Multi-Operator Imaging (MOI) in PyTorch

Unsupervised Learning From Incomplete Measurements for Inverse Problems

Julián Tachella, Dongdong Chen, Mike E. Davies.

CNRS, France; The University of Edinburgh, UK

In NeurIPS 2022

Run the code

1. Requirements: configure the environment by following environment.yml

2. find the implementation of 'Multi-Operator Imaging (MOI)' at moi.py

3. download datasets from the below source, then preprocess (see our paper for details) and move the datasets under the folders: ../dataset/mri, ../dataset/CelebA, and ../dataset/mnist, repectively:

   • mnist: built-in dataset in PyTorch
   • CelebA: https://www.kaggle.com/jessicali9530/celeba-dataset
   • fastMRI (only the subset 'Knee MRI'): https://fastmri.med.nyu.edu/

4. Train: run the below scripts to train/test the models:

   • run demo_train.py to train MOI for CS-MNIST, Inpainting-MNIST, Inpainting-CelebA, and MRI-fastMRI tasks, respectively. All the trained models can be found in the folder './ckp/'
   • or run train_bash.py to train MOI models on all tasks.

   bash train_bash.sh
   

5. Test: run demo_test.py to test the performance (PSNR) of a trained model on a specific task.

   python3 demo_test.py
   

Citation

If you use this code for your research, please cite our papers.

@inproceedings{tachella2021sampling,
title={Unsupervised Learning From Incomplete Measurements for Inverse Problems},
author={Tachella, Juli{\'a}n and Chen, Dongdong and Davies, Mike},
booktitle={Proceedings of the 36th Conference on Neural Information Processing Systems},
year={2022}}