/joint-vae

Pytorch implementation of JointVAE, a framework for disentangling continuous and discrete factors of variation :star2:

Primary LanguageJupyter NotebookMIT LicenseMIT

Learning Disentangled Joint Continuous and Discrete Representations

Pytorch implementation of Learning Disentangled Joint Continuous and Discrete Representations (NIPS 2018).

This repo contains an implementation of JointVAE, a framework for jointly disentangling continuous and discrete factors of variation in data in an unsupervised manner.

Examples

MNIST

CelebA

FashionMNIST

dSprites

Discrete and continuous factors on MNIST

dSprites comparisons

Usage

The train_model.ipynb notebook contains code for training a JointVAE model.

The load_model.ipynb notebook contains code for loading a trained model.

Example usage

from jointvae.models import VAE
from jointvae.training import Trainer
from torch.optim import Adam
from viz.visualize import Visualizer as Viz

# Build a dataloader for your data
dataloader = get_my_dataloader(batch_size=32)

# Define latent distribution
latent_spec = {'cont': 20, 'disc': [10, 5, 5, 2]}

# Build a Joint-VAE model
model = VAE(img_size=(3, 64, 64), latent_spec=latent_spec)

# Build a trainer and train model
optimizer = Adam(model.parameters())
trainer = Trainer(model, optimizer,
                  cont_capacity=[0., 5., 25000, 30.],
                  disc_capacity=[0., 5., 25000, 30.])
trainer.train(dataloader, epochs=10)

# Visualize samples from the model
viz = Viz(model)
samples = viz.samples()

# Do all sorts of fun things with model
...

Trained models

The trained models referenced in the paper are included in the trained_models folder. The load_model.ipynb ipython notebook provides code to load and use these trained models.

Data sources

The MNIST and FashionMNIST datasets can be automatically downloaded using torchvision.

CelebA

All CelebA images were resized to be 64 by 64. Data can be found here.

Chairs

All Chairs images were center cropped and resized to 64 by 64. Data can be found here.

Applications

Image editing

Inferring unlabelled quantities

Citing

If you find this work useful in your research, please cite using:

@inproceedings{dupont2018learning,
  title={Learning disentangled joint continuous and discrete representations},
  author={Dupont, Emilien},
  booktitle={Advances in Neural Information Processing Systems},
  pages={707--717},
  year={2018}
}

More examples

License

MIT