BUG: Contours with LogNorm

Bug report

Bug summary
Contour plots don't interact nicely with matplotlib.colors.LogNorm.
Likely related to #19748

Code for reproduction

import matplotlib.pyplot as plt
from matplotlib.colors import LogNorm
import numpy as np

# BUG: setting levels as a scalar has no effect and no warning or error is printed
fig, ax = plt.subplots()
im = ax.contourf(x, y, data, norm=LogNorm(), levels=4)
fig.colorbar(im, ax=ax)
plt.show()

# kinda buggy too: setting levels as an array does have some of the desired effects
# but ticks and ticklabels are incorrect
fig, ax = plt.subplots()
levels = np.logspace(np.log10(data.min()), np.log10(data.max()), 5)
im = ax.contourf(x, y, data, norm=LogNorm(), levels=levels)
fig.colorbar(im, ax=ax)
plt.show()

Expected outcome

The experience should be comparable with what happens when the norm isn't specified: the levels kwarg should be usable in both cases.
I note that the output from the first snippet is consistent with the case where neither levels or norm are specified, and is correct in that case, so it's definetely not completely broken and I hope the fix is somewhat straightforward.

I'm happy to take a look at the source and see if I can come up with an easy patch, but I'm not sure when I'll have time to dive in there myself. Any piece advice from maintainers or contributors is more than welcome !

Matplotlib version

Operating system: MacOS
Matplotlib version (import matplotlib; print(matplotlib.__version__)): 3.4.1
Matplotlib backend (print(matplotlib.get_backend())): module://ipykernel.pylab.backend_inline
Python version: 3.9.1
Jupyter version (if applicable): NA
Other libraries:

I installed matplotlib using pip.

I'm not sure what you feel the bug is. It would be better if you provided some data. Thanks!

Oh I see you specified the bug in your code.

I believe we use a LogLocator to determine the levels if you use a log norm. It's possible/likely this doesn't interact well with the nlevels argument. But it does provide equal decade levels, which is probably worth it.

I am still not clear what you feel is wrong with the second plot because you didn't specify what levels you set.

I am still not clear what you feel is wrong with the second plot because you didn't specify what levels you set.

the levels are defined as an array, supposedly I'm defining 4 (= n-1, with n=5) of them

levels = np.logspace(np.log10(data.min()), np.log10(data.max()), 5)

I think the plot itself is what I want, but the colorbar is broken in that I only get on tick label, and the minor ticks don't make any sense to me.

What do you think should happen? I guess the other labels not being labeled is bad, but otherwise the ticking seems ok to me. You have put five log-spaced levels with an arbitrary starting point. The minor ticks are just the normal minor ticks.

Note that much of this will change in 3.5 with the colorbar axis overhaul; colorbar axes will behave like normal axes as far as possible.

What do you think should happen? I guess the other labels not being labeled is bad, but otherwise the ticking seems ok to me.

The ticking itself seems off to me because the pattern seen in minor ticks is offset from one level to an other. I'm not sure I'm making myself clear (second language, sorry), so I'll try an alternative phrasing as well:
In logscale, I expect the visual interval between two consecutive minor ticks to be smaller and smaller (in a given decade, or more generally between two major ticks). It's obviously not the case in the first (purple, bottom) level in my second example.

Note that much of this will change in 3.5 with the colorbar axis overhaul; colorbar axes will behave like normal axes as far as possible.

Sounds like fixing this for the 3.4 branch only might be a waste of time then. I'm curious, what's the status of this overhaul ?

The minor ticks are ticking out tenths of decades. Your major ticks are ticking out arbitrary spacing in log space. They are definitely not going to line up.

Oh right, I guess I was expecting the minor ticks to magically retro-engineer my arbitrary spacing and line up somehow (not even sure how actually, there's no obvious generalization of decades that would work with arbitrary spacing)...
I didn't realize how deeply rooted the base 10 log was. Thanks for this !

What I actually want is to have more than one color per decade, maybe there's just something I'm missing here ?

I would just manually set the level, but I'd start with integer decades for the min and max.

So I tried to do just that and, for the specific example I used in the report, I find that this works as intended

levels = np.concatenate(
    [
        10**n * np.arange(1, 10, dtype="float64")
        for n in range(-2, 7)
    ]
)

Thought this doesn't (labels in the colorbar are still missing)

levels = np.concatenate(
    [
        10**n * np.arange(1, 10, dtype="float64")
        for n in range(-2, 6) # <--- changed the upper boundary here
    ]
)

Note that in both cases I'm reaching over the actual maximal value in the data (which is a little under 2e4) by at least one full decade, and I apparently need to add another one to get a sensible result, leading unused colors at the end of the cmap.

Please provide a self-contained runnable script. Thanks!

