chainer-Variational-AutoEncoder

Variational Auto Encoder implemented by Chainer

Requirement

• Chainer

M1 model

Train

Start training the model using train_VAE.py, for example

$python train_VAE.py

Generate data

You can generate data by giving a latent space vector. For example,

$python generated.py --model [model/created_model.pkl]

M2 model

Train

Start training the model using train_VAE_yz_x.py, for example

$python train_VAE_yz_x.py

Generate data set giving 1 sample input.

You can generate data set by giving a sample input. For example,

$python generated_yz_x.py --model [model/created_model.pkl]

Flying through latent space of M2 model

To generate movies of flying through latent-space of the M2 model, run:

$python run_flying.py --dataset [dataset] --model [model/created_model.pkl] --output_file [output file name]

where dataset is 'mnist' or 'svhn', and output_file is the filename to save the movie file to.

NOTE: This script requires ffmpeg to be installed.

NOTE: Unzip sample model saved in model folder

ToDo

• GPU implementation

Reference

• Justin Bayer's Chainer based Variational Auto Encoder http://nbviewer.ipython.org/gist/duschendestroyer/a41fcab5f7f9ffa45387
• http://papers.nips.cc/paper/5352-semi-supervised-learning-with-deep-generative-models.pdf
• https://github.com/dpkingma/nips14-ssl
• http://www.slideshare.net/beam2d/semisupervised-learning-with-deep-generative-models