Introduction

This repository contains the scripts and code of the experiments we conducted for the Machine Learning Reproducibility Challenge Spring 2021. The experiments are named after the respective captions from the results section. The algorithm is implemented in train_abs.py and test_abs.py, the script to obtain the datasets is dataset.py. The file util.py is required to properly plot the results.

References

This repository contains parts of the original implementation to reproduce the reported results.

• The original paper: https://www.ecva.net/papers/eccv_2020/papers_ECCV/papers/123750426.pdf
• The original implementation: https://github.com/CSIPlab/learnPR_reference

Run The Code

Please follow the instructions below before you run the notebooks.

Environment

We run our experiments on an jupyter lab server with the following packages:

• numpy 1.21.0
• scikit-image 0.18.1
• pytorch 1.5.0
• torchvision 0.6.0+cu101

You can install the most important packages using: pip install -r requirements.txt

Download The Data

You need to download and unzip CelebA manually before running some of the notebooks. For this you can download CelebA from the official source

CelebA Google Drive Repository

and unzip the dataset into a folder named ./data/, such that the .png images are accessible via ./data/img_align_celeba. This is required to load CelebA with the provided script.

Learn The References

It is required to first run the notebook reconstructions-using-learned-references.ipynb as this creates the self-trained references that are used in the other notebooks.

Experiments

At this point, you are ready to run all notebooks in any order.