TensorFlow_0.8.0 somehow breaks PIL ?

[liuyifly06]

Environment info

Operating System:

ubuntu@ip-172-31-32-186:~$ uname -a
Linux ip-172-31-32-186 3.13.0-48-generic #80-Ubuntu SMP Thu Mar 12 11:16:15 UTC 2015 x86_64 x86_64 x86_64 GNU/Linux

TensorFlow version and installation method:

ubuntu@ip-172-31-32-186:~$ python -c "import tensorflow; print(tensorflow.__version__)"
0.8.0

I installed with command:
sudo pip install --upgrade https://storage.googleapis.com/tensorflow/linux/cpu/tensorflow-0.8.0-cp27-none-linux_x86_64.whl

My problem: Tensorflow break the PIL image read

Steps to reproduce

Following command verifies that my PIL works good without importing TensorFlow:

ubuntu@ip-172-31-32-186:~$ python
Python 2.7.6 (default, Jun 22 2015, 17:58:13) 
[GCC 4.8.2] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> from PIL import Image
>>> a = Image.open('bubble-counting/images/original/bubble.jpg')
>>> b = a.getdata()
>>> b[0]
(166, 56, 57)

After importing TensorFlow, same code does not work any more:

ubuntu@ip-172-31-32-186:~$ python
Python 2.7.6 (default, Jun 22 2015, 17:58:13) 
[GCC 4.8.2] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> import tensorflow as tf
>>> from PIL import Image
>>> a = Image.open('bubble-counting/images/original/bubble.jpg')
>>> b = a.getdata()
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/usr/local/lib/python2.7/dist-packages/PIL/Image.py", line 1151, in getdata
    self.load()
  File "/usr/local/lib/python2.7/dist-packages/PIL/ImageFile.py", line 235, in load
    raise_ioerror(e)
  File "/usr/local/lib/python2.7/dist-packages/PIL/ImageFile.py", line 59, in raise_ioerror
    raise IOError(message + " when reading image file")
IOError: broken data stream when reading image file

Other info

After uninstall TensorFlow0.8 and install TensorFlow0.7.1, the problem disappeared!

installation code:
sudo pip install --upgrade https://storage.googleapis.com/tensorflow/linux/cpu/tensorflow-0.7.1-cp27-none-linux_x86_64.whl

>>> PIL.VERSION
'1.1.7'
>>> pandas.__version__
u'0.18.0'
>>> dask.__version__
'0.8.2'
>>> numpy.__version__
'1.11.0'
>>> scipy.__version__
'0.17.0'

My questions:

1. Can you reproduce the problem ?
2. What might be the reason of this problem since I am not familiar with PIL or TensorFlow code ?

If there were any information you need, pls let me know.

[martinwicke]

Can you try importing tensorflow after PIL? There are some shared library issues we've been seeing where the version of libpng and libjpeg tensorflow is compiled with conflicts with other libraries that are linked against a different version (we've seen problems with OpenCV).

[liuyifly06]

@martinwicke Import TensorFlow after PIL works. Thanks for the information.

[martinwicke]

I'm closing this issue as a duplicate of #1924, it seems to be the same problem.