Install TensorFlow from the source

The installation is available at TensorFlow. Installation from the source is recommended because the user can build the desired TensorFlow binary for the specific architecture. It enriches the TensoFlow with a better system compatibility and it will run much faster. Installing from the source is available at Installing TensorFlow from Sources link. The official TensorFlow explanations are concise and to the point. However. few things might become important as we go through the installation. We try to project the step by step process to avoid any confusion. The following sections must be considered in the written order.

The assumption is that installing TensorFlow in the Ubuntu using GPU support is desired. Python2.7 is chosen for installation.

NOTE Please refer to this youtube link for a visual explanation.

Prepare the environment

The following should be done in order:

TensorFlow Python dependencies installation

Bazel installation

TensorFlow GPU prerequisites setup

TensorFlow Python Dependencies Installation

For installation of the required dependencies, the following command must be executed in the terminal:

sudo apt-get install python-numpy python-dev python-pip python-wheel python-virtualenv
sudo apt-get install python3-numpy python3-dev python3-pip python3-wheel python3-virtualenv

The second line is for python3 installation.

Bazel Installation

Please refer to Bazel Installation.

WARNING: The Bazel installation may change the supported kernel by the GPU! After that you may need to refresh your GPU installation or update it, otherwise, you may get the following error when evaluating the TensorFlow installation:

kernel version X does not match DSO version Y -- cannot find working devices in this configuration

For solving that error you may need to purge all NVIDIA drivers and install or update them again. Please refer to CUDA Installation for further detail.

TensorFlow GPU Prerequisites Setup

The following requirements must be satisfied:

NVIDIA's Cuda Toolkit and its associated drivers(version 8.0 is recommended). The installation is explained at CUDA Installation.

The cuDNN library(version 5.1 is recommended). Please refer to NIDIA documentation for further details.

Installing the libcupti-dev using the following command: sudo apt-get install libcupti-dev

Creating a Virtual Environment (Optional)

Assume the installation of TensorFlow in a python virtual environment is desired. First, we need to create a directory to contain all the environments. It can be done by executing the following in the terminal:

sudo mkdir ~/virtualenvs

Now by using the virtualenv command, the virtual environment can be created:

sudo virtualenv --system-site-packages ~/virtualenvs/tensorflow

Environment Activation

Up to now, the virtual environment named tensorflow has been created. For environment activation, the following must be done:

source ~/virtualenvs/tensorflow/bin/activate

However, the command is too verbose!

Alias

The solution is to use an alias to make life easy! Let's execute the following command:

echo 'alias tensorflow="source $HOME/virtualenvs/tensorflow/bin/activate" ' >> ~/.bash_aliases
bash

After running the previous command, please close and open terminal again. Now by running the following simple script, the tensorflow environment will be activated.

tensorflow

check the ``~/.bash_aliases``

To double check let's check the ~/.bash_aliases from the terminal using the sudo gedit ~/.bash_aliases command. The file should contain the following script:

alias tensorflow="source $HO~/virtualenvs/tensorflow/bin/activate"

check the ``.bashrc``

Also, let's check the .bashrc shell script using the sudo gedit ~/.bashrc command. It should contain the following:

if [ -f ~/.bash_aliases ]; then
. ~/.bash_aliases
fi

Configuration of the Installation

At first, the Tensorflow repository must be cloned:

git clone https://github.com/tensorflow/tensorflow

After preparing the environment, the installation must be configured. The flags of the configuration are of great importance because they determine how well and compatible the TensorFlow will be installed!! At first, we have to go to the TensorFlow root:

cd tensorflow  # cd to the cloned directory

The flags alongside with the configuration environment are demonstrated below:

