unrolled-nn-restoration

This repository contains the content of the mini-project for the Machine Learning for Time Series course of the MVA master (2023-2024).

The paper implemented here is Unrolled deep networks for sparse signal restoration.

Installation

To install the required packages, run the following command:

pip install -r requirements.txt

Data

The datasets used in this project are the following:

• MassBank (for the mass spectrometry data), with the same script used by (Gharbi et al., 2023) to degrade the dataset using specified kernels that can vary. Generate the dataset using the following command:

python data/generate_dataset.py --Dataset massbank

• Synthetic data generated using the notebook gen_sparse_signal_draft.ipynb.

Note that the final generated datasets used for training the models are extremely large and not directly included in this repository.

We include a small dataset for testing purposes and in order to run the main notebook main.ipynb which shows how to train and test the models rapidly. However, the experiments in the report were run on the full datasets and were run using the scripts src/main.py and src/experiments.py along with a few notebooks for visualization purposes.

Usage

The main notebook 1_presentation_notebook.ipynb shows how to train and test the models rapidly and outline the main idea behind the different objects and functions in the source code.

In order to train a model, choose a config file in configs/ (for example primal_dual.yaml) and run the following command:

python -m src.main --config configs/primal_dual.yaml --data_path data/massbank --n_epochs 10

It is also possible to specify multiple other parameters to script that will be used in the model and override the parameters in the config file or to create your own config file and directly use it. A more detailed description of the parameters can be found in src/main.py.

Experiments

The experiments in the report were run using src/experiments.py and the configurations files are in configs/experiments/. The models are then saved in models/ and the results were uploaded to Weights & Biases.

An experiment consists of training a model on a dataset with specified parameters in the config file (or given as argument to the script) and then testing the model on datasets specified in the config file of the experiment. The results are then uploaded to Weights & Biases and can be visualized in the corresponding project.

In order to run the experiments, run the following command:

python -m src.experiments --config configs/experiments/learn_h_default_fixed_ista.yaml --n_epochs 30

If a model has already been trained and saved, it is possible to test it on a dataset by running the following command:

python -m src.experiments --config configs/experiments/learn_h_default_fixed_ista.yaml --test