Make Tensorflow use the GPU

Hi, Today I was using Keras library with Tensorflow back end, I installed Tensorflow using this documentation (don't follow it the correct method is bellow) and I made sure to install the GPU version but when I was running my code it seemed that Tensorflow is using the CPU.

After few minutes of Googling things up, I find a way to see if Tensorflow is using the CPU or the GPU by running this code in python :

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

In my case I got an error :

tensorflow.python.framework.errors_impl.InvalidArgumentError: Cannot assign a device for operation 'MatMul_1': Operation was explicitly assigned to /device:GPU:0 but available devices are [ /job:localhost/replica:0/task:0/cpu:0 ]. Make sure the device specification refers to a valid device. 

Like the error shows : the GPU wasn't detected, the reason is that my graphics card driver (NVIDIA) wasn't installed, I recently formatted my PC and I am using  Parrot OS as a Linux distribution and didn't need to install the driver before but now I have to.

After using Google again I found the right way to install Tensorflow.

Make sure the NVIDIA Compute Capability of your GPU is more than 3.0 by checking this link :
https://developer.nvidia.com/cuda-gpus
Otherwise, Tensorflow won't work on your GPU.

First, let's install the NVIDIA driver.

I found this link to install the NVIDIA GPU driver it was for Kali Linux but it works with my distribution too, they are all based on Debian.

Using this command I was able to see my card NVIDIA in the list :

[hosni@parrot]─[~]$ sudo lspci -v

and to install the driver and the CUDA toolkit .. I used this command :

[hosni@parrot]─[~]$ sudo apt install -y ocl-icd-libopencl1 nvidia-driver nvidia-cuda-toolkit

We must restart the PC afterward.

Now let's make sure the driver is installed, we type the following command : 

[hosni@parrot]─[~]$ sudo nvidia-smi

We have to install Cuda Toolkit and cuDNN. For Cuda toolkit use this link: https://developer.nvidia.com/cuda-downloads

It's a normal deb package installation.

For cuDNN use this link downloads the file (download version 6 for Cuda 8) and use the following commands :

[hosni@parrot]─[~/Downloads]$ tar xvzf cudnn-8.0-linux-x64-v6.0.tgz
[hosni@parrot]─[~/Downloads]$ sudo cp -P cuda/include/cudnn.h /usr/local/cuda/include
[hosni@parrot]─[~/Downloads]$ sudo cp -P cuda/lib64/libcudnn* /usr/local/cuda/lib64
[hosni@parrot]─[~/Downloads]$ sudo chmod a+r /usr/local/cuda/include/cudnn.h /usr/local/cuda/lib64/libcudnn*

The files are big so it may take a while.

Next step, update your bash file : 

[hosni@parrot]─[~]$ geany ~/.bashrc

With the text editor scroll to the bottom and put in these lines:

export LD_LIBRARY_PATH="$LD_LIBRARY_PATH:/usr/local/cuda/lib64:/usr/local/cuda/extras/CUPTI/lib64"

export CUDA_HOME=/usr/local/cuda

Save and close the file and type this command :

[hosni@parrot]─[~]$ source ~/.bashrc

Now the last step is to install Tensorflow using pip or pip3 depends on your python version, so we just type this command :

 [hosni@parrot]─[/home/hosni/IDE/anaconda3]$ pip install --upgrade tensorflow-gpu

If you are using Conda like me make sure to create Tensorflow environment first by using this commands :

[hosni@parrot]─[/home/hosni/IDE/anaconda3]$ conda create -n tensorflow

[hosni@parrot]─[/home/hosni/IDE/anaconda3]$ source activate tensorflow 

And that's it :

Now in your IDE make sure to select the right python interrupter version where Tensorflow is Installed.

For Intelij I needed to add the environment variables again

Everything should work fine, for me after I spent the whole day trying to make this work I got this message :

Ignoring visible GPU device (device: 0, name: GeForce GT 620M, PCI bus id: 0000:01:00.0) with Cuda compute capability 2.1. The minimum required Cuda capability is 3.0.

It seems my graphics card is old and I have to use the CPU 😢 so make sure to Check the NVIDIA Compute Capability of your GPU card first.

9/01/2017