$ ./configure
Please specify the location of python. [Default is /usr/bin/python]: /usr/bin/python2.7
Please specify optimization flags to use during compilation when bazel option "--config=opt" is specified [Default is -march=native]:
Do you wish to use jemalloc as the malloc implementation? [Y/n] Y
jemalloc enabled
Do you wish to build TensorFlow with Google Cloud Platform support? [y/N] N
No Google Cloud Platform support will be enabled for TensorFlow
Do you wish to build TensorFlow with Hadoop File System support? [y/N] N
No Hadoop File System support will be enabled for TensorFlow
Do you wish to build TensorFlow with the XLA just-in-time compiler (experimental)? [y/N] N
No XLA JIT support will be enabled for TensorFlow
Found possible Python library paths:
  /usr/local/lib/python2.7/dist-packages
  /usr/lib/python2.7/dist-packages
Please input the desired Python library path to use.  Default is [/usr/local/lib/python2.7/dist-packages]
Using python library path: /usr/local/lib/python2.7/dist-packages
Do you wish to build TensorFlow with OpenCL support? [y/N] N
No OpenCL support will be enabled for TensorFlow
Do you wish to build TensorFlow with CUDA support? [y/N] Y
CUDA support will be enabled for TensorFlow
Please specify which gcc should be used by nvcc as the host compiler. [Default is /usr/bin/gcc]:
Please specify the Cuda SDK version you want to use, e.g. 7.0. [Leave empty to use system default]: 8.0
Please specify the location where CUDA 8.0 toolkit is installed. Refer to README.md for more details. [Default is /usr/local/cuda]:
Please specify the cuDNN version you want to use. [Leave empty to use system default]: 5.1.10
Please specify the location where cuDNN 5 library is installed. Refer to README.md for more details. [Default is /usr/local/cuda]:
Please specify a list of comma-separated Cuda compute capabilities you want to build with.
You can find the compute capability of your device at: https://developer.nvidia.com/cuda-gpus.
Please note that each additional compute capability significantly increases your build time and binary size.
[Default is: "3.5,5.2"]: "5.2"

NOTE:

The cuDNN version must be exactly determined using the associated files in /usr/local/cuda
The compute capability is spesified related the available GPU model in the system architecture. For example Geforce GTX Titan X GPUs have compute capability of 5.2.
Using bazel clean is recommended if re-configuration is needed.

WARNING:

In case installation of the TwnsorFlow in a virtual environment is desired, the environment must be activated at first and before running the ./configure script.

Test Bazel (Optional)

We can run tests using Bazel to make sure everything's fine:

./configure
bazel test ...

Build the .whl Package

After configuration of the setup, the pip package needs to be built by the Bazel.

To build a TensorFlow package with GPU support, execute the following command:

bazel build --config=opt --config=cuda //tensorflow/tools/pip_package:build_pip_package

The bazel build command builds a script named build_pip_package. Running the following script build a .whl file within the ~/tensorflow_package directory:

bazel-bin/tensorflow/tools/pip_package/build_pip_package ~/tensorflow_package

Installation of the Pip Package

Two types of installation can be used. The native installation using system root and the virtual environment installation.

Native Installation

The following command will install the pip package created by Bazel build:

sudo pip install ~/tensorflow_package/file_name.whl

Using Virtual Environments

At first, the environment must be activation. Since we already defined the environment alias as tensorflow, by the terminal execution of the simple command of tensorflow, the environment will be activated. Then like the previous part, we execute the following:

pip install ~/tensorflow_package/file_name.whl

WARNING:

By using the virtual environment installation method, the sudo command should not be used anymore because if we use sudo, it points to native system packages and not the one available in the virtual environment.
Since sudo mkdir ~/virtualenvs is used for creating of the virtual environment, using the pip install returns permission error. In this case, the root privilege of the environment directory must be changed using the sudo chmod -R 777 ~/virtualenvs command.

Validate the Installation

In the terminal, the following script must be run (in the home directory) correctly without any error and preferably any warning:

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

Common Errors

Different errors reported blocking the compiling and running TensorFlow.

Mismatch between the supported kernel versions: This error mentioned earlier in this documentation. The naive solution reported being the reinstallation of the CUDA driver.

ImportError: cannot import name pywrap_tensorflow: This error usually occurs when the Python loads the tensorflow libraries from the wrong directory, i.e., not the version installed by the user in the root. The first step is to make sure we are in the system root such that the python libraries are utilized correctly. So basically we can open a new terminal and test TensorFlow installation again.

ImportError: No module named packaging.version": Most likely it might be related to the pip installation. Reinstalling that using python -m pip install -U pip or sudo python -m pip install -U pip may fix it!

Summary

In this tutorial, we described how to install TensorFlow from the source which has the advantage of more compatibility with the system configuration. Python virtual environment installation has been investigated as well to separate the TensorFlow environment from other environments. Conda environments can be used as well as Python virtual environments which will be explained in a separated post. In any case, the TensorFlow installed from the source can be run much faster than the pre-build binary packages provided by the TensorFlow although it adds the complexity to installation process.

Uh oh!