CoCaIn BPG: Fast Inertial Algorithm for Non-convex Optimization

Convex-Concave Backtracking for Inertial Bregman Proximal Gradient Algorithms in Non-Convex Optimization

by Mahesh Chandra Mukkamala, Peter Ochs, Thomas Pock and Shoham Sabach.

Code theme: CoCaIn BPG escapes Spurious Stationary Points

The goal is to minimize the following non-convex objective (as in Page 18)

The objective function visualizations are given below (as in Page 19).

Dependencies

• numpy, matplotlib

If you have installed above mentioned packages you can skip this step. Otherwise run (maybe in a virtual environment):

pip install -r requirements.txt

Reproduce results

To generate results

chmod +x generate_results.sh
./generate_results.sh

Then to create the plots

python contour_plot.py

Now you can check figures folder for various figures.

Results

Citation

@techreport{MOPS19,
title        = {Convex-Concave Backtracking for Inertial Bregman Proximal Gradient Algorithms in Non-Convex Optimization},
author       = {M.C. Mukkamala and P. Ochs and T. Pock and S. Sabach},
year         = {2019},
journal      = {ArXiv e-prints, arXiv:1904.03537},
}

Related work: CoCaIn BPG for Matrix Factorization

Contact

Mahesh Chandra Mukkamala ([email protected])

References

M. C. Mukkamala, P. Ochs, T. Pock, and S. Sabach: Convex-Concave Backtracking for Inertial Bregman Proximal Gradient Algorithms in Non-Convex Optimization. ArXiv e-prints, arXiv:1904.03537, 2019.

License

Check here.