Consider adding nonzero/flatnonzero to dask.array

This would need to eagerly evaluated, or perhaps return an imperative value, because the result shape is not known apriori.

This sort of thing could be useful for https://github.com/pydata/xarray/pull/815 if you want to mask out a small region of a very large array. For example, imagine you have a tiled digital elevation model dataset at high resolution covering the globe, and now you want to extract out the region corresponding to California. There are certainly more intelligent indexing strategies for geospatial data, but I could see something like ds.sel_where(ds.region_enum == CALIFORNIA) being convenient.

[mrocklin]

If you don’t want to evaluate the entire result then you’ll need to know which chunks to look for, in advance. Generally this isn’t doable. Dask.array isn’t good when the structure of the computation depends on the values of the array.

[mrocklin]

Unknown dimension lengths are fine

On Wed, Mar 29, 2017 at 1:10 PM, jakirkham [e-mail] wrote:

Is it only unknown chunks or are unknown dimension lengths for the overall array allowed too?

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/dask/dask/issues/1076#issuecomment-290157722, or mute the thread https://github.com/notifications/unsubscribe-auth/AASszF2g6ovAtr1rPVsj7pWhibmoDO-2ks5rqpCbgaJpZM4H98G7 .