fzergz / slow-generative-features

Code for the paper "Exploring Slow Feature Analysis for Extracting Generative Latent Factors".

Exploring Slow Feature Analysis for Extracting Generative Latent Factors

This code is for the paper "Exploring Slow Feature Analysis for Generative Latent Factors" by Max Menne, Merlin Schüler, and Laurenz Wiskott published at ICPRAM 2021.

Overview

This repository contains the code to reproduce the experiments presented in the paper. The experiments are divided into the following files:

1. analyzing_reconstructability.py contains the code for the experiments in Section 3.3,
2. latent_space_explorations.py contains the code for the latent space explorations in Section 3.4.1,
3. exploring_embeddings.py contains the code for the investigation of the embeddings in Section 3.4.1 & 3.4.2,
4. separated_extraction.py contains the code for the separated extraction of latent factors in Section 3.4.3,
5. fitting_prior_distributions.py contains the code for fitting the defined prior distributions in Section 3.5.1,
6. predicting_latent_samples.py contains the code for the prediction of latent samples in Section 3.5.2.

Furthermore, models.py provides the implementations of the models, pretrained_models contains the trained model weights for the different experiments and core includes classes for the generation of several datasets as well as the implementation of the PowerSFA framework.

Requirements

To install requirements use

pip install -r requirements.txt

Further, make sure to install the newest version of the modular-data-processing toolkit.

Usage

To reproduce an experiment, simply run

python experiment_name.py

The selection and configuration of the individual models and datasets as well as the training procedure of the models can be configured within the setup section at the beginning of the respective script of each experiment.