Table of contents
  1. Install Ubuntu
    1. Create a bootable USB stick.
    2. Install Ubuntu desktop
  2. Set a good app source server
  3. Configure Input Method
  4. Install Chrome
  5. Install PyCharm
  6. Install Git
  7. Install Anaconda
  8. Install CUDA (optional)
  9. Install GPU-driver
    1. Remove the installed driver (if needed)
    2. Add PPA repository
    3. Install driver
  10. Install Pytorch/TensorFlow
    1. Pytorch
    2. TensorFlow
  11. Configure PyCharm
  12. Install TeamViewer (optional)

Introduction

This post documents how I configure a brand new computer into a work station that could be used for deep learning. It was first posted on the ZhiHu in Chinese, which is no longer updated. This page is maintained by me to be up-to-date.

Lastest Update: 27 April 2022.

Install Ubuntu

Create a bootable USB stick.

Requirement: A 8GB or larger USB stick/flash drive

Follow the the offical tutorials for creating a bootable USB stick on Windows, masOS, or Ubuntu.

Install Ubuntu desktop

Follow the offical tutorial. We can start from its step-4 since we have already created a bootable USB stack.

Configuration adivces:

  • Don’t use dual boot, i.e., install two operating systems (e.g., Ubuntu and Windows) in your computer.
  • Install system on SSD drive if possible.
  • “Erase disk and install Ubuntu” is good unless you are famialr with disk partition. Make sure that important data in the disk have been backed up.
  • Use short <user name> and <computer's name> to have a nice prefix in the terminal: image
  • If you accidently configured some wrong/unwanted settings, reinstallation could be one of the simplest solution.
  • Do not waste time on trying to install Windows apps on Ubuntu.
  • Install the version of English.
Got BIOS problems? **F12 not working?** 1. *"F12 is the most common key for bringing up your system’s boot menu, but Escape, F2 and F10 are common alternatives."* 2. *"If you’re unsure, look for a brief message when your system starts – this will often inform you of which key to press to bring up the boot menu."* 3. If there is not brief message guiding you to press which key to get into the system’s boot menu (BIOS), Google the BIOS key of your motherboard. 4. If pressing the right BIOS key (repeatedly/and holding) is still not working, it's probably because the Fast Boot is enable on your computer, which will skip the BIOS access. Google how to disable the Fast Boot for your motherboard. **Cannot boot from USB in BIOS?** Different motherboards have BIOS of different styles. Find and get into the setting of `Boot Menu`, and there should be the name of your USB flash in the list. Select your USB and continue the booting.

Set a good app source server

Choose a good proxy/mirror server could make your apt and snap run much faster if you are outsite United States, like China mainland:

  1. Speed up apt: Software & Updates --> --> Ubuntu Software --> Download from --> Other... --> Select Best Server
  2. Speed up snap (optional):
    sudo snap install snap-store-proxy
sudo snap install snap-store-proxy-client

Configure Input Method

Take the configuration of input source of Chinese as an example.

  1. Settings --> Region&Language --> Manage Installed Language --> Install/Remove Language --> check "Chinese (simplified)" and "Chinese (traditional)" --> Apply Screenshot from 2022-04-27 14-39-44
  2. Re-login or Open a terminal and run ibus restart: Screenshot from 2022-04-27 14-40-57
  3. Settings --> Keyboard --> Input Sources --> Add Input Source... --> Chinese (Intelligent Pinyin) --> Add: Screenshot from 2022-04-27 14-42-04
  4. Configuration completed. Now you can use the shortcut Super + Space to switch between the input sources. The status is shown in the right-up corner： Screenshot from 2022-04-27 14-42-41

Install Chrome

wget https://dl.google.com/linux/direct/google-chrome-stable_current_amd64.deb
sudo dpkg -i google-chrome-stable_current_amd64.deb

Install PyCharm

Search Pycharm in Software/App Center and install the one you like: Screenshot from 2022-04-27 12-17-48

I installed pro because I have free license of student.

External reading: Ubuntu Software vs Software

Install Git

sudo apt install git

Install Anaconda

Here we install the miniconda

wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
sh Miniconda3-latest-Linux-x86_64.sh 
sudo rm Miniconda3-latest-Linux-x86_64.sh
conda update -n base -c defaults conda  # restart the terminal before run this line

Install CUDA (optional)

Nowadays, you don NOT have to install cuda-toolkit or cuDNN because they will come with the installation of pytorch/tensorflow using conda. Therefore, you only need to install a NVIDIA driver. Skip this section if you don’t need cuda.

