Cuda and Torch installation guide

A setup guide for installing python, nvidia-driver, cuda, cuDNN, and Pytorch from the new installed ubuntu

• Python
• Nvidia driver
• Cuda
  • (Optional) Remove Cuda
  • Install Cuda
  • Install Cudnn
• Pytorch
  • Install PyTorch by pip
  • Check whether PyTorch is installed

Python

Make a hard link to ensure that you use python3 as a default python, and there is no python path problem while running shell script.

sudo rm -rf /usr/bin/python
sudo ln /usr/bin/python3 /usr/bin/python

Get Python-pip

sudo apt-get install python3-pip
python3 -m pip install --upgrade pip

Nvidia Graphics Driver

https://www.geforce.com.tw/drivers
Choose and download the driver which corresponding to your gpu, then install it.
For example:

sudo sh NVIDIA-Linux-x86_64-440.82.run

If you meet Nouveau kernel problem, please type these 2 command to blackout nouveau:

$ sudo bash -c "echo blacklist nouveau > /etc/modprobe.d/blacklist-nvidia-nouveau.conf"
$ sudo bash -c "echo options nouveau modeset=0 >> /etc/modprobe.d/blacklist-nvidia-nouveau.conf"

After you finish the installation process, you can type

nvidia-smi

to check whether you have already installed graphic driver.
You can see the response, like this:

Tue Apr 28 15:24:34 2020       
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 440.82       Driver Version: 440.82       CUDA Version: 10.2     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  GeForce GTX 1080    Off  | 00000000:01:00.0 Off |                  N/A |
| 26%   38C    P0    35W / 180W |      0MiB /  8119MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+
                                                                               
+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID   Type   Process name                             Usage      |
|=============================================================================|
|  No running processes found                                                 |
+-----------------------------------------------------------------------------+

Cuda

(Optional) Remove Cuda

sudo apt-get purge cuda

sudo apt-get purge cuda-XX-YY
sudo apt-get autoremove

For example, if you installed cuda 10.1, then you type

sudo apt-get purge cuda-10-1

maybe there are some other versions of cuda in /usr/local/ you can remove all of it by

sudo rm -r /usr/local/cuda*

Install Cuda

https://developer.nvidia.com/cuda-toolkit-archive
Choose the cuda version that you want to install (say, 10.2), and assume the machine environment is Ubuntu 18.04.
I recommend install cuda by runfile (local) because it has good command-line prompts that can help you to install cuda, and set PATH environment for cuda automatically.
For example:

sudo sh cuda_10.2.89_440.33.01_linux.run

While you are installing cuda, a prompt will be displayed, asking you whether to install "Driver", "CUDA Toolkit", "CUDA Samples",...
The "Driver" here means "Nvidia Graphics Driver"
Since we have already installed Driver in the previous step, do NOT install Driver (i.e., unselect Driver) while running cuda installation runfile.
(I have tried to install Driver via cuda installation runfile but it usually leads to installation problem.)

Add Cuda to Path

Type the command below, which appends two lines of code into ~/.bashrc to get cuda path

echo >> ~/.bashrc '
export PATH=/usr/local/cuda/bin:$PATH
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda/lib64'

then type the command which will source ~/.bashrc

source ~/.bashrc

Check whether Cuda is installed

By just type command below

nvcc -V

You can see the response:

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2019 NVIDIA Corporation
Built on Wed_Oct_23_19:24:38_PDT_2019
Cuda compilation tools, release 10.2, V10.2.89

Install Cudnn

https://developer.nvidia.com/rdp/cudnn-download
After you download all three files, you can install them via:
(You cannot download these files by curl or wget because nvidia needs users to login)

sudo dpkg -i libcudnn7_7.6.5.32-1+cuda10.2_amd64.deb
sudo dpkg -i libcudnn7-dev_7.6.5.32-1+cuda10.2_amd64.deb
sudo dpkg -i libcudnn7-doc_7.6.5.32-1+cuda10.2_amd64.deb # not requirement, doc is for verifying installation

Check whether CuDNN is installed

type in command below

CUDNN_H_PATH=/usr/include/cudnn.h
cat ${CUDNN_H_PATH} | grep CUDNN_MAJOR -A 2

you can see the response:

#define CUDNN_MAJOR 7
#define CUDNN_MINOR 6
#define CUDNN_PATCHLEVEL 5
--
#define CUDNN_VERSION (CUDNN_MAJOR * 1000 + CUDNN_MINOR * 100 + CUDNN_PATCHLEVEL)

#include "driver_types.h"

which means you are using CuDNN 7.6.5

Test Cudnn via mnist training example

If you installed libcudnn-doc, you can exploit mnist examples from doc to verify whether cudnn is installed correctly.

cd /usr/src/cudnn_samples_v7/mnistCUDNN
sudo make clean
sudo make 
./mnistCUDNN

PyTorch

Install PyTorch by pip

pip install torch torchvision

Check whether PyTorch is installed

$ python
Python 3.6.9 (default, Apr 18 2020, 01:56:04) 
[GCC 8.4.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch
>>> torch.cuda.is_available()
True