BUG: fix np.ma.median with only one non-masked value and an axis argument

[lesteve]

Fix #8029.

[lesteve]

np.copyto(low, high, where=odd) doesn't unmask the values in contrary to low[odd] = high[odd].

[charris]

Now there are indexing errors because odd is boolean...

[lesteve]

Not sure I am following, do you have an example where things go wrong?

[charris]

There were problems when odd was a scalar or scalar array. Perhaps that no longer happens with the other changes...

[charris]

So now that we are using the count function, that may be fixed. In fact, by the time we get there 0-d inputs should have been raveled to 1-d and 1-d arrays have been handled, so the asorted.ndim > 1 check should not be needed either.

I don't think any of the values should be masked except for zero values of count, which brings up the question of negative indices in creating low.

[lesteve]

I am afraid I haven't got a grasp of the issue as good as yours, but what about replacing h - 1 by np.maximum(h - 1, 0) ?

[lesteve]

This way you can keep using np.copyto as previously.

[charris]

I think that should work also, I was just a bit concerned about speed. OTOH, it is a bit more obvious. Does is work?

[charris]

If we do that, can probably remove low._sharedmask = False also.

[lesteve]

I think that should work also, I was just a bit concerned about speed. OTOH, it is a bit more obvious. Does is work?

The tests pass on my machine, I pushed the change. Let's see what Travis and AppVeyor have to say!

If we do that, can probably remove low._sharedmask = False also.

OK I can try that too afterwards.

[charris]

When (if) this passes the tests, would you squash the commits?

[lesteve]

When (if) this passes the tests, would you squash the commits?

The tests were passing so I squashed the commits. BTW just in case you didn't know about it you can do squash + merge via the github UI these days.

[charris]

I don't really trust git to do it right after trying their revert button ;) Haven't tried squashing though.

[charris]

Thanks @lesteve. I still want to revisit ma.median at some point, but I think we now have something that works. Of course, I've thought that before. If nothing else, the tests need to have better coverage.

[lesteve]

Great!

If nothing else, the tests need to have better coverage.

If you have suggestions of weird edge cases to add I'd be happy to give it a go.

[charris]

If nothing else, there should be tests for different values of ndim, say up to at least three, with all different valid axis values tested with different amounts of masking from all to none, and maybe empty arrays also (which will currently fail for some cases). There is also the out and overwrite_input keywords. It is a large combinational mess, but if we had had such tests we wouldn't have run into so many problems.

[lesteve]

Thanks for the suggestions, I'll try to find some time to add more tests.