Don't let margins expand polar plots to negative radii by default. #13980
Conversation
| @image_comparison(baseline_images=['streamplot_masks_and_nans'], | ||
| tol=0.04 if on_win else 0, | ||
| remove_text=True, style='mpl20') | ||
| tol=.1, remove_text=True, style='mpl20') |
There was a problem hiding this comment.
Why bump the tolerance instead of replacing the test image?
There was a problem hiding this comment.
There was a problem hiding this comment.
We definitely shouldn't bump tolerances because the actual test image has changed, as this means any small changes to the images in future will not be picked up.
There was a problem hiding this comment.
Sure, I just forced the axes limits to their old values instead...
lib/matplotlib/axes/_base.py
Outdated
| # tolerances (whose value come from isclose()) must be used due to | ||
| # floating point issues with streamplot. | ||
| # Index of largest element < x0 + tol, if any. | ||
| i0 = stickies.searchsorted(x0 + 1e-5 * abs(x0) + 1e-8) - 1 |
There was a problem hiding this comment.
What is the 1e-8 doing for us here that the 1e-5*abs(x0) is not?
There was a problem hiding this comment.
x0 could be equal to zero and we would still need some slop... the actual value comes from what isclose() uses, I didn't really want to bother finding something better (really the correct thing to do should be to fix streamplot() so that we don't need slop at all, but that's a separate project).
There was a problem hiding this comment.
Can you factor that out into a function? This is used twice and the function removes the magic numbers.
def tol(x, rtol=1e-5, atol=1e-8):
"""A tolerance for x such that x+/-tol is considered close to x."""
return rtol * abs(x) + atol
There was a problem hiding this comment.
That doesn't really work because in one case it's plus and in the other it's minus. I moved the values to their own variables, though.
There was a problem hiding this comment.
But that would just be x0 + tol(x0) and x1 - tol(x1) in the code.
lib/matplotlib/axes/_base.py
Outdated
| if self.get_xscale().lower() == 'log': | ||
| x_stickies = x_stickies[x_stickies > 0] | ||
| if self.get_yscale().lower() == 'log': | ||
| y_stickies = y_stickies[y_stickies > 0] | ||
| else: # Small optimization. | ||
| x_stickies, y_stickies = [], [] | ||
| x_stickies = y_stickies = np.array([]) |
There was a problem hiding this comment.
Is it a problem that these are now the same object? I assume we do not mutate these values below?
There was a problem hiding this comment.
Technically, this is probably working. However, I'd prefer
x_stickies = np.array([])
y_stickies = np.array([])
First, we don't know if these might be written to for some reason in the future. Second, it keeps the symmetry.
There was a problem hiding this comment.
I moved the conditional higher to avoid this need.
cce31fa to
0a035aa
Compare
|
I also reverted the sole image change by instead adding a check that the "correct limits" are applied, and then forcing the limits to the old values. |
This is implemented by altering the semantics of sticky edges as
documented in the changelog. As it turns out, this change also improves
consistency for streamplot():
- in test_streamplot.py::test_{linewidth,mask_and_nans}, the change
is necessary because the axes limits would now be (-3, 3) (matching
the vector field limits), whereas they were previously
xlim=(-3.0, 2.9999999999999947), ylim=(-3.0000000000000004, 2.9999999999999947)
(they can be inspected with
`plt.gcf().canvas.draw(); print(plt.gca().get_xlim(), plt.gca().get_ylim())`).
- in test_streamplot.py::test_maxlength, note that the previous version
expanded the axes limits *beyond* (-3, 3), whereas the current doesn't
do so anymore. It doesn't actually make much sense if the vector
field limits are applied if the streamplot goes all the way to the
edges, but are ignored otherwise.
This is implemented by altering the semantics of sticky edges as
documented in the changelog. As it turns out, this change also improves
consistency for streamplot():
in test_streamplot.py::test_{linewidth,mask_and_nans}, the small bump
in tolerance is necessary because the axes limits are now (-3, 3)
(matching the vector field limits), whereas they were previously
xlim=(-3.0, 2.9999999999999947), ylim=(-3.0000000000000004, 2.9999999999999947)
(they can be inspected with
plt.gcf().canvas.draw(); print(plt.gca().get_xlim(), plt.gca().get_ylim())).in test_streamplot.py::test_maxlength, note that the previous version
expanded the axes limits beyond (-3, 3), whereas the current doesn't
do so anymore. It doesn't actually make much sense if the vector
field limits are applied if the streamplot goes all the way to the
edges, but are ignored otherwise.
Closes #13292.
PR Summary
PR Checklist