Learning Sparsifying Transforms using closed-form gradients

Table of Contents

• One-time Setup
• Updating
• Working with the Code
• Learning the Transform on 1D Signals
• Learning the Transform on 2D Images
• Gradients and Time Comparisons

One-time Setup

1. Install a working version of conda.
2. Create a conda environment: conda env create --file environment.yml
3. Activate the conda environment: conda activate learnreg1d

Updating

1. Pull the latest changes: git pull
2. Update the conda environment: conda env update -f environment.yml

Working with the Code

Activate the conda environment before working with the code: conda activate learnreg1d

Learning the Transform on 1D Signals

Run the following command to learn the transform on 1D signals:

python scripts/1D_learn.py --learning_rate 0.001 --num_steps 1e6 --sign_threshold 0.0001 --signal_type piecewise_constant --n 64 --noise_sigma 0.01 --forward_model_type identity --num_training 10000 --transform_type identity --transform_scale 0.01 --num_testing 5 --seed 0 --batch_size 5

Learning the Transform on 2D Images

Run the following command to learn the transform on 2D images:

python scripts/example_image.py --filename "barbara_gray.bmp" --patch_size 8 --forward_model_type identity --noise_sigma 0.01 --transform_type identity --transform_scale 0.01 --SEED 0 --learning_rate 0.001 --num_steps 100000 --sign_threshold 0.0001 --batch_size 5

Gradients and Time Comparisons

Run the notebook scripts/check_gradients.ipynb for gradient and time comparisons with autodiff solvers.

Citation and Paper

This work has been published in ICASSP-2022 and SIAM Journal on Imaging Sciences. You can read the full paper on arXiv.

To cite this work, please use the following BibTeX entry:

@article{ghosh2022learning,
  title={Learning Sparsity-Promoting Regularizers using Bilevel Optimization},
  author={Ghosh, Avrajit and McCann, Michael T and Mitchell, Madeline and Ravishankar, Saiprasad},
  journal={arXiv preprint arXiv:2207.08939},
  year={2022}
}