https://github.com/musyoku/ddgm

Chainer implementation of Deep Directed Generative Models with Energy-Based Probability Estimation
https://github.com/musyoku/ddgm

Last synced: 11 days ago
JSON representation

Chainer implementation of Deep Directed Generative Models with Energy-Based Probability Estimation

• Host: GitHub
• URL: https://github.com/musyoku/ddgm
• Owner: musyoku
• Created: 2016-09-27T12:40:03.000Z (about 8 years ago)
• Default Branch: master
• Last Pushed: 2016-11-20T05:59:43.000Z (almost 8 years ago)
• Last Synced: 2024-08-02T12:22:06.309Z (3 months ago)
• Language: Python
• Homepage:
• Size: 131 KB
• Stars: 2
• Watchers: 3
• Forks: 1
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
## Deep Directed Generative Models with Energy-Based Probability Estimation

code for the [paper](https://arxiv.org/abs/1606.03439)

[この記事](http://musyoku.github.io/2016/10/28/Deep-Directed-Generative-Models-with-Energy-Based-Probability-Estimation/)で実装したコードです。

### Requirements

- Chainer 1.17

- PIL

- pylab

Contains the following repository:

- [chainer-sequential](https://github.com/musyoku/chainer-sequential)

## 2D datasets

Train generator to generate 10 dimensional Gaussian mixture distribution and swiss-roll distribution.

![gaussian](https://github.com/musyoku/musyoku.github.io/blob/master/images/post/2016-10-28/gaussian.png?raw=true)

![swiss_roll](https://github.com/musyoku/musyoku.github.io/blob/master/images/post/2016-10-28/swissroll.png?raw=true)

See videos:

- [https://gfycat.com/DarlingShowyHypsilophodon](https://gfycat.com/DarlingShowyHypsilophodon)

- [https://gfycat.com/UnrulyMisguidedHornedviper](https://gfycat.com/UnrulyMisguidedHornedviper)

### Running

run `train_2d/train.py` to train the model.

run `train_2d/gif_gaussian.py` or `train_2d/gif_swissroll.py` to generate gif frames.

## MNIST

run `train_mnist/train.py`

If there is no MNIST image, it will be downloaded automatically.

### Genereted images

![result](https://github.com/musyoku/musyoku.github.io/blob/master/images/post/2016-10-28/mnist_success.png?raw=true)

## killmebaby（キルミーベイベー）

Download 686 images from [http://killmebaby.tv/special_icon.html](http://killmebaby.tv/special_icon.html) and resize all to 64x64 pixels.

run `train_killmebaby/train.py` 

### Original images

![original](https://github.com/musyoku/musyoku.github.io/blob/master/images/post/2016-10-28/kb_original.png?raw=true)

### Images generated by Deep Generative Model

![gen](https://github.com/musyoku/musyoku.github.io/blob/master/images/post/2016-10-28/kb_gen.png?raw=true)

Since the position of the face of the training data is not constant, I think it is difficult to train the generator, but relatively clean images are generated.

### When learning of Generator did not go well

![gen](https://github.com/musyoku/musyoku.github.io/blob/master/images/post/2016-10-28/kb_fail.png?raw=true)

Whichever noise z is used to generate an image, an average is generated.