UG-VAE

KO | EN

gitlite — search

#javascript #python #hacktoberfest #react #ai #typescript #llm #go #golang #android #machine-learning #rust #deep-learning #linux

★ 8 Open GitHub ↗

Official code for UG-VAE

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# UG-VAE
This repository contains the official Pytorch implementation of the Unsupervised Global VAE (**UG-VAE**) model proposed
in the  paper 
[Unsupervised Learning of Global Factors in Deep Generative Models](https://arxiv.org/abs/2012.08234).

Please, if you use this code, cite the [preprint](https://arxiv.org/abs/2012.08234) using:
```
@article{peis2020unsupervised,
  title={Unsupervised Learning of Global Factors in Deep Generative Models},
  author={Peis, Ignacio and Olmos, Pablo M and Art{\'e}s-Rodr{\'\i}guez, Antonio},
  journal={arXiv preprint arXiv:2012.08234},
  year={2020}
}
```

### Dependencies
```
torch 1.7.0
torchvision 0.8.1
matplotlib 3.3.3
numpy 1.19.4
pandas 1.1.4
scikit-learn 0.23.2
```

## Usage
UG-VAE is implemented as a class that inherits from Pytorch <code>nn.Module</code> in <code>models.py</code>. You can 
train UG-VAE using the <code>train.py</code> script. A few examples are included
below:
```
# example for training celeba:
python3 train.py --dataset celeba --arch beta_vae --epochs 10 --model_name celeba

# example for training mnist
python3 train.py --dataset mnist --arch k_vae --dim_z 10 --dim_beta 20 --K 10 --model_name mnist

# example for training mixed celeba+faces
python3 train.py --dataset celeba_faces --arch beta_vae --dim_z 40 --dim_beta 40 --K 40 --model_name celeba_faces
```
The script will create a log dir in <code>results/[model_name]</code>. In <code>/checkpoints/</code>, model checkpoints for 
each log interval are stored. In <code>/figs/</code> you will have reconstructions and samples at the end of each 
log interval, and a plot with the losses during the training procedure. Some remarkable arguments are:
- <code>--dataset</code>: you can choose among <code>celeba</code>, <code>mnist</code>, <code>celeba_faces</code>, 
 <code>cars_chairs</code> and some others included in <code>datasets.py</code>.
 For [CelebA](http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html), 
 [3D FACES](https://faces.dmi.unibas.ch/bfm/index.php?nav=1-1-1&id=scans),
 [Cars dataset](https://ai.stanford.edu/~jkrause/cars/car_dataset.html),
 and [Chairs](https://www.di.ens.fr/willow/research/seeing3Dchairs/)
 you have to download from the given links and put the images in <code>data/[dataset]/img/</code>
- <code>--arch</code>: The architecture for encoder and decoder networks: use <code>beta_vae</code> for convolutional
 networks that work with 64x64 images (as CelebA), or <code>k_vae</code> for fully-connected networks that work with 
 28x28 images (as MNIST). 
- <code>--model_name</code> will be the name of the logdir stored in <code>results/</code> folder.
- <code>--no_cuda</code> for disabling GPU training.

## Examples
UG-VAE learns both local and global disentanglement from random batches of data in a fully unsupervised manner, which 
leads to a promising performance in domain alignment and discovering non-trivial underlying structures. Some 
experiments are included in the <code>experiments/</code> folder.
### Interpolation
 In <code>experiments/interpolation.py</code> you have an implementation of the Experiment 4.1 of the 
 [paper](https://arxiv.org/abs/2012.08234). By running:
```
python3 interpolation.py --dataset celeba --arch beta_vae --epochs 10 --model_name celeba 
```
you will store in <code>results/[model_name]/figs/interpolation/</code>) figures like the following:

### Domain alignment
In <code>experiments/domain_alignment.py</code> you have an implementation of experiment 4.2 of the 
[paper](https://arxiv.org/abs/2012.08234). By running:
```
python3 interpolation.py --dataset celeba --arch beta_vae --epochs 10 --model_name celeba 
```
you will store in <code>results/[model_name]/figs/interpolation/</code>) a similar figure than the following:

### Global structure
In <code>experiments/celeba_attributes.py</code> and <code>experiments/mnist_series.py</code> you have an implementation
of experiment 4.3 of the [paper](https://arxiv.org/abs/2012.08234). By running:
```
python3 celeba_attributes.py
```
you will obtain figures like the following:
 

<img width="900" src="imgs/exp3.png">

If you run:
```
python3 mnist_series.py
```
you will obtain figures like the following:
 

<img width="400" src="imgs/exp32.png">

## Contributors
[Ignacio Peis](http://www.tsc.uc3m.es/~ipeis/index.html),
[Pablo M. Olmos](http://www.tsc.uc3m.es/~olmos/)
and [Antonio Artés-Rodríguez](http://www.tsc.uc3m.es/~antonio/antonio_artes/Home.html).

For further information: <a href="mailto:ipeis@tsc.uc3m.es">ipeis@tsc.uc3m.es</a>

UG-VAE

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README