GitHub repository to reproduce experiments from the paper: Ambiguity in solving imaging inverse problems with deep learning based operators. Pre-print available on arXiv.org at: https://arxiv.org/abs/2305.19774.
Feel free to cite us:
@article{evangelista2023ambiguity,
title={Ambiguity in solving imaging inverse problems with deep learning based operators},
author={Evangelista, Davide and Morotti, Elena and Piccolomini, Elena Loli and Nagy, James},
journal={arXiv preprint arXiv:2305.19774},
year={2023}
}
The use the code, simply clone the GitHub repository locally with
git clone https://github.com/devangelista2/UnderstandingStabilizers.git
Moreover, the IPPy library is required to execute portions of the code. Please refer to IPPy documentation for an explanation on how to install it.
Please note that the functions requires a specific folders and files structure to work. Since, due to memory constraint, it was not possible to upload the whole project on GitHub, the user is asked to create some folders to follow the required structure. This can be obtained by simply creating the data and the model_weights folders by running:
mkdir data
mkdir gaussian_blur/model_weights
mkdir motion_blur/model_weights
Into the main project folder. For informations about how to download the data (to be placed inside the data folder), and the pre-trained model weights, please refer to the following.
To run the experiments, the training and the test set has to be downloaded. A copy of the data used to train the models and get the results for the paper is available on HuggingFace. To get it, simply create a folder named data into the main project directory, move into that and run the following command:
git lfs install
git clone https://huggingface.co/datasets/TivoGatto/celeba_grayscale
which will download the data, in .npy format, used in the experiments. It is a slighly modified version of the GoPro dataset, where the images has been cropped to be
The pre-trained models will be available soon.