MD-AOT-GAN

Overview

The Multilayer distilled AOT-GAN (MD-AOT-GAN) model was used for my CSC2541 final project. It is a multilayer distilled GAN based on the SOTA AOT-GAN model. Below is an image of the general overview and how the model is structured.

The resulting model contains about 2x less parameters. Here are the results of the model.

Note that the majority of the AOT_GAN submodule was provided by Zeng et al (https://github.com/researchmm/AOT-GAN-for-Inpainting). There were a few modifications made to

1. Support Quantization (AOT_GAN/src/model/aotgan.py)
2. Support multilayer knowledge distillation (AOT_GAN/src/model/aotgan.py)

Files

train.py

This is the main file which is used for training the MD-AOT-GAN. This performs multilayer knowledge distillation and the parameters for training can be configured here. The main function call is the last function call in the file.

test.py

This file is used to generate samples using your trained model. The train.py script will save the file to desired location. Use this location to load your model and generate samples that will be evaluated on.

test_non_neural.py

This is used to test non-neural inpainting methods like telea and naive stokes. Implementation is provided by OpenCV. This script is much like test.py as it will generate new samples.

eval.py

This script is used with arguments --real_dir /real/dir, --fake_dir /fake/dir, --metric mae psnr ssim fid. This is used to evaluate your generated samples against a ground truth.

quantize.py

This script is our attempt at Eager mode Post training dynamic quantization.