There is discrenpancy between GCP and the prod NWP data. the x,y grid is not the same and therefore cannot use the same region-masking model. For now we have a quick workaround by quickly altering the shape of the prod data so that it matches the training data. This allows the same mask to be used and the inference datapipe to work - BUT does not provide the same results.
TODO - Change this before final productionization by interpolating onto training axis or change the training data to more accurately reflect prod data.
There is discrenpancy between GCP and the prod NWP data. the x,y grid is not the same and therefore cannot use the same region-masking model. For now we have a quick workaround by quickly altering the shape of the prod data so that it matches the training data. This allows the same mask to be used and the inference datapipe to work - BUT does not provide the same results.
TODO - Change this before final productionization by interpolating onto training axis or change the training data to more accurately reflect prod data.