Implementation of a 1D conditional Invertible Neural Network
Explore the docs »
Table of Contents
The OScINN implements the invertible- and conditional neural networks from the paper "Investigation of inverse design of multilayer thin-films with conditional invertible Neural Networks" [1, 2]. The architecture is heavily based on the contribution of Ardizzone et al [3] who published the FrEIA package which enabled our work.
The Repo contains only the architecture for the network, which was evaluated in the aforementioned paper but also contains a short Introduction jupyter notebook which gives an example how to use the networks.
For replicating the results of the paper, the thin-film computation/dataset generation can be done via the TMM-Fast(https://github.com/MLResearchAtOSRAM/tmm_fast) package, which also contains convenience routines for dataset generation. Have fun!
Since the scope of this package is quite narrow, it is best to follow the Introduction.ipynb. The package provides encapsulates the functionality provided by the FrEIA package to make it easy to replicate the results of the publication [1]. It is of course possible to use the architecture for other 1D problems to but then, it might be wise to implement the networks yourself in order to gain access to the full capabilities of conditional Invertible Neural Networks.
You only need a running Python environment with Python >= 3.7. Its recommended to use Conda to set up the environment.
- Clone the repo by typing into your prompt
git clone https://github.com/MLResearchAtOSRAM/OsCINN.git
- Install requirements via pip
pip install -r requirements.txt
- Train some Network!
Distributed under the MIT License. See LICENSE.txt for more information.
Thanks to
- Heribert Wankerl for contributing to the development
- Maike Stern for running the Repo
- Daniel Grünbaum just for fun
[1] Luce, A. et. al. (2022). Investigation of inverse design of multilayer thin-films with conditional invertible Neural Networks. Preprint.
[2] Luce, A. et. al. (2022). Investigation of inverse design of multilayer thin-films with conditional invertible Neural Networks. Publication.
[3] Ardizzone, L. et. al. (2020). Conditional Invertible Neural Networks for Diverse Image-to-Image Translation. https://doi.org/10.48550/arXiv.2105.02104