Closed AUdaltsova closed 2 weeks ago
By this comment I actually mean should we lock into np.float32
or np.float64
etc of the values in the DataArray.
Right now I'm leaning towards we shouldn't, although controlling the dtype has bitten us before. For example, the training UKV data is float16 which means that when we have visibility with a value higher than 2*16-1 ~= 65000 it was overflowing to infinity. The visibility is measured in metres and can be more than 65km often enough. We didn't realise this when preparing the training data. Its the reason for this awful hack in PVNet.
In production we had the UKV data in a different format which wasn't overflowing. Therefore we had to hack that into the app as well to resave the data as float16
change to not planned
Detailed Description
Responding to this TODO in ecmwf provider:
I think having "channel" dimension basically does that for you? All providers are guarnteed to have it, and it guarantees that the variables are flattened into a DA