If you do have to install CUDA, since the software is being updated frequently, here I provide some websites for reference: Compatibility check:

After the confirmation of the CUDA version to install, go to the CUDA archive website and follow the instruction to install the CUDA. Below is an example: image

  • I suggest using the network installation because it always follows the latest guidance from the NVIDIA.
  • Instead of running sudo apt-get -y install cuda, it’s suggested to first running sudo apt-cache search cuda to get list of available version of CUDA, and then run sudo apt-get install cuda-12.1 (for example) to install the specific version.
  • It happens that the command sudo apt-cache search cuda does NOT show the wantted version if you are looking for a older version of CUDA, e.g., CUDA-10.2. In this case, you are suggested to install CUDA in via runfile (local) in the above figure, to download a CUDA-10.2 installer on you computer and then excute it.

Verify the installation of CUDA:

export PATH=$PATH:/usr/local/cuda/bin
nvcc -V

Install GPU-driver

Nowadays, you don NOT have to install cuda-toolkit or cuDNN because they will come with the installation of pytorch/tensorflow using conda. Therefore, you only need to install a NVIDIA driver. Check the compatibility of GPU-Driver (optional).

Remove the installed driver (if needed)

The below commands are dangerous

### remove all cuda/nvidia related softwares, be very careful.
# sudo rm /etc/apt/sources.list.d/cuda*
# sudo apt-get --purge remove "*cublas*" "cuda*" "nsight*" 
# sudo apt-get --purge remove "*nvidia*"
# sudo apt-get autoremove
# sudo apt-get autoclean
# sudo rm -rf /usr/local/cuda*

Add PPA repository

sudo add-apt-repository ppa:graphics-drivers
sudo apt update
ubuntu-drivers devices  # list all effetive drivers
# sudo apt list nvidia-driver*  # (alternative) list all effective drivers

Output (1080ti, 27 April 2022):

Screenshot from 2022-04-27 13-17-11

Install driver

Choose one appropriate version, install it:

sudo apt install nvidia-driver-510

Reboot and check the installation:

reboot
nvidia-smi

Outputs:

Screenshot from 2022-04-27 15-42-16

Note that the CUDA Version: 11.6 in the above picture does NOT mean that a CUDA is installed.

Install Pytorch/TensorFlow

Follow the official installation tutorials of Pytorch and Tensorflow. Alternatively, you may run the following given commands to install pytorch/tensorflow in a new conda env (cuda and cuDNN are all installed automatically).

Pytorch

conda create -n pyt  # replace 'pyt' with name you like
conda activate pyt
conda install pytorch torchvision torchaudio -c pytorch
conda list

TensorFlow

conda create -n tf-gpu
conda activate tf-gpu
conda install tensorflow-gpu 
conda list

Noted that you can specify version that you want, e.g., tensorflow-gpu=2.4.1 cudatoolkit=10.1 and pytorch=1.11.0 cudatoolkit=11.3.

  • Conda virtual enviroment is flexible and simple to use. It can install cuda-toolkit and cuDNN easily. We can configure multiple envs with different versions of cuda/pytorch/tensorflow and switch among them quickly.
  • If you have installed a global cuda in the system, and a different version of the cuda-toolkit in the virtual env, you then might encounter the error of version conflict. You can install pytorch/tensorflow solely by using pip install (instead of conda install). In this case, your virtual env will use the cuda installed in the system.

Configure PyCharm

To use the created conda virtual in Pycharm, you need to configure it as the Python Interpreter of your project. Open the File --> Settings --> Project --> Python Interpreter： Screenshot from 2022-04-27 16-06-48

Click the gear icon at the right-up corner, chose ‘Add’: Screenshot from 2022-04-27 16-11-18

Choose Conda Enviroment --> Exsiting enviroment, then configure the executable python path of your conda env: Screenshot from 2022-04-27 16-23-22

The executable python path will be automatically detected by the Pycharm. In case not, it normally locates at /home/<user_name>/miniconda3/env/<env_name>/bin/python.

Install TeamViewer (optional)

Control your computer remotely:

wget https://download.teamviewer.com/download/linux/teamviewer_amd64.deb
sudo apt install -y ./teamviewer_amd64.deb
sudo rm teamviewer_amd64.deb

TeamViewer didn’t restart with the system even after I checked the box Extras --> Options --> Start TeamViewer with system. The solution I found:

sudo systemctl start teamviewerd.service
sudo systemctl enable teamviewerd.service