Closed joeloskarsson closed 3 months ago
@sadamov Hope it's ok that I put you to review PRs like this :) I think it's valuable to get a second pair of eyes to look at the changes, and also good for you to get an update on small things I am changing.
The changes to the MEPS Dataset class here are not very important, this is really motivated by moving away from things being too specific for that data.
Thanks for taking a look!
I just realized I forgot to change create_parameter_weights.py
, as the Dataset class is used in there also. Will fix that (should only be a tiny change of index) and then merge.
The batch-static tensor contained forcing that differed between initialization times, but stayed static for all lead times of a forecast. For the MEPS data we used this for the land-water-mask, as this could be different throughout the year, but we could not produce separate values per lead time (as all other forcing).
This PR removes the batch-static features as an explicit extra input. The motivation is:
None
.This PR changes: