softmin/ReHLine-python
Regularized Composite ReLU-ReHU Loss Minimization with Linear Computation and Linear Convergence
What's novel
Regularized Composite ReLU-ReHU Loss Minimization with Linear Computation and Linear Convergence
Code Analysis
11 files read · 4 roundsA high-performance C++/Python solver for empirical risk minimization with convex piecewise linear-quadratic losses and linear constraints, featuring a novel coordinate descent algorithm with free variable sets and shrinking heuristics.
Strengths
Genuinely novel optimization algorithm with strong theoretical foundation (NeurIPS 2023), high-performance C++ core with clean Python API layers, comprehensive test suite validating against sklearn, and support for complex problems including constraints and ElasticNet regularization.
Weaknesses
Some code duplication in the Python layer (similar fit methods across classes), the sklearn mixin module is quite large and could benefit from refactoring, and there's some commented-out code in the path solution module.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
95
Contributors
9
Files
89
Active weeks
28
Repository
Language
Python
Stars
18
Forks
6
License
MIT