void nmf(viennacl::matrix_base<NumericT> const & V,
viennacl::matrix_base<NumericT> & W,
viennacl::matrix_base<NumericT> & H,
viennacl::linalg::nmf_config const & conf)
{
viennacl::hsa::context & ctx = const_cast<viennacl::hsa::context &>(viennacl::traits::hsa_context(V));
const std::string NMF_MUL_DIV_KERNEL = "el_wise_mul_div";
viennacl::linalg::opencl::kernels::nmf<NumericT, viennacl::hsa::context>::init(ctx);
vcl_size_t k = W.size2();
conf.iters_ = 0;
if (viennacl::linalg::norm_frobenius(W) <= 0)
W = viennacl::scalar_matrix<NumericT>(W.size1(), W.size2(), NumericT(1), ctx);
if (viennacl::linalg::norm_frobenius(H) <= 0)
H = viennacl::scalar_matrix<NumericT>(H.size1(), H.size2(), NumericT(1), ctx);
viennacl::matrix_base<NumericT> wn(V.size1(), k, W.row_major(), ctx);
viennacl::matrix_base<NumericT> wd(V.size1(), k, W.row_major(), ctx);
viennacl::matrix_base<NumericT> wtmp(V.size1(), V.size2(), W.row_major(), ctx);
viennacl::matrix_base<NumericT> hn(k, V.size2(), H.row_major(), ctx);
viennacl::matrix_base<NumericT> hd(k, V.size2(), H.row_major(), ctx);
viennacl::matrix_base<NumericT> htmp(k, k, H.row_major(), ctx);
viennacl::matrix_base<NumericT> appr(V.size1(), V.size2(), V.row_major(), ctx);
NumericT last_diff = 0;
NumericT diff_init = 0;
bool stagnation_flag = false;
for (vcl_size_t i = 0; i < conf.max_iterations(); i++)
{
conf.iters_ = i + 1;
{
hn = viennacl::linalg::prod(trans(W), V);
htmp = viennacl::linalg::prod(trans(W), W);
hd = viennacl::linalg::prod(htmp, H);
viennacl::hsa::kernel & mul_div_kernel = ctx.get_kernel(viennacl::linalg::opencl::kernels::nmf<NumericT>::program_name(), NMF_MUL_DIV_KERNEL);
viennacl::hsa::enqueue(mul_div_kernel(H, hn, hd, cl_uint(H.internal_size1() * H.internal_size2())));
}
{
wn = viennacl::linalg::prod(V, trans(H));
wtmp = viennacl::linalg::prod(W, H);
wd = viennacl::linalg::prod(wtmp, trans(H));
viennacl::hsa::kernel & mul_div_kernel = ctx.get_kernel(viennacl::linalg::opencl::kernels::nmf<NumericT>::program_name(), NMF_MUL_DIV_KERNEL);
viennacl::hsa::enqueue(mul_div_kernel(W, wn, wd, cl_uint(W.internal_size1() * W.internal_size2())));
}
if (i % conf.check_after_steps() == 0) //check for convergence
{
appr = viennacl::linalg::prod(W, H);
appr -= V;
NumericT diff_val = viennacl::linalg::norm_frobenius(appr);
if (i == 0)
diff_init = diff_val;
if (conf.print_relative_error())
std::cout << diff_val / diff_init << std::endl;
// Approximation check
if (diff_val / diff_init < conf.tolerance())
break;
// Stagnation check
if (std::fabs(diff_val - last_diff) / (diff_val * NumericT(conf.check_after_steps())) < conf.stagnation_tolerance()) //avoid situations where convergence stagnates
{
if (stagnation_flag) // iteration stagnates (two iterates with no notable progress)
break;
else
// record stagnation in this iteration
stagnation_flag = true;
} else
// good progress in this iteration, so unset stagnation flag
stagnation_flag = false;
// prepare for next iterate:
last_diff = diff_val;
}
}
}
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);
}
}
}
