void test_nmf(std::size_t m, std::size_t k, std::size_t n)
{
std::vector<ScalarType> stl_w(m * k);
std::vector<ScalarType> stl_h(k * n);
viennacl::matrix<ScalarType> v_ref(m, n);
viennacl::matrix<ScalarType> w_ref(m, k);
viennacl::matrix<ScalarType> h_ref(k, n);
fill_random(stl_w);
fill_random(stl_h);
viennacl::fast_copy(&stl_w[0], &stl_w[0] + stl_w.size(), w_ref);
viennacl::fast_copy(&stl_h[0], &stl_h[0] + stl_h.size(), h_ref);
v_ref = viennacl::linalg::prod(w_ref, h_ref); //reference
// Fill again with random numbers:
fill_random(stl_w);
fill_random(stl_h);
viennacl::matrix<ScalarType> w_nmf(m, k);
viennacl::matrix<ScalarType> h_nmf(k, n);
viennacl::fast_copy(&stl_w[0], &stl_w[0] + stl_w.size(), w_nmf);
viennacl::fast_copy(&stl_h[0], &stl_h[0] + stl_h.size(), h_nmf);
viennacl::linalg::nmf_config conf;
viennacl::linalg::nmf(v_ref, w_nmf, h_nmf, conf);
viennacl::matrix<ScalarType> v_nmf = viennacl::linalg::prod(w_nmf, h_nmf);
float diff = matrix_compare(v_ref, v_nmf);
bool diff_ok = fabs(diff) < EPS;
long iterations = static_cast<long>(conf.iters());
printf("%6s [%lux%lux%lu] diff = %.6f (%ld iterations)\n", diff_ok ? "[[OK]]":"[FAIL]", m, k, n, diff, iterations);
if (!diff_ok)
exit(EXIT_FAILURE);
}
void nmf(viennacl::matrix<ScalarType> const & v,
viennacl::matrix<ScalarType> & w,
viennacl::matrix<ScalarType> & h,
std::size_t k,
ScalarType eps = 0.000001,
std::size_t max_iter = 10000,
std::size_t check_diff_every_step = 100)
{
viennacl::linalg::kernels::nmf<ScalarType, 1>::init();
w.resize(v.size1(), k);
h.resize(k, v.size2());
std::vector<ScalarType> stl_w(w.internal_size1() * w.internal_size2());
std::vector<ScalarType> stl_h(h.internal_size1() * h.internal_size2());
for (std::size_t j = 0; j < stl_w.size(); j++)
stl_w[j] = static_cast<ScalarType>(rand()) / RAND_MAX;
for (std::size_t j = 0; j < stl_h.size(); j++)
stl_h[j] = static_cast<ScalarType>(rand()) / RAND_MAX;
viennacl::matrix<ScalarType> wn(v.size1(), k);
viennacl::matrix<ScalarType> wd(v.size1(), k);
viennacl::matrix<ScalarType> wtmp(v.size1(), v.size2());
viennacl::matrix<ScalarType> hn(k, v.size2());
viennacl::matrix<ScalarType> hd(k, v.size2());
viennacl::matrix<ScalarType> htmp(k, k);
viennacl::matrix<ScalarType> appr(v.size1(), v.size2());
viennacl::vector<ScalarType> diff(v.size1() * v.size2());
viennacl::fast_copy(&stl_w[0], &stl_w[0] + stl_w.size(), w);
viennacl::fast_copy(&stl_h[0], &stl_h[0] + stl_h.size(), h);
ScalarType last_diff = 0.0f;
for (std::size_t i = 0; i < max_iter; i++)
{
{
hn = viennacl::linalg::prod(trans(w), v);
htmp = viennacl::linalg::prod(trans(w), w);
hd = viennacl::linalg::prod(htmp, h);
viennacl::ocl::kernel & mul_div_kernel = viennacl::ocl::get_kernel(viennacl::linalg::kernels::nmf<ScalarType, 1>::program_name(),
NMF_MUL_DIV_KERNEL);
viennacl::ocl::enqueue(mul_div_kernel(h, hn, hd, cl_uint(stl_h.size())));
}
{
wn = viennacl::linalg::prod(v, trans(h));
wtmp = viennacl::linalg::prod(w, h);
wd = viennacl::linalg::prod(wtmp, trans(h));
viennacl::ocl::kernel & mul_div_kernel = viennacl::ocl::get_kernel(viennacl::linalg::kernels::nmf<ScalarType, 1>::program_name(),
NMF_MUL_DIV_KERNEL);
viennacl::ocl::enqueue(mul_div_kernel(w, wn, wd, cl_uint(stl_w.size())));
}
if (i % check_diff_every_step == 0)
{
appr = viennacl::linalg::prod(w, h);
viennacl::ocl::kernel & sub_kernel = viennacl::ocl::get_kernel(viennacl::linalg::kernels::nmf<ScalarType, 1>::program_name(),
NMF_SUB_KERNEL);
//this is a cheat. i.e save difference of two matrix into vector to get norm_2
viennacl::ocl::enqueue(sub_kernel(appr, v, diff, cl_uint(v.size1() * v.size2())));
ScalarType diff_val = viennacl::linalg::norm_2(diff);
if((diff_val < eps) || (fabs(diff_val - last_diff) < eps))
{
//std::cout << "Breaked at diff - " << diff_val << "\n";
break;
}
last_diff = diff_val;
//printf("Iteration #%lu - %.5f \n", i, diff_val);
}
}
}
