A Julia module to perform orthogonal matching pursuit.

From the Julia REPL, type

Pkg.clone("https://github.com/klkeys/OMP.jl")

OMP.jl exports two functions, omp and cv_omp. To construct a k-sparse approximation to a response y from a design matrix x, the call is

output = omp(x, y, k)

The result output is a SparseMatrixCSC with one column for each sparsity level 1, 2, ..., k.

To compute the sparsity level of the best predictive model, the call is

output = cv_omp(x, y)

By default cv_omp tests sparsity levels 1, 2, ..., 20 in a q = 5-fold crossvalidation scheme. Use the keyword argument k = MODEL_SIZE to change the maximum sparsity level to MODEL_SIZE. The keyword argument q = NUM_FOLDS sets the number of folds to NUM_FOLDS. Also by default, cv_omp spawns crossvalidation folds on all available processors. The processes involved in the crossvalidation are controled by the keyword argument pids. For example, pids = [1,2,3] will only spawn folds on process IDs 1, 2, and 3.

Signal Recovery From Random Measurements Via Orthogonal Matching Pursuit, Joel A. Tropp and Anna C. Gilbert, IEEE Transactions on Information Theory 53:12, December 2007.