Ubuntu installation Tensorflow-gpu + Keras

Reprint Please specify:

Look at Daniel's small freshness : http://www.cnblogs.com/luruiyuan/

This article original website : http://www.cnblogs.com/luruiyuan/p/6660142.html

The Ubuntu version I used was 16.04, and using Gnome as the desktop (which doesn't matter) has gone through a lot of twists and turns and finally completed the installation of Keras with TensorFlow as the back end.

Installation of the TENSORFLOW-GPU version:

1. Download CUDA 8.0

Address: Https://developer.nvidia.com/cuda-downloads

Install the following version:

2. Download CuDNN v5 (Here i download is V6, but it turns out that TensorFlow does not directly support V6, unless you build the V6 yourself, otherwise the following tutorial installed TensorFlow can only load v5)

Address: HTTPS://DEVELOPER.NVIDIA.COM/CUDNN

You need to login/register before you can download

3. Install the NVIDIA driver:

First open the terminal, enter the instructions, update the application list

sudo apt-get update

Thanks to the strong support of the Linux community, we can install the driver in a very concise way:

Press the win button to open the menu,

Then enter in the above type to search: Additional drivers Open "Additional drivers--additional Driver" and select the NVIDIA driver that matches your graphics card, and here I am

Using Nvidia binary drivers (375)

Then choose Apply Changes, where the installation may fail, at which point you can continue to choose to apply the changes, install multiple times

When the installation is complete, select restart

4. GCC down version

CUDA8.0 does not support GCC 5.0 and above compilers, and the system comes with 5.4 and later, so we need to reduce the version, or will be in the back error

Enter the following command in terminal to reduce the GCC version to 4.9

sudoApt-getInstallg++-4.9sudoUpdate-alternatives--Install/usr/bin/GCC GCC/usr/bin/GCC-4.9  -sudoUpdate-alternatives--Install/usr/bin/GCC GCC/usr/bin/GCC-5 TensudoUpdate-alternatives--Install/usr/bin/g++ g++/usr/bin/g++-4.9  -sudoUpdate-alternatives--Install/usr/bin/g++ g++/usr/bin/g++-5 TensudoUpdate-alternatives--Install/usr/bin/cc cc/usr/bin/GCC  -sudoUpdate-alternatives--setcc/usr/bin/GCCsudoUpdate-alternatives--Install/usr/bin/c++ C + +/usr/bin/g++ -sudoUpdate-alternatives--set C + +/usr/bin/g++

5. Installing CUDA 8.0

cd/media/your username/toshiba\ ext/alu/cuda/8.0sudo dpkg-i cuda-repo-ubuntu1604-8- 0-rc_8. 0.27-1_amd64? Debsudo apt-get updatesudoinstall cuda?

6. Installing CuDNN

CD  cd/media/your user name/toshiba\ ext/alu/cuda/ # Enter the path to the CuDNN installation file tar xvzf cudnn-8.0- Linux-x64-v6. 0 . tgz # Unzip sudo CP cuda/include/cudnn.h/usr/local/cuda/include # Copy to include sudocp cuda/ lib64/libcudnn*/usr/local/cuda/lib64 # Copy to lib64 in sudochmod a+r/usr/local/cuda/ include/cudnn.h/usr/local/cuda/lib64/libcudnn* # header file copied in

7. Configuring CUDA Environment variables

Turn on GPU support:

According to the official website tutorial

We type the following command in Terminal:

sudo gedit ~/.bash_profile # Open. Bash_profile This is the user's environment variable, not the global

Then, at the end of the open text, add:

Export ld_library_path="$LD _library_path:/usr/local/cuda/lib64:/usr/local/cuda/extras/cupti/lib64  "export cuda_home=/usr/local/cuda

After saving and closing, enter the following command to make the environment variable effective:

SOURCE ~/.bash_profile # makes the changed environment variable effective

After the installation is complete, the driver must be inspected by the following two commands:

1. nvidia's setup interface

Nvidia-settings # Open NVIDIA Setup interface

This command opens the following interface:

2. NVIDIA GPU List

Nvidia-smi

This command generates a list of GPUs in the terminal, for example, I have only one GPU here

Some people online to copy others ' blog, said nvcc-v can verify, after I measured, there is nvcc-v normal output but the driver is still installed failure phenomenon, therefore, the above verification method is not credible .

8. Installing python3.5.2

Because tensorflow1.0 is better for python3 support and currently only supports python3.5.2, we choose Python 3.5.2.

Install using the Linux version of anaconda3-4.2.0, with the following address:

https://repo.continuum.io/archive/.winzip/

After the installation is complete, add the environment variable and set it as the default Python interpreter

First open the file for the environment variable

Gedit ~/.BASHRC

Then add the path to the Anaconda3 at the end of the file

Export path=/home/your path/anaconda3/bin: $PATH

And finally make our changes effective

SOURCE ~/.BASHRC

This way, we enter Python in terminal and the default is open Anaconda3

So we can use the Python3 safely.

9. Installing Keras and TensorFlow

With the above installation process, the default PIP in our system will be the PIP in Anaconda3, so we only need to use PIP to install Keras and TensorFlow to Anaconda.

Execute the following command:

 Install TENSORFLOW-GPU Keras # installs GPU version of TensorFlow and Keras

Once the installation is complete, we can verify the success by using the following command:

" Import Keras "

If you see the following output, it means that the installation was successful

Of course, I installed here CuDNN due to the version is too high, temporarily can not be supported by PIP installation TensorFlow, if changed to CuDNN V5 will be able to support the normal.

I hope this article can help the novice like me.

Resources:

[1]: ubuntu16.04 installation of TensorFlow (GPU acceleration)----Detailed Graphic tutorial

[2]: Ubuntu16.04+cuda8.0+caffe installation Tutorial

Ubuntu installation Tensorflow-gpu + Keras