Comprehensive Examination of Unrolled Networks for Solving Linear Inverse Problems
This repository includes the accompanying code for the paper "Comprehensive Examination of Unrolled Networks for Solving Linear Inverse Problems," which contains a novel generalized deep architecture for solving linear inverse problems called Deep Memory Unrolled Networks (DeMUNs), as well as a comprehensive ablation study on the impact of various hyperparameter choices and on robustness under various sampling conditions. Below are the details regarding the dataset, model architecture, results, and additional resources used.
-
Denoising_Algorithms/Memory_Network/: The model architecture for the Long Memory Unrolled Network is located here. -
Results/: Contains all results, including:- Model weights
- Training histories
- Evaluation losses
-
Auxiliary_Functions/&Denoising_Algorithms/DL_Training/: Houses auxiliary functions for image processing, training, and other utility scripts.
We utilize a subset of the 2012 ImageNet Object Large Scale Visual Recognition Challenge (ILSVRC 2012) - Object Localization Challenge validation dataset, which contains 50,000 images. The code to generate our training dataset can be found in Auxiliary_Functions/image_preprocessing.ipynb, or can be made avialable upon request.
🔗 Dataset link: Kaggle - ILSVRC 2012 Validation Data