imshow interpolation uses masked values

Bug report

I recently upgraded to v2.0.0 and noticed that the new imshow interpolation is not honoring masked values. This is likely related to the new edge handling, described here: http://matplotlib.org/examples/pylab_examples/image_interp.html

Code for reproduction

a = np.random.random((7,7))*1000
ndv = -9999.0
a[2:4,1:5] = ndv
am = np.ma.masked_equal(a, ndv)

plt.figure()
#Plot with no interpolation, looks fine
plt.imshow(am, interpolation='none')
plt.title("interpolation='none'")
#Plot with bilinear interpolation, the -9999.0 values are used 
plt.imshow(am, interpolation='bilinear')
plt.title("interpolation='bilinear'")

Actual outcome

Expected outcome

The interpolation should ignore masked values as in previous versions.

Matplotlib version

Matplotlib v2.0.0, Python 2.7.13, OSX
installed from pip

It is honoring it, just in weird ways, will have a PR shortely:

Note that in carves out a larger area because the region of screen pixels that get input from the masked data values is larger.

The fix seems to break something else, will get to this tomorrow.

This is way nastier than I expected 😞

@tacaswell

@tacaswell Doesn't this still require some tweaks (cf. the last comment in #8024 and #8183)?

Yes, on my to-do list for tonight

Note that in carves out a larger area because the region of screen pixels that get input from the masked data values is larger.

This seems still to be the case, would It be possible for interpolation to not change the size of the masked area?

This seems still to be the case, would It be possible for interpolation to not change the size of the masked area?

No, because the way the interpolation works is by (effectively) convolving the input data with a fixed size kernel. Any place you evaluate the kernel that overlaps with a masked point also needs to be poisoned (e.g. if you drop a np.nan into any interpolation method then the output is always np.nan).

This seems still to be the case, would It be possible for interpolation to not change the size of the masked area?

No, because the way the interpolation works is by (effectively) convolving the input data with a fixed size kernel. Any place you evaluate the kernel that overlaps with a masked point also needs to be poisoned (e.g. if you drop a np.nan into any interpolation method then the output is always np.nan).

Yes I understand. My problem is not that I have np.nan, but simply that I want to mask some voxels from the image being shown. One solution could be to add a mask option to imshow, then the interpolation could be done before masking. Or another option would be to detect if the data is masked, then remove the mask (perhaps also replace masked values with zero), do the interpolation and then reapply the mask.

If you want the masked pixel to be transparent, you are stuck, if you want to make it white, you can do something like

mask = np.ones((*data.shape, 4))  # RGBA of the right size
mask[~to_hide] = 0                # make the pixels you want to keep transparent
im_data = ax.imshow(data)
im_mask = ax.imshow(mask, zorder=im1.get_zorder() + 1, interpolation='nearest')

Navigation Menu

Repository navigation

imshow interpolation uses masked values #8012

Bug report

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

imshow interpolation uses masked values #8012

Description

Bug report

Activity

tacaswell commented on Feb 3, 2017

tacaswell commented on Feb 3, 2017

tacaswell commented on Feb 3, 2017

tacaswell commented on Feb 5, 2017

QuLogic commented on Mar 13, 2017

tacaswell commented on Mar 14, 2017

gurraburra commented on Aug 29, 2024

tacaswell commented on Aug 29, 2024

gurraburra commented on Aug 29, 2024

tacaswell commented on Aug 29, 2024

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions