CuDNN Image Filtering via PyTorch

A tutorial for basic spatial filtering of imagery on the GPU using PyTorch. This is an easy way to complement and accelerate traditional numpy/scipy/OpenCV image processing or image synthesis workflows.

If you want to skip to the code, try running the associated notebook in Google Colab! It's free and let's you skip the installation steps below, and in a reproducable, GPU accelerated runtime environment. Click the badge!

Installation

Requirements

This project requires the following for GPU support:

• Python3.6 or greater
• An NVIDIA graphics card that supports CUDA 10 and is greater than ... in compute power (link to table of existing cards to aid user in checking)
• The ability to install necessary CUDA toolkits, libraries, and SDKs as needed by PyTorch

Install CUDA/CUDA Toolkit, and CuDNN

https://developer.nvidia.com/cuda-toolkit

https://developer.nvidia.com/cudnn

The versions you install will inform which PyTorch local installation you should go with.

https://pytorch.org/

Project installation and Virtual Environment Setup

On Linux

# Navigate to desired project location
cd my/project/directory

# Checkout project
git checkout https://github.com/RyanHartzell/cudnn-image-filtering.git
cd cudnn-image-filtering

# Set up Python virtual environment assuming Python 3.6 or greater (X below refers to your major revision of Python)
python3.X -m venv venv_gpufiltering
source venv_gpufiltering/bin/activate
pip install -r requirements.txt

On Windows

# Navigate to desired project location
chdir my/project/directory

# Checkout project
git checkout https://github.com/RyanHartzell/cudnn-image-filtering.git
chdir cudnn-image-filtering

# Set up Python virtual environment assuming Python 3.6 or greater (X below refers to your major revision of Python)
python3.X -m venv venv_gpufiltering
venv_gpufiltering/Scripts/activate.bat
pip install -r requirements.txt

Execution

Run the Jupyter Notebook cell by cell from Jupyter!

jupyter notebook "notebooks/CuDNN Image Filtering Tutorial Using PyTorch.ipynb"

Or try running the notebook in a free, GPU accelerated Google Colab runtime as noted above!