Adversarial Non-linear Independent Component Analysis

This is an implementation of the work descriped in the ICML 2017 workshop on implicit models titled "Maximizing Independence with GANs for Non-linear ICA"

ArXiv version (with a slightly different title): [https://arxiv.org/abs/1710.05050]

Dependencies

• numpy, tensorflow, visdom

Optional dependencies:

• matplotlib for plotting results
• scipy for reading the audio data
• the audio data itself

Installation

Clone the repository and you should be good to go.

Training a model

First, start a visdom server using python -m visdom.server for visualizing the results.

Next, run python train.py -c ./examples/gan_mlp_example.conf --vd_server=http://127.0.0.1 to train a model with the settings from one of the example configurations. The settings in the configuration files can be overridden using the command line.

python train.py -h will print the available command line and configuration options.

The folder ./examples/best contains the hyper-parameters found using a random search.

License

MIT