Some nan functions don't exist for certain versions of NumPy about dask HOT 5 CLOSED

Yes, we should probably guard these as we guard nanprod

from dask.

That seems perfectly reasonable. I started doing that, but I think they were being imported dask/array/init.py#3f5e6e69137be3b4ffffc7cc3390d19d6c6b73e5. Is it ok if this is eliminated?

I noticed that you were just using nansum. So, I figured it would be fine to have your functions and skip the wrappings, which I did here ( https://github.com/jakirkham/dask/tree/numpy_1_7_support ).

However, I ran into another surprise, which is nansum does not supply the keepdims argument. So, whatever solution is settled upon should probably have a workaround for keepdims, as well.

from dask.

They should probably be imported in __init__.py if they are available. Not sure on the least ugly way to do this (maybe __all__ in reductions.py?).

Skipping just the wrappings sounds great.

Arg, not supporting keepdims is a pain. At the moment I don't see a clear way around this other than by fundamentally changing how we do reductions. This might require some cleverness but is probably possible. Thoughts?

from dask.

Using __all__ seems reasonable.

Good to hear.

As for the last bit, I think this should be doable. I tried adding a wrapper function called keepdims_wrapper, which appears to work ( jakirkham@89b9395 ). It could be applied at the reduction stage (as done in the commit), which guarantees that all functions used have keepdims of some sort. Alternatively, it could be applied earlier when wrapping all the numpy functions, which would reduce the performance hit of doing the actual wrapping and inspection.

from dask.

Nevermind, it appears following the last case is required.

from dask.