Skip to content

Latest commit

 

History

History
45 lines (36 loc) · 2.45 KB

README.md

File metadata and controls

45 lines (36 loc) · 2.45 KB
Raw

admm-l1-2-logistic-regression

Description

This repo is the final project of COMP633 at UNC-Chapel Hill. We used Alternating Direction Method of Multipliers (ADMM) optimization methods to solve the L-1/L-2 regularized binary logistic regression. In this repo, we provide efficient squential and distributed implementations of ADMM powered by GNU Scientific Library (GSL) and Message Passing Interface (MPI) such that the program can handle millions of samples within couple of minutes.

Dependencies

  • GCC
  • GNU Scientific Library (GSL)
  • Message Passing Interface (MPI)

Usage

Compile

We use gcc/icc to compile the whole project. The standard Makefile has been provided. You may need to change the Makefile slightly to accommodate your local include path and library path. The program has been compiled successfully in MacOS and Linux environment.

Runner

  • logit: Sequential Runner
  • mpi_logit: Distributed Runner
usage: logit/mpi_logit [-A Feature Matrix] [-b Response Vector] [-e Relative Tolerance]
                       [-E Absolute Telerance] [-r Regularization Type] 
                       [-t Maximum Iteration] [-o Output File] [-p Print Progress]
arguments:
  -A:       The path to feature matrix,  e.g. data/input.dat;
  -b:       The path to response vevtor, e.g. data/output.dat;
  -e:       The relative precision tolerance in ADMM algorithm;
  -E:       The absolute precision tolerance in ADMM algorithm;
  -r:       Regularization type to use: only support L-1 and L-2, e.g. -r l2;
  -t:       The maximum number of iterations;
  -o:       The path to the result file, e.g. data/solution.dat
  -p:       Whether to print the optimization progress.

Note: Matrix and vector uses the standard Matrix Market File Format, i.e. Matrix Market Exchange Formats.

Distributed Runner

mpirun -np {\# of cores} ./mpi_logit [Arguments]

Note: For the distributed runner, we follow the data-parallel principle. Every process reads the part of feature matrix and response vector (partitioned by sample). Therefore, the argument -A and b should be the prefix path of data file, e.g. we use -A data/A when A1.dat--A{N}.dat are located in data.

Performance

Reference

Boyd, Stephen, et al. "Distributed optimization and statistical learning via the alternating direction method of multipliers." Foundations and Trends® in Machine learning 3.1 (2011): 1-122.