tensorflow 1.4.0 for ubuntu

Posted on

Recently I found that the workstation in my office had a NVIDIA GPU. It is Quodra K620 with 2G memory, so I decided to make it in good use. Here is a note for how to install tensorflow 1.4.0 for ubuntu 16.04 LTS. For tensorflow 1.4.0, I find that CUDA 8.0 and cudnn v6.0 is the most recommended combination.

I maily follow the documents on official Tensorflow website and some NVIDIA documents.

Install CUDA Toolkit 8.0

The official documents are very detailed but have too many details. To sum them up, 3 main steps are involved.

Step1: Pre-installation actions

This is just a pre-check to see whether your computer can install CUDA.

Verify You Have a CUDA-Capable GPU

1
$ lspci | grep -i nvidia

If it returns nothing, we can give up installing it.

Check you have gcc installed

This is for compilation.

1
$ gcc --version

Remember your architecture

x86_64 or others.

1
$ uname -m && cat /etc/*release

Step2: Download .deb file and install

First I choose a proper .deb to download [here].(https://developer.nvidia.com/cuda-80-ga2-download-archive)
cuda_1

Next, we enter the following 3 commands in terminal.

1
2
3
sudo dpkg -i cuda-repo-ubuntu1604-8-0-local-ga2_8.0.61-1_amd64.deb
sudo apt-get update
sudo apt-get install cuda

Step3: Post-installing

Last, change environment variables: add these 2 lines to .bashrc or .zshrc.

1
2
3
4
$ export PATH=/usr/local/cuda-8.0/bin${PATH:+:${PATH}}
$ export LD_LIBRARY_PATH=/usr/local/cuda-8.0/lib64\
                         ${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}

After this, we should check 2 things:

  1. If the driver is properly installed by nvidia-smi

  2. Run some samples in /usr/local/cuda-8.0/samples to check CUDA.

Install cudnn v6.0

The cudnn is the nvidia’s deep learning library. You can find cudnn v6.0 here and download cuDNN v6.0 Library for Linux.

Then unpack and copy some files to /usr/local/cuda directory.

1
2
3
4
5
tar -zxvf cudnn-8.0-linux-x64-v6.0.tgz
sudo cp cuda/include/cudnn.h /usr/local/cuda/include/
sudo cp cuda/lib64/libcudnn* /usr/local/cuda/lib64/ -d 
sudo chmod a+r /usr/local/cuda/include/cudnn.h
sudo chmod a+r /usr/local/cuda/lib64/libcudnn*

Install tensorflow-gpu 1.4.0

This the last step before we can use tensorflow.

  1. You may need $ sudo apt-get install libcupti-dev to install this NVIDIA CUDA Profile Tools Interface.

  2. Use pip3 install tensorflow-gpu to install the tensorflow gpu version.

Check with examples

Finally, we can try following codes:

1
2
3
4
5
6
7
8
# Creates a graph.
a = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[2, 3], name='a')
b = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[3, 2], name='b')
c = tf.matmul(a, b)
# Creates a session with log_device_placement set to True.
sess = tf.Session(config=tf.ConfigProto(log_device_placement=True))
# Runs the op.
print(sess.run(c))