HeavyTailed-VAE

Matlab implementation of a Variational Auto-Encoder using the following distributions to learn a latent space

• Normal
• LogNormal
• Inverse Gaussian

Installation

git clone https://github.com/francescogrillea/HeavyTailed-VAE
cd HeavyTailed-VAE/matlab
mkdir embeddings model_dumps

Usage

The main script will create a new model and train it from scratch, according to the configuration specified in config.json file, and then evaluate the reconstruction error on the test set. If you want to run some configuration example, copy the desired file from matlab/config_examples/ to matlab folder. For each configuration file, a folder will be created in model_dumps/ and named with its runID, generated as [dataset]-[samplingLayerType]-[timestamp].

> main("config.json")

More configurations can be executed in a single run, simply main("config1.json", "config2.json").

To run test separately, execute:

> run_test("runID")

where runID is the folder name in model_dumps/ for a specific run.

To use the model as a generative model, you need first to generate data embeddings for each class, then can be used to generate new samples.

> generateEmbeddings("runID")
> generate("runID")

Models and results are stored in model_dumps/runID.

Results

Read report for more details.