Repository navigation
too large file size created by the errorbar of matplotlib #3345
Description
I was tring to plot a lot of data with errors using errobar of matplotlib. The eps or pdf file is very large, much larger than created by IDL. I have tried to use rasterized=True, but it does not work. If I only use plt.plot(x,y) to plot the data without errors, it can create a much smaller file. I am using OSX 10.9 and Anaconda. Here is the example of my code
import numpy as np
import matplotlib.pyplot as plt
np.random.seed(0)
x = np.linspace(-5., 5., 5000)
y = 3 * np.exp(-0.5 * (x - 1.3)2 / 0.82)
y += np.random.normal(0., 0.2, x.shape)
f = plt.figure(0)
plt.errorbar(x,y, yerr=0.2,fmt='')
f.savefig('test.pdf')
Activity
tacaswell commentedon Aug 6, 2014
what does
matplotlib.__version__give?ellot commentedon Aug 14, 2014
I use matplotlib v1.3.1.
I have tried this in Fedora with matplotlib v1.3.1, it also has the same problem.
Is it fixed in higher version?
jenshnielsen commentedon Aug 14, 2014
Could you please give some numbers for reference?
I would suggest testing the Cairo backend as a work around:
These are the results that I get from testing your plot (MPL 1.4.x branch). I don't have access to IDL so I have no clue about the sizes that it generates.
Another thing is that the errorbars in this plot makes no sense IMHO since they are all on top of each other. I would reduce the number of errorbars with the
erroreveryargument to make them separable. That will also likely reduce the file size.ellot commentedon Aug 14, 2014
The original backend is MacOSX, the file size is about 1.5 MB. If I create an eps file and then use ps2pdf, the pdf file size is about 232 KB.
ellot commentedon Aug 14, 2014
I have tried use cairo backend with OSX 10.9, but I got an error
"Fatal Python error: PyThreadState_Get: no current thread
Abort trap: 6"
jenshnielsen commentedon Aug 14, 2014
How large is the IDL file?
You probably don't have the right dependencies for the cairo backend. You will need either cairocffi or pycairo installed and working but without some more details about how you installed matplotlib and python it is impossible to know. The cairo backend certainly works on OS X 10.9
tacaswell commentedon Aug 14, 2014
@jkseppan I have a hunch that is related to a bug that was reported to me verbally, but hasn't made it into the tracker yet.
When saving collections the pdf backend creates a new nested group (not sure if that is the correct pdf term) for every element instead of putting all of the elements in a single group.
I also suspect that this is related to the some what funny way we use graphic contexts.
ellot commentedon Aug 15, 2014
IDL can not create PDF file directly. It can create an eps file, and then I use ps2pdf to get a PDF file about 180 KB.
ellot commentedon Aug 15, 2014
I use Anaconda, the matplotlib is included in Anaconda.
jenshnielsen commentedon Aug 15, 2014
Thanks
It does't look like conda packs pycairo or cairocffi so that will not work right away.
jkseppan commentedon Aug 18, 2014
It seems that the errorbars get rendered using separate XObjects instead of just outputting the moveto and lineto operations in the page, and that blows up the file size.
mdboom commentedon Aug 25, 2014
@jkseppan: That makes sense. Markers (which include ticks) all use XObjects, which normally results in a reduction in file size when those markers are used frequently. But maybe ticks are simple enough that not using XObjects is actually better.
Use the uses_per_path data in pdf backend
Use the uses_per_path data in pdf backend
jkseppan commentedon Sep 7, 2014
See #3485 for a proposed fix.
jkseppan commentedon Sep 7, 2014
There could be a deeper problem somewhere. Why is draw_path_collection being called with collections of one path each? It seems that there is code in Collection.draw that is intended to call draw_markers for such collections.
Reduce the emitting of object definitions in ps and svg