HongminWu/VariationalRecurrentAutoEncoder

VRAE Variational Recurrent Autoencoder

Variational Recurrent Autoencoder (VRAE) in TensorFlow

forked from Fabius, Otto

Implementation of VRAE paper: "Fabius, Otto, and Joost R. van Amersfoort. "Variational recurrent auto-encoders." arXiv preprint arXiv:1412.6581 (2014)." in Tensorflow on MIDI data.

Original Paper: https://arxiv.org/abs/1412.6581

Primary Requirements:
Tensorflow: Release 1.4
Python 3.0

Tutorial

1. git clone this repo
2. sudo pip install DEPENDANCES
3. new a folder named "MIDI_Data" and download the source MIDI dataset
4. python midi_io.py (convert the original dataset into the *npy format )
5. python vrae.py data_set_idx=0 z_dim=20 time_steps=50

Summary

The Variational Recurrent Auto-Encoder (VRAE) [1] is a generative model for unsupervised learning of time-series data. It combines the strengths of recurrent neural networks (RNNs) and stochastic gradient variational bayes (SGVB) [2]. In the typical VAE, the log-likelihood for a data point $i$ can be written as:

The evidence lower bound (ELBO) is:

Where, the variational distribution and likelihood function are parametrized with recurrent neural networks. In the standard VAE, these networks are dense feed forward networks. VRAE extends this framework by replacing the encoder and decoder dense feed forward neural networks with encoder and decoder recurrent neural nets.

TensorBoard Graph

Data

The MIDI data used is available here

[1] Fabius, Otto, and Joost R. van Amersfoort. "Variational recurrent auto-encoders." arXiv preprint arXiv:1412.6581(2014).

[2] Kingma, Diederik P., and Max Welling. "Auto-encoding variational bayes." arXiv preprint arXiv:1312.6114 (2013).