https://github.com/bshall/vectorquantizedvae

A PyTorch implementation of "Continuous Relaxation Training of Discrete Latent Variable Image Models"
https://github.com/bshall/vectorquantizedvae

generative-models pytorch vae vq-vae

Last synced: 5 months ago
JSON representation

A PyTorch implementation of "Continuous Relaxation Training of Discrete Latent Variable Image Models"

• Host: GitHub
• URL: https://github.com/bshall/vectorquantizedvae
• Owner: bshall
• License: mit
• Created: 2019-09-10T08:17:55.000Z (about 6 years ago)
• Default Branch: master
• Last Pushed: 2020-03-25T18:26:31.000Z (over 5 years ago)
• Last Synced: 2025-04-08T14:45:48.776Z (6 months ago)
• Topics: generative-models, pytorch, vae, vq-vae
• Language: Jupyter Notebook
• Size: 975 KB
• Stars: 73
• Watchers: 1
• Forks: 16
• Open Issues: 1
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

          
# Vector Quantized VAE

A PyTorch implementation of [Continuous Relaxation Training of Discrete Latent Variable Image Models](http://bayesiandeeplearning.org/2017/papers/54.pdf).

Ensure you have Python 3.7 and PyTorch 1.2 or greater. 

To train the `VQVAE` model with 8 categorical dimensions and 128 codes per dimension 

run the following command:

```

python train.py --model=VQVAE --latent-dim=8 --num-embeddings=128

``` 

To train the `GS-Soft` model use `--model=GSSOFT`. 

Pretrained weights for the `VQVAE` and `GS-Soft` models can be found 

[here](https://github.com/bshall/VectorQuantizedVAE/releases/tag/v0.1).



    



The `VQVAE` model gets ~4.82 bpd while the `GS-soft` model gets ~4.6 bpd.

# Analysis of the Codebooks 

As demonstrated in the paper, the codebook matrices are low-dimensional, spanning only a few dimensions:



    



Projecting the codes onto the first 3 principal components shows that the codes typically tile 

continuous 1- or 2-D manifolds: