vin-nag/XAI_GAN

This work was accepted at the AAAI-21 Explainable Agency Workshop. It implements a new class of GAN that leverages recent advances in explainable AI (xAI) systems to make training more data-efficient.

XAI-GAN

Table of Contents

• Introduction
• Usage
• Report

Introduction

Generative Adversarial Networks (GANs) are a revolutionary class of Deep Neural Networks (DNNs) that have been successfully used to generate realistic images, music, text, and other data. However, GAN training presents many challenges, notably it can be very resource-intensive. A potential weakness in GANs is that it requires a lot of data for successful training and data collection can be an expensive process. Typically, discriminator DNNs provide only one value (loss) of corrective feedback to generator DNNs (namely, the discriminator's assessment of the generated example). By contrast, we propose a new class of GAN we refer to as xAI-GAN that leverages recent advances in explainable AI (xAI) systems to provide a "richer" form of corrective feedback from discriminators to generators.

Source Code

The code of the project is found in the /src/ directory and run using the main.py file.

Reproduce Results

To run any of the GANs specified (with or without xAI):

1. Go to src/experiment_enums.py and fill out the enums(s) of experiments to run (the file shows what to include)
2. Run python3 main.py
3. The results for each experiment (except for the FID scores) will be found under src/results. In order to evaluate (using FID scores), do the following:
4. For CIFAR10, clone https://github.com/bioinf-jku/TTUR/blob/master/fid_example.py and pass the folder of the generated images.
5. For MNIST or Fashion MNIST, clone https://github.com/mseitzer/pytorch-fid and pass the paths of numpy files generated.
6. In order to run Differentiable Augmentation technique, copy the https://github.com/mit-han-lab/data-efficient-gans/DiffAugment_pytorch.py file and follow the instructions in https://github.com/mit-han-lab/data-efficient-gans.