imshow no longer correctly handles 'bad' (nan) values

[tacaswell]

To help us understand and resolve your issue please check that you have provided
the information below.

• [ v2.x ] Matplotlib version, Python version and Platform (Windows, OSX, Linux ...)
• [ source ] How did you install Matplotlib and Python (pip, anaconda, from source ...)
• If possible please supply a Short, Self Contained, Correct, Example
  that demonstrates the issue i.e a small piece of code which reproduces the issue
  and can be run with out any other (or as few as possible) external dependencies.

plt.imshow(np.ones((5, 5)) * np.nan)

• If this is an image generation bug attach a screenshot demonstrating the issue.
  

Also got some user warnings

 /home/tcaswell/.virtualenvs/dd_35/lib/python3.5/site-packages/numpy/ma/core.py:4089: UserWarning: Warning: converting a masked element to nan.
  warnings.warn("Warning: converting a masked element to nan.")
/home/tcaswell/source/p/matplotlib/lib/matplotlib/colors.py:581: RuntimeWarning: invalid value encountered in less
  cbook._putmask(xa, xa < 0.0, -1)
/home/tcaswell/source/p/matplotlib/lib/matplotlib/cm.py:275: RuntimeWarning: invalid value encountered in less
  rgba[:, :, 3][x < 0.0] = 0

• [ always ] If this is a regression (Used to work in an earlier version of Matplotlib), please
  note where it used to work.

[tacaswell]

It looks like new image re-sampling code loses the mask that gets added on the way into AxesImage which then does not propagate properly through norm / color map.

[mdboom]

Lemme know if you need help with this one.

[mankoff]

I think this bug still exists w/ RGB images. An MWE based on the test_image in the above fix.

fig, (ax1, ax2) = plt.subplots(1, 2)

foo = np.ones((5,5))
foo[1:2, 1:2] = np.nan
ax1.imshow(A)

bar = np.zeros((5,5,3))
bar[:,:,0] = 1
bar[:,:,1] = 0
bar[:,:,2] = 0
bar[1:2,1:2,:] = np.nan
ax2.imshow(bar)

[tacaswell]

In the case of RGBA there is no color mapping. The black coming through is a bit odd.

@mankoff What version of mpl and what backend are you using?

[mankoff]

mpl: 1.5.1
backend: Qt4Agg
OS: OS X.

I'm not sure what you mean by "no color mapping". Is it possible to have transparent values in an RGB displayed with imshow?

[tacaswell]

This original bug in this issue with the new image code which is only on 2.x/master (not released yet).

What I mean by 'no color mapping' is that if you pass a (M, N, 3) or (M, N, 4) array to imshow it does not get run through the normalization / color mapping code path. The [0,1] -> R3 step has a mapping, plus three special colors 'over', 'under', and 'bad'.

I am actually not sure what to expect for an RGB/RGBA input with nans in it. Did this used to do something different? In any case, can you make a new issue?

[efiring]

Something might have gotten lost. We are using safe_masked_invalid in set_data, so I suspect that the mask was being handled prior to the big refactoring. What was being rendered in the masked locations, I don't know.

[efiring]

Contrary to my last comment, I don't think anything has been lost. Looking at v1.4.1, I don't see any indication in the code or the docstring that invalid or masked data has ever been supported for RGB or RGBA arrays. Even though they get passed through safe_masked_invalid, their mask is never checked or used. Maybe the thing to do is to warn if any values are actually masked, and then use .filled(0) to turn them back into ordinary ndarrays. Or raise an exception if any values are masked. I think it should remain the responsibility of the supplier of RGB or RGBA arrays to handle bad values before passing the arrays in to imshow.

[benjimin]

I think it is a useful feature that regions of NaN appear like holes when visualised by imshow. Unfortunately (in matplotlib 2.0.0) imshow raises ValueError if the raster array does not contain any finite values whatsoever.

[tacaswell]

Can you try with 2.0.2 or provide a MWE in a new issue?

I have a memory of fixing something like that recently.