HOWTO: Install tensorflow & remote access to jupyter notebook

Assumptions

You have access to a linux server with

1. GPUs,
2. CUDA Toolkit (for details: NVIDIA’s documentation),
3. NVIDIA drivers associated with CUDA Toolkit,
4. cuDNN (for details: NVIDIA’s documentation),
5. python 3.x, and
6. jupyter are installed.

You want to install tensorflow with Anaconda and run a jupyter notebook remotely.

Installation

1. Download Anaconda

   wget http://repo.continuum.io/archive/Anaconda3-4.3.0-Linux-x86_64.sh
   

2. Install Anaconda

   sh Anaconda3-4.3.0-Linux-x86_64.sh
   

3. Create a new virtual environment

   conda create -n ENVIRONMENT-NAME
   

   Replace ENVIRONMENT-NAME with your a name of your choice.

4. Activate the new environment

   source activate ENVIRONMENT-NAME
   

   To see which enviromnet is run the following command and look for *

   conda info --envs
   

5. Set enviroment variables

   export CUDA_HOME=/usr/local/cuda-9.0
   export PATH=$PATH:/usr/local/cuda-9.0/bin:/usr/local/cuda-9.0/lib64
   export LD_LIBRARY_PATH=/usr/local/cuda-9.0/bin:/usr/local/cuda-9.0/extras/CUPTI/lib64
   

   in case of cuda-9.0

6. Install tensorflow-gpu

   pip install --ignore-installed --upgrade tensorflow-gpu  
   

7. OR install tensorflow

   pip install --ignore-installed --upgrade tensorflow  
   

Checking Installation

1. Run nvcc and you should see release 9.0 on the last line

nvcc --version  

2. Run nvidia-smi to see what is happening on the GPU(s)

nvidia-smi

3. Run python

python

and then paste this code (taken from here)

# Python
import tensorflow as tf
hello = tf.constant('Hello, TensorFlow!')
sess = tf.Session()
print(sess.run(hello))

Start a jupyter notebook

1. Here is the script I use to start a jupyter notebook after activating the new environment

   #!/bin/bash
   
   export CUDA_HOME=/usr/local/cuda-9.0
   export PATH=$PATH/usr/local/cuda-9.0/bin:/usr/local/cuda-9.0/lib64
   export LD_LIBRARY_PATH=/usr/local/cuda-9.0/bin:/usr/local/cuda-9.0/extras/CUPTI/lib64
   
   # after setting the environment variables start the notebook
   jupyter-notebook --port=8899 --no-browser
   

   You will see something like:

   http://localhost:8899/?token=YOUR-TOKEN
   

2. Tunnel

   ssh -L 8899:127.0.0.1:8899 your-username@server-address 
   

3. Open a browser and copy paste the address from 1.

Possible error

I ran in a libcublas.so.9.0 error. In my case the problem was that
/usr/local/cuda was pointing at /usr/local/cuda-8.0/

You can check that with

ls -lh /usr/local/cuda

In case your are using my script, be sure to start the script with source. Why? answer