Explain2Me-Framework-Example

This repo showcases the Explain Me This: Salience-Based Explainability for Synthetic Face Detection Models research codebase bundled with the tools and best practices developed by the frameworks group.

Getting Started

To use this repo, clone it and run the following commands:

pdm install
pdm run dvc repro

Documentation can be contributed to by editing the nbs/ files then running pdm run nbdev_prepare.

Frameworks Approach

The frameworks team has added the following tools for explainability and reproducability:

  • PDM: for python package management
  • DVC: for pipeline reproduceability
  • nbdev: for code explainability

DVC Pipeline and Experiments

Rather than re-creating an entire DVC Pipeleine, the frameworks team focused in on nbs/noise.ipynb to display the DVC experiments feature to allow code iterataions from the CLI or vscode plugin. To view the full configuration please view dvc.yaml.

Running Experiments

Experiments can be run using the following methods:

Command Line Interface

Run the following command with the -S argument will run the pipeline with the specified parameters.

pdm run dvc exp run -S 'noise.amount=0.01'

VSCode Plugin (Recommended)

The DVC plugin for VSCode can be installed to provide an IDE for running experiments, viewing plots, and changing parameters. Check out this quick tutorial for an overview here.