I have a problem with non-negative matrix factorization in octave. I try to estimate synergies from Emg-data, but octave only lets me do this for two or more synergies, but not for one. I was able to reproduce the problem with the following code. nmf_bpas is from the linear-algebra pkg from octave-forge.
V=rand(4, 20);
k=1; k2=2;
[W, H, Iter, HIS]=nmf_bpas(V,k);
with this input I get the following error:
error: vertical dimensions mismatch (4x1 vs 1x20)
error: called from
nmf_bpas>getStopCriterion at line 373 column 19
nmf_bpas at line 266 column 26
when I define k>1 as in the following it works
[W2, H2, Iter2, HIS2]=nmf_bpas(V,k2);
with this input, it works fine, and the output matrices are W2 (4x2) and H2 (2x20).
same problem is when I try it with other codes for nnmf and specify the input data Winit and Hinit for matrices (4x1) and (1x20) (eg.:[W, H] = nmf_pg (V, Winit, Hinit, tol, timelimit, maxiter))
in matlab it works with the nnmf function.
I would be glad about help
This seems to be a bug in nmf_bpas.
From what I can tell, the bug is on line 373. Change
pGrad = [gradW(gradW<0|W>0); gradH(gradH<0|H>0)];
to
pGrad = [gradW(gradW<0|W>0); gradH(gradH<0|H>0)(:)];
and it will work.