Stephan Hoyer

> Is it possible to preserve dtype when persisting xarray Datasets/DataArrays to disk? Unfortunately, there is a frustrating disconnect between string data types in NumPy and netCDF. This could be...

I'm pretty sure this was intentional -- `mean` of a string is not well defined. Could we make this opt-in, by adding a new keyword argument -- perhaps something like...

The main reason why we used a batch size of 1 per device (and also model parallelism) is that we found training scaled poorly with batch size -- we saw...

I think the tests will need to be updated to add new cached images. Is that something I can do, or that happens automatically?

Turning a record array into a dict of arrays is pretty straightforward, e.g., `arrays = {name: my_recarray[name] for name in my_recarray.dtype.names}` You could then pass this into `xr.Dataset`, but you'll...

We could potentially pick dimension names automatically, but it's not an entirely obvious this to do since passing a dict of numpy arrays into the `xarray.Dataset` constructor isn't supported (but...

Thanks for the report! Since this artifact only shows up in the latitudinal direction, my guess is that it is somehow related to errors in the spherical harmonic transform. Potentially...

We intended to open source these details but it looks like they were accidentally dropped from the checked in configuration files on GitHub. I'll see if we can dig these...

As I recall, netcdf-C has always been OK _reading_ scalar attributes (we have compatibility tests for this). It just doesn't write them. So I think this would pronbably be OK...

One choice we could make here would be to automatically read length one vectors as scalars, similar to netCDF4-python. This would be propagating a poor design choice (there are use-cases...