Use GPU installation of tensorflow/cuda in spyder under ubuntu 14.04

1
votes

I am running ubuntu 14.04 with an anaconda2 installation and would like to use tensorflow in combination with CUDA. So far the steps I performed are:

  1. Installed CUDA 7.5 and cudnn
  2. Installed tensorflow (GPU version) through a DEB package. Note that I don't want to use the conda package of tensorflow since that one is not the GPU version.
  3. Added Anaconda, CUDA and cudnn to path.
  4. Created a conda environment for tensorflow (conda create -n tensorflow python=2.7)

Now if I start python or IDLE from the terminal, I can import tensorflow and it will find all the CUDA dependencies, great!

...however, if I start ipython or spyder from the same terminal, running "import tensorflow as tf" gives me a cold-hearted "ImportError: No module named tensorflow".

My question: How can I get ipython and spyder to find the tensorflow library just like in an IDLE and a python instance?

1
pythonanacondatensorflowspyder
Are you sure that the they use the same python environment? – Bonifacio2
Spyder is pointing to ~/anaconda2/bin/python as its python executable. My tensorflow is installed in ~/anaconda2/envs/tensorflow/bin. Seems like it's the same environment? Or am I missing something? – Mustafa Radha
Update. Ok, I think I got it. So apparently within the tensorflow environment I was able to install ipython again and now it does recognize tensorflow. Same trick did not solve it for spyder though. – Mustafa Radha
Do you have spyder and ipython installed in your tensorflow environment? What is the output of source activate tensorflow && conda list? – Eric Dill

1 Answers

0
votes

To solve your issue you have 3 options here:

1 just start spyder from terminal

2 move PATH variable definition from .bash_profile to session init scripts

3 Duplicate your PATH in spyder's run configuration