segmentation fault with tensorflow 0.8.0 and matplotlib

[Daniel451]

When I build a simple script:

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

plt.plot(np.arange(100))
plt.show()

I get a segmentation fault with TensorFlow 0.8.0. Tested with both, Matplotlib 1.5.1 and 1.4.3.

When I downgrade TensorFlow to 0.7.1 everything runs fine, with both, Matplotlib 1.5.1 and 1.4.3.

TensorFlow is installed via pip inside a virtual environment. I've not installed it outside of a virtual-env, because I do not want to break anything on my system.

Environment info

Operating System: Manjaro Linux

uname --all
Linux 4.4.6-1-MANJARO #1 SMP PREEMPT Sat Mar 19 06:00:37 CET 2016 x86_64 GNU/Linux

pip freeze
backports.shutil-get-terminal-size==1.0.0
cycler==0.10.0
decorator==4.0.9
funcsigs==1.0.1
ipython==4.2.0
ipython-genutils==0.1.0
louis==2.6.5
matplotlib==1.5.1
mock==2.0.0
netsnmp-python==1.0a1
nose==1.3.7
numpy==1.11.0
packaging==16.6
pathlib2==2.1.0
pbr==1.9.1
pexpect==4.0.1
pickleshare==0.7.2
protobuf==3.0.0b2
ptyprocess==0.5.1
pwquality==1.3.0
pyparsing==2.1.1
python-dateutil==2.5.3
python-libtorrent==1.0.9
python-musicbrainz==0.0.0
pytz==2016.4
simplegeneric==0.8.1
six==1.10.0
team==1.0
tensorflow==0.7.1
traitlets==4.2.1
virtualenv==15.0.0

[yaroslavvb]

What if you switch the order and do

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

Seems related to #2034

[Daniel451]

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

plt.plot(np.arange(100))
plt.show()

Produces a segmentation fault, too.

Surprisingly:

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

plt.plot(np.arange(100))
plt.show()

runs fine, no issues at all.

What kind of dark magic is this?^^
Does TensorFlow actually need dependencies to be imported before TensorFlow is imported itself?
Why isn't this happening in TensorFlow 0.7.1?

[lizhitwo]

Not sure if this is of any help, but I ran into the same apparent problem -- importing tf and matplotlib.pyplot gives a segfault, but for me the import order does not matter. I resolved my issue by installing the tensorflow and matplotlib inside a virtual environment without the --system-site-packages flag (so everything is installed anew).

I am using Archlinux w/ kernel 4.5.0-1.

[Daniel451]

Mysterious...
Can someone confirm this issue? Does it only happen on some systems?

[yaroslavvb]

Does it still happen at head? (a similar-sounding segfault issue was fixed in be092f4)

[zsxgb]

I have the same problem. And that's how it is solved.
Before import tensorflow .I import numpy. It have no problem. And Why This？？？

[girving]

Closing as a duplicate of #2034. This is fixed in 0.9.