I think quickly checking how to fix the isfinite version, we forgot that for histogram with automatic bins np.linspace would have to be fixed as well (not sure how straight forward that is). Although, I guess this probably still doesn't hurt, just does not quite fix the linked issue.

@seberg : That's true. I looked around a bit, and found the following:

• np.linspace is not the only thing stopping np.histogram from working with datetime64 and timedelta64. There are at least these issues:
  • np.histogram._get_outer_edges tries to do addition/subtraction between first_edge/last_edge and floats in order to avoid a division by zero.
  • np.histogram here tries to divide an int by last_edge - first_edge
• np.linspace has only one problem with timedelta64: a multiplication between two timedelta64.
• np.linspace has at least this problems with datetime64:
  • Multiplication with float.

As per my novice understanding, both linspace and histogram needs special logic to deal with datetime64 and timedelta64. And maybe some ufunc level changes might help, e.g. multiply between timedelta64 and float.

I have made an attempt to get linspace to work with datetime64 and timedelta64. Added (passing) tests as well. I had to pepper linspace with datetime specific extra logic in a few places.

Should I submit a PR?

Needs a release note.

@debsankha ping. All this needs is a release note (doc/release/1.17.0-notes.rst), and maybe use Eric's shorter type spec.

OK, done.

Thanks @debsankha .