This package solves the Sparse Programming problem described in the paper. The aforementioned paper solves cardinality constrained or penalized optimization problem. The algorithm in the paper works for non-convex objective functions but this repo only implements linear regression and logistic regression loss functions.
All the package dependencies are included in requirements.txt
The jupyter notebook sample_usage.ipynb walks through a simple case of how to set up a problem and how to solve it.
synthetic_expt.py is the code used to generate synthetic problems and solve them via the procedure described in the paper. To run synthetic_expt.py script, type
python synthetic_expt.py --dim 100,30,10 --loss l2 --l2reg constr --l0reg constr --sparsity 0.1 --gamma 1 --lowrank soft --iterate_over rank --param_grid 1,5,10 --save
Look at the module docstring for more information.