Invertible RIM

Official Pytorch implementation of the code accompanying the 2019 NeurIPS paper Invert to Learn to Invert by Patrick Putzky and Max Welling.

This package supplies methods for defining invertible models in Pytorch, a wrapper class that enables invert to learn on such invertible models, and a reference implementation for the i-RIM. A number of examples demonstrate how to use the package.

This work was used for the model that won the singlcoil track in the fastMRI challenge. Training and evaluation code for the fastMRI problem can befound here.

Installation

This package requires pytorch>=1.3. In order to install invertible_rim, run the following commands

git clone https://github.com/pputzky/invertible_rim.git
cd invertible_rim
pip install --user -r requirements.txt
python setup.py install

Package structure

1. examples

   A number of examples for demonstrating the usage of this package. The focus is on the usage of irim.core.

2. irim.core

   This is the core of the package. It implements all components necessary to build an i-RIM and to train it using invert to learn. This module can easily be extended with other invertible layers and modules.

3. irim.rim

   A reference implementation of the Recurrent Inference Machines (RIM)

4. irim.test A number of tests for module invertible_rim.irim. You can run

   pytest

   to confirm that invert to learn is numerically stable on your system.

Reference

If you use this code or derivatives thereof please cite the associated paper

@incollection{pputzky2019,
title = {Invert to Learn to Invert},
author = {Putzky, Patrick and Welling, Max},
booktitle = {Advances in Neural Information Processing Systems 32},
editor = {H. Wallach and H. Larochelle and A. Beygelzimer and F. d\textquotesingle Alch\'{e}-Buc and E. Fox and R. Garnett},
pages = {444--454},
year = {2019},
publisher = {Curran Associates, Inc.},
url = {http://papers.nips.cc/paper/8336-invert-to-learn-to-invert.pdf}
}