Skip to content
master
Switch branches/tags
Code

Reproducing Neural Discrete Representation Learning

Course Project for IFT 6135 - Representation Learning

Project Report link: final_project.pdf

Instructions

  1. To train the VQVAE with default arguments as discussed in the report, execute:
python vqvae.py --data-folder /tmp/miniimagenet --output-folder models/vqvae
  1. To train the PixelCNN prior on the latents, execute:
python pixelcnn_prior.py --data-folder /tmp/miniimagenet --model models/vqvae --output-folder models/pixelcnn_prior

Datasets Tested

Image

  1. MNIST
  2. FashionMNIST
  3. CIFAR10
  4. Mini-ImageNet

Video

  1. Atari 2600 - Boxing (OpenAI Gym) code

Reconstructions from VQ-VAE

Top 4 rows are Original Images. Bottom 4 rows are Reconstructions.

MNIST

png

Fashion MNIST

png

Class-conditional samples from VQVAE with PixelCNN prior on the latents

MNIST

png

Fashion MNIST

png

Comments

  1. We noticed that implementing our own VectorQuantization PyTorch function speeded-up training of VQ-VAE by nearly 3x. The slower, but simpler code is in this commit.
  2. We added some basic tests for the vector quantization functions (based on pytest). To run these tests
py.test . -vv

Authors

  1. Rithesh Kumar
  2. Tristan Deleu
  3. Evan Racah

About

Vector Quantized VAEs - PyTorch Implementation

Topics

Resources

Releases

No releases published

Packages

No packages published

Languages