It would be nice to take advantage of sparse matrix multiplication when cov_type='poisson'. With 100 cause and 100 effect bins I find that about 80% of compute time is spent in
mixer.get_MC_err() # unfold.py line 214
and about 85% of that time is spent doing
Vc1 = dcdP.dot(CovPP).dot(dcdP.T) # mix.py line 242
Of course, due to the rapid scaling of the problem, these will get much more expensive going up to even say 200 bins each. But if the covariance is poisson, the covariance should be sparse, and a full dot product over the full [nbins^2 x nbins^2] matrix is not necessary.
jvavrek
commented
4 years ago
It would be nice to take advantage of sparse matrix multiplication when cov_type='poisson'. With 100 cause and 100 effect bins I find that about 80% of compute time is spent in
and about 85% of that time is spent doing
Of course, due to the rapid scaling of the problem, these will get much more expensive going up to even say 200 bins each. But if the covariance is poisson, the covariance should be sparse, and a full dot product over the full [nbins^2 x nbins^2] matrix is not necessary.