Demo: Segmentation on CPU and GPU

This is a demonstration of the performance benefits of using SciPy, scikit-learn and scikit-image on GPUs (AMD or NVIDIA) using Array API.

The changes to SciPy, scikit-learn, scikit-image and CuPy can be seen here:

• CuPy
• SciPy
• scikit-learn
• scikit-image

Running the demo

First login to Docker Registry

export CR_PAT=YOUR_PERSONAL_ACCESS_TOKEN_FOR_GITHUB
echo $CR_PAT | docker login ghcr.io -u USERNAME --password-stdin

On NVIDIA GPU

docker run --gpus all -it -p 8788:8788 ghcr.io/quansight-labs/array-api-gpu-demo-cuda:latest bash

and then run jupyterlab inside the container:

jupyter lab --ip=0.0.0.0 --port=8788 --allow-root

On AMD GPU

docker run -it --device=/dev/kfd --device=/dev/dri --security-opt seccomp=unconfined --group-add video ghcr.io/quansight-labs/array-api-gpu-demo-rocm:latest bash

and then run jupyterlab inside the container:

jupyter lab --ip=0.0.0.0 --port=8788 --allow-root

Now run the plot_coin_segmentation.ipynb notebook.

Running Segmentation Performance script

Get into the docker container for either of the above mentioned GPU platform and run the following script:

python segmentation_performance.py

This will run the segmentation for various proportions of the greek coins image for cupy and numpy array, i.e. on CPU and GPU.

This will also create a plot of the performance comparison between numpy and cupy, which would look something like:

On NVIDIA GPU

On AMD GPU