Version: 1.0.0

Copyright: Maximilian Behr

License: The software is licensed under under MIT. See LICENSE for details.

cuNMF is a CUDA library implementing multiplicative update rules for the nonnegative matrix factorization $X\approx WH$, using the $\beta$-divergence as an error measure.

cuNMF supports real single and double precision matrices.

int cunmf_info_create(cunmf_info* info);
int cunmf_info_destroy(cunmf_info info);

int cunmf_options_screate(cunmf_options* opt);
int cunmf_options_sdestroy(cunmf_options opt);

int cunmf_sMUbeta_buffersize(int m, int n, int k, double beta, size_t* bufferSize);
int cunmf_sMUbeta(int m, int n, int k, double beta, const float* X, void* buffer, const cunmf_options opt, float* W, float* H, cunmf_info info);

int cunmf_options_dcreate(cunmf_options* opt);
int cunmf_options_ddestroy(cunmf_options opt);

int cunmf_dMUbeta_buffersize(int m, int n, int k, double beta, size_t* bufferSize);
int cunmf_dMUbeta(int m, int n, int k, double beta, const float* X, void* buffer, const cunmf_options opt, float* W, float* H, cunmf_info info);

cuNMF implements the multiplicative update rules to minimize the $\beta$-divergence for the nonnegative matrix factorization $X\approx W H$, where $X$ is of size $m\times n$, $W$ is of size $m\times k$ and $H$ is of size $k\times n$.

In more details, we consider the optimization problem

where $\epsilon$ is a small nonegative constant and the $\beta$-divergence $d_{\beta}(x,y)$ is given by

The case $\beta=0$ gives the Itakura–Saito divergence, $\beta = 1$ gives the Kullback–Leibler divergence, and $\beta=2$ gives the Frobenius norm distance $\frac{1}{2}||\cdot||_F^2$. For more details on the multiplicative update rule see Theorem 8.8 and Theorem 8.9 in

Gillis, Nicolas. Nonnegative matrix factorization. Society for Industrial and Applied Mathematics, 2020.

Prerequisites:

• CMake >= 3.23
• CUDA >= 11.4.2

  mkdir build && cd build
  cmake ..
  make
  make install

The multiplicate update algorithm is an iterative ones. The initial iterates $W_0$ and $H_0$ must be nonnegative. The parameter $k$ (number of columns of $W$ / rows of $H$) must be specified by the user. The user can also specifiy stopping criteria based on the

• number of iterations (maxiter)
• computational time (maxtime)
• relative change of the iterates $W$ and $H$ (tol_relchange_WH)
• relative change of the objective $D_{\beta}$ (tol_relchange_objective).

See example_cunmf_MUbeta.cu for an example using double precision data.