Conditional Normalizing Flows for Super-Resolution

This repository contains an implementation in Pytorch of conditional normalizing flows applied to super-resolution. This master thesis project was conducted at the Amsterdam Machine Learning Lab 2019.

This Code corresponds to the pre-print: Likelihood Learning with Conditional Normalizing Flows.

Results of the conditional normalizing flow model compared to a factorized likelihood baseline:

Requirements

The requirements for the conda environment in which we have tested this code are started in environment_nfsr.yaml.

The main dependencies are:

• python 3.8.3
• pytorch 1.8.0

The conda environment can be installed via

conda env create -f environment_nfsr.yaml 

And used by running

source activate cnf-sr

Cite

@misc{winkler2019learning,
      title={Learning Likelihoods with Conditional Normalizing Flows}, 
      author={Christina Winkler and Daniel Worrall and Emiel Hoogeboom and Max Welling},
      year={2019},
      eprint={1912.00042},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}