Repository navigation
Plotting zero values with logarithmic axes triggers OverflowError, Matplotlib hangs permanently #7493
Description
I write an application with a graphical user interface to load and plot data files (https://github.com/FilipDominec/plotcommander). When I simultaneously select the y-axis to be logarithmic and load a file containing zeros, the Ticker class from Matplotlib encounters an error and the plotting window stops working permanently, even if I try to open some "safe" file thereafter. In such a case of error, the whole window of my program has to be closed and launched again, which is obviously not acceptable for a user.
As a mere downstream programmer who only uses the library, I have no clue how to fix this issue myself. I tried to always reset the plotting by self.fig.clf(), but it did not help. I believe that correct catching of the OverflowError would make Matplotlib more robust.
- Matplotlib version 1.5.1-1ubuntu1, Python version2.7.12, Platform Ubuntu 16.04 x64
- same experience under python 3
- I installed Matplotlib and Python from apt repository
Error message:
Traceback (most recent call last):
File "/usr/lib/python3/dist-packages/matplotlib/backends/backend_gtk3cairo.py", line 45, in on_draw_event
self._render_figure(w, h)
File "/usr/lib/python3/dist-packages/matplotlib/backends/backend_gtk3cairo.py", line 34, in _render_figure
self.figure.draw (self._renderer)
File "/usr/lib/python3/dist-packages/matplotlib/artist.py", line 61, in draw_wrapper
draw(artist, renderer, *args, **kwargs)
File "/usr/lib/python3/dist-packages/matplotlib/figure.py", line 1091, in draw
self.tight_layout(renderer, **self._tight_parameters)
File "/usr/lib/python3/dist-packages/matplotlib/figure.py", line 1754, in tight_layout
rect=rect)
File "/usr/lib/python3/dist-packages/matplotlib/tight_layout.py", line 349, in get_tight_layout_figure
pad=pad, h_pad=h_pad, w_pad=w_pad)
File "/usr/lib/python3/dist-packages/matplotlib/tight_layout.py", line 126, in auto_adjust_subplotpars
tight_bbox_raw = union([ax.get_tightbbox(renderer) for ax in subplots])
File "/usr/lib/python3/dist-packages/matplotlib/tight_layout.py", line 126, in <listcomp>
tight_bbox_raw = union([ax.get_tightbbox(renderer) for ax in subplots])
File "/usr/lib/python3/dist-packages/matplotlib/axes/_base.py", line 3683, in get_tightbbox
bb_yaxis = self.yaxis.get_tightbbox(renderer)
File "/usr/lib/python3/dist-packages/matplotlib/axis.py", line 1073, in get_tightbbox
ticks_to_draw = self._update_ticks(renderer)
File "/usr/lib/python3/dist-packages/matplotlib/axis.py", line 949, in _update_ticks
tick_tups = [t for t in self.iter_ticks()]
File "/usr/lib/python3/dist-packages/matplotlib/axis.py", line 949, in <listcomp>
tick_tups = [t for t in self.iter_ticks()]
File "/usr/lib/python3/dist-packages/matplotlib/axis.py", line 892, in iter_ticks
majorLocs = self.major.locator()
File "/usr/lib/python3/dist-packages/matplotlib/ticker.py", line 1538, in __call__
return self.tick_values(vmin, vmax)
File "/usr/lib/python3/dist-packages/matplotlib/ticker.py", line 1565, in tick_values
numdec = math.floor(vmax) - math.ceil(vmin)
OverflowError: cannot convert float infinity to integer
Activity
[-]Logarithmic axes with zero values should catch OverflowError[/-][+]Logarithmic axes with zero values trigger OverflowError, Matplotlib hangs permanently[/+]tacaswell commentedon Nov 21, 2016
Can you provide a minimal code snippet to reproduce this?
FilipDominec commentedon Nov 21, 2016
I surely would if it was easy to reproduce in some script. But it occurs in GUI and so far I am not sure that it is fully deterministic. I will try to find a minimum example and post it here, though.
[-]Logarithmic axes with zero values trigger OverflowError, Matplotlib hangs permanently[/-][+]Plotting zero values with logarithmic axes triggers OverflowError, Matplotlib hangs permanently[/+]FilipDominec commentedon Nov 21, 2016
I think I have traced the source of error. I plotted a line using
ax.set_yscale('log'), and then I switched to a contour plot, keeping the y-axis in the logarithmic mode. Then, the bottom border of the contour plot would be positioned at log(0), which is obviously a nonsense.So the error is on my side. Still I believe matplotlib should be more robust and recover from my impossible demand, so that one can switch back to logarithmic lines. Currently, it just keeps printing an error message over and over, without plotting even a line.
tacaswell commentedon Nov 21, 2016
mpl should raise when you set the invalid state, not at draw time (because drawing tends to happen in GUI callback threads which are hard to debug / catch errors from).
This is hitting an API design edge, user facing / interface code should never raise, libraries should raise early and often. These two uses cases historically been mixed in mpl, but we should be moving more towards the 'raise early and often' side of things and leave the exception handling to your application.
anntzer commentedon Nov 21, 2016
Currently you just have to catch the error yourself in the GUI thread (which as @tacaswell says is not optimal, but not so easy to fix...), see #6983 for how it's done by mpl itself.
FilipDominec commentedon Nov 22, 2016
I must admit that I am not able to catch the error myself. It happens in the library and does not write out the offending line of my code.
Naturally, I have had the
try .. exceptclause around the plotting commands, but the error handler is never executed and matplotlib keeps "freezing" as I described above. Then the program ceases to plot anything, no matter that I return the settings to original safe values, and has to be restarted.Without running a debugger and plunging deep into the matplotlib code, I have probably no means of tracing the source of error, nor even preventing it from happening.
anntzer commentedon Nov 22, 2016
Ah, now I remember I did write something about that. Can you try with #7197?
FilipDominec commentedon Nov 22, 2016
@anntzer, to test your fix, I have run
The error message is different from that I got previously, but occurs at the same situation and causes the same troublesome behaviour. (Note that I also diffed the following error message against that one from the latest commit of the matplotlib trunk - and both messages are almost the same except for the matplotlib revision numbers
+2909.gd36e499!=+2910.g1bb648c.)Maybe I just did something wrong, I will be thankful for some tips for testing it further.
anntzer commentedon Nov 22, 2016
You probably didn't, something else is going on I guess...
temporary fix for the log-contour error matplotlib/matplotlib#7493
jfmoulin commentedon Dec 8, 2016
I have the same problem in a script which could be reduced to
(This crashes IPython beyond repair by the way, with errors coming up in an apparently random fashion from that point.
FilipDominec commentedon Dec 8, 2016
@jfmoulin In my case (see configuration in the first post), I could not reproduce the crashing in Python nor IPython, version 2 nor 3. Instead, matplotlib always correctly reports the error.
efiring commentedon Dec 8, 2016
I'm working on this and related problems. I hope to have a PR within a day or two.
BUG: fix minpos handling and other log ticker problems
BUG: fix minpos handling and other log ticker problems
BUG: fix minpos handling and other log ticker problems
QuLogic commentedon Dec 13, 2016
Closed by #7598.