spams-cython

This package exposes convenient interfaces of some functions of the SPArse Modeling Software (SPAMS) C++ library, allowing them to be used from Cython code by releasing the Python Global Interpreter Lock (GIL)

Installation from PyPI

pip install spams-cython

Installation from source

git clone https://github.com/getspams/spams-cython.git
cd spams-cython
pip install .

Use spams-cython into your project

pyproject.toml

[build-system]
requires = [
    "setuptools",
    "Cython",
    "spams-cython"
]

setup.py

from setuptools import setup, Extension
from Cython.Build import cythonize
import cyspams

extensions = [
    Extension(
        mymodule,
        sources['mymodule.pyx'],
        include_dirs=cyspams.get_include()
    )
]

setup(ext_modules=cythonize(extensions))

mymodule.pyx

from cyspams.interfaces cimport nnls, lasso

Note

To build your project with spams-cython you need to have a BLAS/LAPACK library on your system (e.g. OpenBLAS, Intel MKL)

Exposed functions

nnls

void nnls(const double *A, const double *y, const int m, const int n, double *x, double &rnorm)

Args in:
A → Matrix 'A = (m, n)' stored as 1D contiguous array (column-major order)
y → Vector 'y = (m)'
m → Dimension 'm'
n → Dimension 'n'

Args out:
x → Solution vector 'x = (n)'
rnorm → Squared Euclidean norm of the final residual vector

---

lasso

void lasso(double *A, double *y, const int m, const int p, const int n, double *x)
void lasso(double *A, double *y, const int m, const int p, const int n, double *x, const double lambda1, const double lambda2)
void lasso(double *A, double *y, const int m, const int p, const int n, double *x, const double lambda1, const double lambda2, const int mode, const bint pos)
void lasso(double *A, double *y, const int m, const int p, const int n, double *x, const double lambda1, const double lambda2, const int mode, const bint pos, const bint ols, const int max_length_path, const int L, const bint cholesky, const int n_threads, const bint verbose)

Args in:
A → Matrix 'A = (m, p)' stored as 1D contiguous array (column-major order)
y → Matrix 'y = (m, n)' stored as 1D contiguous array (column-major order)
m → Dimension 'm'
p → Dimension 'p'
n → Dimension 'n'
lambda1 → Regularization parameter (default = 0.0)
lambda2 → Parameter for solving the Elastic-Net (default = 0.0)
mode → 0 = L1COEFFS, 1 = L2ERROR, 2 = PENALTY, 3 = SPARSITY, 4 = L2ERROR2, 5 = PENALTY2, 6 = FISTAMODE (default = 2)
pos → Adds non-negativity constraints on the coefficients (default = true)
ols → Perform an orthogonal projection before returning the solution (default = false)
max_length_path → Maximum length of the path (default = -1)
L → Maximum number of steps of the homotopy algorithm. Can be used as a stopping criterion (default = -1)
cholesky → Choose between Cholesky implementation or one based on the matrix inversion Lemma (default = false)
n_threads → Number of threads to use (default = 1)
verbose → Verbose mode (default = false)

Args out:
x → Solution matrix 'x = (p, n)' stored as 1D contiguous array (column-major order)

---