Shared Interest (Article | Demo)

A method to explore model behavior by comparing ground truth regions and model explanations. Check out the demo to see shared interest applied to ImageNet classification and melanoma prediction tasks. Read the article presented at VISxAI 2020.

Getting Started

Before cloning this repo, install git lfs. When you clone the repo, the data files will automatically download.

From the root:

1. conda env create -f environment.yml
2. conda activate shared-interest
3. pip install -e .
4. cd client; npm i; npm run build

The distill article is available at / while the main demo is available at /demo.

Running Locally

To start the server for development, run:

uvicorn backend.server:app --reload

For production, run:

uvicorn backend.server:app

This will run on a single worker, which should be sufficient for this. By default this will run on 127.0.0.1:8000. To change the host or the port, run:

uvicorn backend.server:app --host <host> --port <port>

Creating Data Files

The code in data/ is used to create the data files consumed by Shared Interest. To apply it to your own data, models, and explanation methods, modify data/generate_datasets.py and data/explanation_methods.py.

Once you have created your own data file, you can incorporate it into the interface, by adding it to backend/server/api/main.py and to the case study selection bar in client/src/ts/etc/selectionOptions.ts.