Shape Mismatch Between Input Batch and Predicted Output

[MarinoMarrocu]

I noticed that when running Aurora with ERA5 at 0.25° (but also with HRES T0 at 0.1°), the shape of the input batch tensors is 721×1440 (or 1801×3600), while the shape of the predicted outputs is 720×1440 (or 1800×3600). It seems that the South Pole is missing, as the metadata containing the latitude coordinates does not include the latitude of the North Pole (lat = -90°). Perhaps I’m missing something, but I wonder: in what kind of grid should we consider the model’s predictions to be represented if we want to validate the prediction? Is it a grid that is staggered relative to the one specified in the latitude and longitude metadata? Thanks

[wesselb]

Hey @MarinoMarrocu! Thanks for opening an issue. :)

The model indeed internally drops the last row (corresponding to the south pole) so the resulting tensor can be patched with a patch size of 4x4.

For evaluation, you can also drop the last row in the reference data, meaning that the south pole is not included in the evaluation. This is also what we did.

[MarinoMarrocu]

Thanks!