Enable GPU acceleration on Win10 with Notebook

Summary: Records about enabling GPU acceleration on Windows 10 with Jupyter Notebook.

Contents:

• 1. Introduction and basics
• 2. Installing Jupyter lab and setting default directory
• 3. Installing Cuda and Cudnn
  • 3.1 Installing Cuda
  • 3.2 Installing Cudnn
  • 3.3 Adding directory to environment path
  • 3.4 Check if it’s installed successfully
• 4. Creating tensorflow-gpu virtual environment and testing
• 5. Creating pytorch-gpu virtual environment and testing
• 6. Adding virtual env to notebook
• 7. NVDashboard
• Reference:

1. Introduction and basics

Both Tensorflow and PyTorch provide GPU acceleration option, the Google Colab and Kaggle Notebook also provide GPU/TPU acceleration options.

This blog is about my recent configure about GPU acceleration on Windows 10 desktop.

Suppose Anaconda has been intalled.

The common commands about creating virtual environment is following:

#the below commands run in anaconda prompt

# list virtual environment
conda env list

# create new virtual env, 
conda create -n env_names package_names   # conda create -n tfgpu python=3.8

# activate virtual env
conda activate tfgpu

# deactivate current environment
conda deactivate

# delete virtual env
conda env remove -n env_name

Hardware:

• Nvidia GeForce RTX 2060 super
• Windows 10

2. Installing Jupyter lab and setting default directory

• Installing Jupyter lab.

  pip install jupyterlab

• Generating config file jupyter_notebook_config.py:

  jupyter notebook --generate-config

  on Window, the file will be in:

  C:\Users\username\.jupyter/jupyter_notebook_config.py

  on Ubuntu, the file will be in:

  ~/.jupyter/jupyter_notebook_config.py

• Setting default start directory

  on Windows:

  c.NotebookApp.notebook_dir = u'D:\\OneDriveFolder\\Notebook_OneDrive'
  

  on Ubuntu:

  c.NotebookApp.notebook_dir = '/home/ycheng/Documents/Notebook'
  

3. Installing Cuda and Cudnn

3.1 Installing Cuda

Download Cuda.

Version: cuda_11.5.1_496.13_windows.

Install it with default settings.

3.2 Installing Cudnn

Login or Create account and download Cudnn

Version: cudnn-windows-x86_64-8.3.1.22_cuda11.5-archive.

Unzip the file and place the file to C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.5.

3.3 Adding directory to environment path

When installing Cuda, some directory will be added to path automatically, some are not.

C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.5
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.5\libnvvp
C:\Program Files\NVIDIA Corporation\Nsight Compute 2021.3.1\
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.5\extras\CUPTI\lib64
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.5\cudnn-windows-x86_64-8.3.1.22_cuda11.5-archive\bin

3.4 Check if it’s installed successfully

Open command window and type: nvcc --V

The above figure means Cuda and Cudnn have been installed successfully.

4. Creating tensorflow-gpu virtual environment and testing

• Creating virtual env:

  conda create -n tfgpu python=3.8

• Activate tfgpu

  conda activate tfgpu

• Installing tensorflow-gpu

  pip install tensorflow-gpu

• Activating python and testing

  import tensorflow as tf
  tf.config.list_physical_devices('GPU')
  

  

5. Creating pytorch-gpu virtual environment and testing

• Creating virtual env:

  conda create -n pytorch_gpu python=3.8

• Activate tfgpu

  conda activate pytorch_gpu

• Installing pytorch, torchvision, torchaudio, cudatoolkit

  conda install pytorch torchvision torchaudio cudatoolkit=10.2 -c pytorch

  

  Note that the cuda version 10.2 was selected, because cuda 11.5 is not available, which could be a problem.

• Activating python and testing

  import torch
  print(torch.cuda.is_available())
  

  

6. Adding virtual env to notebook

• Activate tfgpu environment

  conda activate tfgpu

• Installing ipykernel

  pip install --user ipykernel

• Adding virtual env to notebook

  python -m ipykernel install --user --name=tfgpu
  #same for pytorch_gpu
  python -m ipykernel install --user --name=pytorch_gpu
  

  Now the new virtual env will be available in notebook

  

  Note: If you forgot to activating new env, and adding the new env to jupyter kernel, you could meet the error:

  you can import new installed libraries for the new env in shell, but you can import it in notebook

  Run

  import sys
  sys.executable
  

  in shell and notebook to check that the path are same, if not, please activate the new env and adding the new env to jupyter kernel.

• List the virtual env in notebook

  jupyter kernelspec list

  Note: since jupyter hasn’t been installed in new virtual env, I run above command under base environment.

  

• Removing virutal env from notebook

  jupyter kernelspec uninstall env_name

7. NVDashboard

https://developer.nvidia.com/blog/gpu-dashboards-in-jupyter-lab/

Reference:

1. https://zhuanlan.zhihu.com/p/420694711
2. https://www.jianshu.com/p/ed57d1dbf3f9