mknw / nf-synth

Framework for analysis of Normalizing Flows based Generative models. Analyses include: similarity between classes, dimensionality reduction (PCA, UMAP), experimental image compression.

NF Synth

The present code is in prototype stage.

This repository contains the code for the MSc Thesis: "Normalising Flow: Seeing Latent Space". Written and coded by Michael Accetto and supervised by Jakub Tomczak.

Project

It is intended to compress images with continuous, Normalising Flow (NF) generative models.

The NF architectures are:

• Glow (link to paper)
• Real-NVP (link to paper)

Usage

Network training:

$ python3 train.py [--config /path/to/config.yml]

Compression visualisation:

$ python3 reduce.py [--config /path/to/config.yml]

Similarity analysis:

$ python3 similarity.py [--config /path/to/config.yml]

Models

Synthesizer

Abstraction over step based compression module comprising:

• Normalising Flow Architecture (Glow or Real NVP)
• Principal Component Analysis (PCA)
• Uniform Manifold Approximation Projection (UMAP, optional)

Synthesizer is a scikit compliant transformer class implementing the methods: fit, transform, inverse_transform, fit_transform.

A more detailed description will be soon pushed to main.

Architectures

Glow

Trainable on CelebA-128 and FFHQ-128. Additionally, the model contains a learned prior at the end of each flow-step.

Glow samples, FFHQ dataset: