Ubuntu21. Install TensorFlow and configure GPU support under 10 (cuda11.1+cudnn8.0.5)

Posted by webent on Sat, 29 Jan 2022 19:04:31 +0100

1, Foreword

Installing TensorFlow under Ubuntu is not a simple installation using pip. If you can't use GPU, the efficiency will be very low. The methods here are found out by stepping on some pits under the system. I hope they will be helpful.

  • System: Ubuntu 21 ten
  • CUDA version: cuda11 one
  • Cudnn version: cudnn8 two point one

2, Install NVIDIA driver

1. Add PPA source

Add PPA source:

sudo add-apt-repository ppa:graphics-drivers/ppa

Install dependencies for the system to build kernel modules:

sudo apt-get install dkms build-essential

2. Check the hardware model of the graphics card

Open the terminal and enter the following command

ubuntu-drivers devices

The output interface is as follows. You can see that my hardware graphics card model here is GP108M [GeForce MX250], and the recommended driver version number is: Driver: nvidia-driver-470 - distro non free recommended

3. Install NVIDIA driver

Here we can download the driver version recommended by our computer:

sudo ubuntu-drivers autoinstall

The system will automatically install the recommended version driver, and then restart the system.

3, Installing cuda

The cuda version I use here is 11.1 run file installation, the latest version 11.6 tried to have some problems, so I changed to version 11.1.

1. Download related dependencies

The command is as follows:

sudo apt-get install freeglut3 freeglut3-dev libxi-dev libxmu-dev

2. Download the installation file

Just use wget to download directly. Here I provide the download link of version 11.1:

cd ~/Downloads
wget https://developer.download.nvidia.com/compute/cuda/11.1.0/local_installers/cuda_11.1.0_455.23.05_linux.run

There are probably more than three gigabytes. Pay attention to the network connection.

3. Install CUDA Toolkit

Here we only use the installation file to install CUDA toolkit. The driver installation has been completed in the above operations, so the command is as follows:

sudo sh ./cuda_11.1.0*.run --toolkit --silent --override

4. Configure environment variables

Modify ~ / bashrc file:

vim ~/.bashrc

Add the following:

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda/lib64
export PATH=$PATH:/usr/local/cuda/bin
export CUDA_HOME=$CUDA_HOME:/usr/local/cuda

Make changes effective immediately:

source ~/.bashrc

5. Inspection and installation

The command is as follows:

nvcc --version

4, Install cudnn

1. Download the installation file

Here I download version 8.2.1. The installation of cudnn needs to correspond to your cuda version. The download link on the official website is as follows: https://developer.nvidia.com/rdp/cudnn-archive , you may need to register to download (domestic email can register).

2. Install cudnn8 two point one

Enter the download directory:

cd ~/Downloads

Unzip tgz package:

mkdir cudnn-11.1
tar -zxf cudnn-11.1*.tgz -C ./cudnn-11.1

Enter the extracted file directory:

cd cudnn-11.1

Install the executable and grant permissions:

sudo cp cuda/include/cudnn.h /usr/local/cuda/include
sudo cp cuda/lib64/libcudnn* /usr/local/cuda/lib64
sudo chmod a+r /usr/local/cuda/include/cudnn.h /usr/local/cuda/lib64/libcudnn*

3. Configure environment variables

Edit ~ / bashrc file:

vim ~/.bashrc

Add the following:

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

source:

source ~/.bashrc

5, Install TensorFlow

1. Install dependent packages

Before installing TensorFlow, we need to install a dependency package. Here, my CUDA version is 11.1 and the download dependency package is libcudnn8_8.0.5.39-1+cuda11.1_amd64.deb, the official website link is as follows: https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/

Here I use wget to download:

cd ~/Downloads
wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/libcudnn8_8.0.5.39-1+cuda11.1_amd64.deb

The installation commands are as follows:

sudo dpkg -i libcudnn8_8.0.5.39-1+cuda11.1_amd64.deb

2. Install TensorFlow

You can install it directly with pip. Here, Python version is required to be 3.6-3.9, pip > = 19.0. First upgrade your pip tool with the following command:

python3 -m pip install --upgrade pip

Install TensorFlow using pip. Here I choose Alibaba cloud:

pip install -i https://mirrors.aliyun.com/pypi/simple tensorflow

3. Test

Open Python and test with the following command:

import tensorflow as tf
print(tf.__version__)

The output is as follows:

Next, test our GPU support:

tf.config.list_physical_devices("GPU")

be accomplished!

6, Summary

Topics: Python Ubuntu TensorFlow Deep Learning