherBench leaderboard.
Thanks, I am working on it. Since learning rate need be re-tuned, it cost some time to test all the cases(t850 and z500, 3d/5d).
Hi, buddies,
Just for your information, we released a weatherbench toolkit for model development in PyTorch, and yesterday we upgrade the project
Hope this could help the community.
Mingli