[Bug] masking of NaN Z values in pcolormesh

[andim]

In pcolormesh non-finite values (e.g. NaNs) are not currently masked. The function allows the input of masked arrays, so there is an easy workaround of masking the array before passing it in (see http://stackoverflow.com/questions/7778343/pcolormesh-with-missing-values). It would make more sense IMHO, though, to mask the array internally as is done in imshow. This would fix what I believe is a bug, namely that the autoscaling of the Z axis does not work if the array contains NaNs.

[efiring]

That's reasonable, but I think that instead of putting the safe_masked_invalid call in pcolormesh it should go in two places (check--this is based on a very quick look) in _pcolor_args so that it will be applied to the whole pcolor family.

[andim]

Thanks for the quick reply!

You mean putting the check before both of the returns in _pcolor_args I suppose? I was afraid that some of the other pcolor functions might not be able to handle a masked array, but if this is not the case then I'll indeed move the call to safe_masked_invalid up there.

[efiring]

In the development of mpl, masked array support predates nan support, which is the reason there are still places like this that need safe_masked_invalid added.

[tacaswell]

@efiring I could go either way on backporting this to 2.0.x

[efiring]

@andim Looks good. Would you mind squashing the two commits down to one and force-pushing?

[andim]

I had some trouble at first (I'm new to this more advanced git features), but it should be ok now with the latest commit. Thanks again for your comments and looking forward to seeing this fixed.

[efiring]

@andim Thanks for your contribution!

[Alexander-Barth]

@tacaswell It would be great if this feature could be back-ported to 2.0 (except when 2.1 will be released soon). It would help a lot of users coming from Matlab.
(I have to change the coded in my local copy of matplotlib and it does exactly what I wanted. But if I want to share code with my colleagues using features from an not yet released version of matplotlib is not really an option.)

[tacaswell]

The current goal is 2.1 RC in the next month or so.

[Alexander-Barth]

OK, great! Thank you for the information.