The TensorFlow package contains a helper module called mnist.py. I've verified that the file exists on my machine, and manually added the directory to sys.path. When I try to import the module, python can't seem to find it. I can import tensorflow as tf just fine, but python seems to be looking for the module in the wrong directory although I've explicitly given it the correct location:
>>> import tensorflow.contrib.learn.python.learn.datasets.mnist
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/home/rh/Applications/Anaconda/anaconda3/lib/python3.5/site-packages/tensorflow/contrib/__init__.py", line 25, in <module>
from tensorflow.contrib import cudnn_rnn
File "/home/rh/Applications/Anaconda/anaconda3/lib/python3.5/site-packages/tensorflow/contrib/cudnn_rnn/__init__.py", line 21, in <module>
from tensorflow.contrib.cudnn_rnn.python.ops.cudnn_rnn_ops import CudnnGRU
File "/home/rh/Applications/Anaconda/anaconda3/lib/python3.5/site-packages/tensorflow/contrib/cudnn_rnn/python/ops/cudnn_rnn_ops.py", line 28, in <module>
resource_loader.get_path_to_datafile("_cudnn_rnn_ops.so"))
File "/home/rh/Applications/Anaconda/anaconda3/lib/python3.5/site-packages/tensorflow/contrib/util/loader.py", line 42, in load_op_library
ret = load_library.load_op_library(path)
File "/home/rh/Applications/Anaconda/anaconda3/lib/python3.5/site-packages/tensorflow/python/framework/load_library.py", line 64, in load_op_library
None, None, error_msg, error_code)
tensorflow.python.framework.errors_impl.NotFoundError: /home/rh/Applications/Anaconda/anaconda3/lib/python3.5/site-packages/tensorflow/contrib/cudnn_rnn/python/ops/_cudnn_rnn_ops.so: cannot open shared object file: No such file or directory
Verifying the file
ls -l ~/Applications/Anaconda/anaconda3/lib/python3.5/site-packages/tensorflow/contrib/learn/python/learn/datasets
total 44
-rw-rw-r-- 2 rh rh 4660 Dec 1 12:09 base.py
drwxrwxr-x 3 rh rh 4096 Jan 12 19:57 data
-rw-rw-r-- 2 rh rh 2126 Dec 1 12:09 __init__.py
-rw-rw-r-- 2 rh rh 1334 Dec 1 12:09 load_csv_test.py
-rw-rw-r-- 2 rh rh 8308 Dec 1 12:09 mnist.py
-rw-rw-r-- 2 rh rh 1317 Dec 1 12:09 produce_small_datasets.py
drwxrwxr-x 2 rh rh 4096 Jan 12 19:57 __pycache__
-rw-rw-r-- 2 rh rh 2745 Dec 1 12:09 text_datasets.py
Verifying the path
>>> sys.path
['', '/home/rh/Applications/Anaconda/anaconda3/lib/python35.zip',
'/home/rh/Applications/Anaconda/anaconda3/lib/python3.5',
'/home/rh/Applications/Anaconda/anaconda3/lib/python3.5/plat-linux',
'/home/rh/Applications/Anaconda/anaconda3/lib/python3.5/lib-dynload',
'/home/rh/Applications/Anaconda/anaconda3/lib/python3.5/site-packages',
'/home/rh/Applications/Anaconda/anaconda3/lib/python3.5/site-packages/Sphinx-1.4.6-py3.5.egg',
'/home/rh/Applications/Anaconda/anaconda3/lib/python3.5/site-packages/setuptools-27.2.0-py3.5.egg',
'/home/rh/Applications/Anaconda/anaconda3/lib/python3.5/site-packages/tensorflow/contrib/learn/python/learn/datasets']
Tensorflow Info
$ conda list | grep tensorflow
tensorflow 0.12.0rc0 py35_0 jjhelmus
Only 1 Python Environment
$ conda env list
# conda environments:
#
root * /home/rh/Applications/Anaconda/anaconda3
$ type -p python
/home/rh/Applications/Anaconda/anaconda3/bin/python
My source code comes from a tensorflow tutorial that runs perfectly on OSX, something about the Ubuntu install/environment is wonky. I've been beating on this problem for a while without success. This related question doesn't address the issue either. Any suggestions?
Found a clue in the answer to this related question, and @gunan's comment.
Although the machine in question doesn't have a python-gpu interface, python kept looking for cudnn files. tensorflow 0.12.0rc0 (installed from a private repo) offers gpu support by default, and I couldn't disable it by resetting environment variables.
The solution was to remove tensorflow 0.12.0rc0 and replace it with tensorflow 0.10.0rc0 from the default Anaconda repositories.
$ conda remove tensorflow
$ conda install tensorflow
$ conda list | grep tensorflow
tensorflow 0.10.0rc0 np111py35_0
----------
Several tensorflow functions were deprecated between versions 10 and 12. At present time, version 12 is the recommended tensorflow package for OSX Anaconda Python (installed from private repos with Anaconda pip). Any script that needs to run on both Mac and non-gpu Linux should use the deprecated version 10 functions. Mac users will need to suppress the resulting warnings.