VQ-VAE project: re-implementation and results' confirmation

Objectifs

The objective is to explore the paper called Neural Discrete Representation Learning by Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu published in 30 May 2018 arxiv link.

More preciselly, the objective is to understand in depht the proposed method, to re-implement using newer version of python and more reable code to be able to reproduce results and confirm (or not) the results shown in the paper.

Structures

The structure of this project is the following one:

• Folder Models contains the VQ-VAEs model for MNIST and CIFAR10 dataset and the PixelCNN models for generating new samples.
• VQ_VAE.py is the file containing the main training algorithm and testing for VQ-VAE models.
• generation.py is the dile containing the training and testing for generating new samples.
• utils.py is a python file full of usefull general funtctions.
• It is assumed to have a folder Datasets containing MNIST and CIFA10 dataset download via torchvision.datasets.
• It is assumed to have a folder saves containing saves of the different models.