[Bug]: scalar mappable format_cursor_data crashes on BoundarNorm

[jklymak]

Bug summary

In 3.5.0 if you do:

import matplotlib.pyplot as plt
import numpy as np
import matplotlib as mpl

fig, ax = plt.subplots()
norm = mpl.colors.BoundaryNorm(np.linspace(-4, 4, 5), 256)
X = np.random.randn(10, 10)
pc = ax.imshow(X, cmap='RdBu_r', norm=norm)

and mouse over the image, it crashes with

File "/Users/jklymak/matplotlib/lib/matplotlib/artist.py", line 1282, in format_cursor_data
    neighbors = self.norm.inverse(
  File "/Users/jklymak/matplotlib/lib/matplotlib/colors.py", line 1829, in inverse
    raise ValueError("BoundaryNorm is not invertible")
ValueError: BoundaryNorm is not invertible

and interaction stops.

Not sure if we should have a special check here, a try-except, or actually just make BoundaryNorm approximately invertible.

Matplotlib Version

main 3.5.0

[anntzer]

I think the correct fix is to specifically check for BoundaryNorm there and implement special logic to determine the positions of the neighboring intervals (on the BoundaryNorm) for that case.

[yumasheta]

I tried returning the passed in value instead of the exception at

matplotlib/lib/matplotlib/colors.py

Line 1829 in b2bb7be

raise ValueError("BoundaryNorm is not invertible")

return value

and it seems to work. At least the error is gone and the values displayed in the plot (bottom right) when hovering with the mouse seem also right. But the numbers are rounded to only 1 digit in sci notation. So 19, 20 or 21 become 2.e+01.
Hope this helps.

[yumasheta]

Maybe a more constructive suggestion for a change in

matplotlib/lib/matplotlib/artist.py

Line 1280 in b2bb7be

if np.isfinite(normed):

that tries to fix the error:

ends = self.norm(np.array([self.norm.vmin, self.norm.vmax]))
if np.isfinite(normed) and np.allclose(ends, 0.5, rtol=0.0, atol=0.5):

This way, because BoundaryNorm doesn't map to 0...1 range but to the indices of the colors, the call to BoundaryNorm.inverse() is skipped and the default value for g_sig_digits=3 is used.
But I'm not sure if there can be a case where BoundaryNorm actually maps its endpoints to 0...1, in which case all breaks down.

[raphaelquast]

hey, I just ran into this while plotting data... (interactivity doesn't stop but it plots a lot of issues) any updates?

[tacaswell]

@raphaelquast no, this still needs someone to work on it.

Labeled as a good first issue as there in no API design here (just returning an empty string is better than the current state).

[yumasheta]

I can give a PR, that is based on my last comment a try. But my concerns/questions from then are still valid.

[tacaswell]

As this is all internal code, I think we should just do an isinstance(norm, BoundaryNorm) check and then act accordingly. There is already a bunch of instances of this in the colorbar code as precedence.

[yumasheta]

After thinking about my suggestion again, I now understand why you prefer an isinstace check.
(In my initial suggestion one would have to check if ends maps to (0,1) and if normed is in [0, 1] for the correct way to implement my idea. But these conditions are taylored to catch BoundaryNorm anyway, so why not do it directly.)

However, I suggest using a try block after

matplotlib/lib/matplotlib/artist.py

Line 1305 in c33557d

if np.isfinite(normed):

# Midpoints of neighboring color intervals.
try:
    neighbors = self.norm.inverse(
        (int(normed * n) + np.array([0, 1])) / n)
except ValueError:
    # non-invertible ScalarMappable

because BoundaryNorm raises a ValueError anyway and I assume this to be the case for all non-invertible ScalarMappables.
(And the issue is the exception raised from inverse).

The question is:
What to put in the except? So what is "acting accordingly" in this case?
If BoundaryNorm isn't invertible, how can we reliably get the data values of the midpoints of neighboring colors?

I think it should be enough to get the boundaries themselves, instead of the midpoints (though both is possible) with something like:

cur_idx = np.argmin(np.abs(self.norm.boundaries - data))
cur_bound = self.norm.boundaries[cur_idx]
neigh_idx = cur_idx + 1 if data > cur_bound else cur_idx - 1
neighbors = self.norm.boundaries[
    np.array([cur_idx, neigh_idx])
]

Problems with this code are:

• boundaries property is specific to BoundaryNorm, isn't it? So we gained nothing by using try .. except
• more checks are needed to cover all cases for clip and extend options of BoundaryNorm such that neigh_idx is always in bounds
• for very coarse boundaries compared to data (ie: `boundaries = [0,1,2,3,4,5]; data = random(n)*5) the displayed values are always rounded to one significant digit. While in principle, this is expected (I believe), it results in inconsistency. In my example 0.111111 would be given as 0.1 and 0.001 as 0.001 and 1.234 would be just 1.

[tacaswell]

boundaries property is specific to BoundaryNorm, isn't it? So we gained nothing by using try .. except

This is one argument in favor of doing the isinstance check because then you can assert we have a BoundaryNorm and can trust that you can access its state etc. One on hand, duck typeing is in general a Good Thing in Python and we should err on the side of being forgiving on input, but sometimes the complexity of it is more trouble than it is worth if you really only have 1 case that you are trying catch!

I assume this to be the case for all non-invertible ScalarMappables.

However we only (at this point) are talking about a way to recover in the case of BoundaryNorm. Let everything else continue to fail and we will deal with those issues if/when they get reported.