Team: Daniel Galperin, Jonas Hellgoth, Alexander Kunkel
This repository contains our final project for the lecture "Advanced Machine Learning" at the University of Heidelberg.
Using conditional invertible neural networks for image-to-image translation with landscape photos and Monet paintings. Our source code heavily draws on https://github.com/VLL-HD/conditional_INNs and the corresponding publication https://arxiv.org/abs/1907.02392.
All relevant python files for the final project can be found in the source folder:
| config.py | hyperparameters and paths |
| data.py | load data from path specified in config.py |
| models.py | includes final architecture MonetCINN_squeeze |
| train.py & eval.py | train and evaluate models |
You can safely ignore all other directories and files.
The trained model used to generate all figures in the report can be downloaded here: https://drive.google.com/file/d/1obP2slgHca-HhP31gpaQIm5Qs374-4pT/view?usp=sharing
For loading, set appropriate paths in config.py.
We were not sure if we are allowed to make the data sets public here. Therefore, please contact us if you wish to have access to the data we used.
You can find animations of images linearly interpolating between the reconstruction z and -z in latent space for 64 test images under 'test animations'.
All version numbers are only the minimum version required to run the code. Probably, most other versions will work too.
| Package | Version |
| Pytorch | 1.8.0 |
| Numpy | 1.19.5 |
| Matplotlib | 3.2.2 |
| scikit-learn | 0.22.2 |
| PIL | 0.1.12 |
| albumentations | 7.1.2 |
| cv2 | 4.1.2 |