Here you go !
(I realised while compiling this that I didn't include the data generation in the initial report, sorry !)

import matplotlib.pyplot as plt
import matplotlib.colors import LogNorm
import numpy as np

x, y = np.mgrid[1:10:0.1, 1:10:0.1]
data = np.abs(np.sin(x)*np.exp(y))


# CASE 0,a: leave norm unspecified, pass an int value to the `levels` keyword argument
# RESULT: works fine
fig, ax = plt.subplots()
im = ax.contourf(x, y, data, levels=4)
fig.colorbar(im, ax=ax)
plt.savefig("/tmp/mpl_logcontours_bug0a.png")


# CASE 0,b: specify a log norm, leave levels unspecified
# RESULT: works fine, but only one level is used per decade
fig, ax = plt.subplots()
im = ax.contourf(x, y, data, norm=LogNorm())
fig.colorbar(im, ax=ax)
plt.savefig("/tmp/mpl_logcontours_bug0b.png")


# CASE 1: pass an int value to the `levels` keyword argument
# RESULT: the argument is ignored
fig, ax = plt.subplots()
im = ax.contourf(x, y, data, norm=LogNorm(), levels=4)
fig.colorbar(im, ax=ax)
plt.savefig("/tmp/mpl_logcontours_bug1.png")


# CASE 2: construct arbitrary levels that don't match whole decades
# RESULT: major ticks' labels are not displayed, except for the second one 
fig, ax = plt.subplots()
levels = np.logspace(np.log10(data.min()), np.log10(data.max()), 5)
im = ax.contourf(x, y, data, norm=LogNorm(), levels=levels)
fig.colorbar(im, ax=ax)
plt.savefig("/tmp/mpl_logcontours_bug2.png")


# CASE 3: construct arbitrary levels, matching whole decades
# RESULT: this one works fine, but the max value is more than one decade above the
# actual max value in the data, so the extreme colors in the colormap are not used.
levels = np.concatenate(
    [
        10**n * np.arange(1, 10, dtype="float64")
        for n in range(-2, 7)
    ]
)

fig, ax = plt.subplots()
im = ax.contourf(x, y, data, norm=LogNorm(), levels=levels)
fig.colorbar(im, ax=ax)
plt.savefig("/tmp/mpl_logcontours_bug3.png")


# CASE 4: same as case 3, but max level is closer to global max in the data
# RESULT: colorbar labels are missing (except for the min value, 1e-2)
levels = np.concatenate(
    [
        10**n * np.arange(1, 10, dtype="float64")
        for n in range(-2, 6) # <--- changed the upper boundary here
    ]
)

fig, ax = plt.subplots()
im = ax.contourf(x, y, data, norm=LogNorm(), levels=levels)
fig.colorbar(im, ax=ax)
plt.savefig("/tmp/mpl_logcontours_bug4.png")

Thanks, thats super helpful.

You are kind of reporting a couple of issues here:

levels=4 doesn't work for LogNorm. That seems a bug to me. If you do

im = ax.contourf(x, y, data, norm=LogNorm(), locator=mticker.LogLocator(numticks=4))

it works fine, so I wonder if in def _autolev(self, N): we should be calling the LogLocator with numticks=N? Right now it is called with no arguments.

the labelling doesn't work. I think that will be closed by Enh better colorbar axes #18900, though I note that Enh better colorbar axes #18900 doesn't properly make the formatter scientific, so that seems a bug.

I'll make 1 as a good first issue, though it may have some back-combat issues that will need to be tested.

I didn't realize how deeply rooted the base 10 log was.

You can provide an arbitrary base to LogFormatter (and subclasses):

matplotlib/lib/matplotlib/ticker.py

Lines 879 to 881 in 612cfae

    
           def __init__(self, base=10.0, labelOnlyBase=False, 
        
                        minor_thresholds=None, 
        
                        linthresh=None):

This will (probably) also show the ticks, which seems to be removed by:

matplotlib/lib/matplotlib/ticker.py

Lines 1073 to 1075 in 612cfae

    
           # only label the decades 
        
           fx = math.log(x) / math.log(b) 
        
           is_x_decade = is_close_to_int(fx)

and

matplotlib/lib/matplotlib/ticker.py

Lines 1081 to 1082 in 612cfae

    
           if self.labelOnlyBase and not is_x_decade: 
        
               return ''

Navigation Menu

Repository navigation

BUG: Contours with LogNorm #19856

Bug report

3 remaining items

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

	def __init__(self, base=10.0, labelOnlyBase=False,
	minor_thresholds=None,
	linthresh=None):

	# only label the decades
	fx = math.log(x) / math.log(b)
	is_x_decade = is_close_to_int(fx)

BUG: Contours with LogNorm #19856

Description

Bug report

Activity

jklymak commented on Apr 3, 2021

jklymak commented on Apr 3, 2021

neutrinoceros commented on Apr 3, 2021

jklymak commented on Apr 3, 2021

neutrinoceros commented on Apr 3, 2021

jklymak commented on Apr 3, 2021

neutrinoceros commented on Apr 3, 2021

jklymak commented on Apr 3, 2021

neutrinoceros commented on Apr 4, 2021

jklymak commented on Apr 5, 2021

neutrinoceros commented on Apr 6, 2021

jklymak commented on Apr 6, 2021

jklymak commented on Apr 6, 2021

3 remaining items

oscargus commented on Jan 19, 2022

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions