Timing the vectorized pruning algorithm vs. the normal one

[LinguList]

I modified (refactored?) the vectorized pruning algorithm and added a more general function for the calculations in order to see how much impact the changes make.

$ time python test_pruning.py time pv

real	0m14,346s
user	0m21,164s
sys	0m22,824s

$ time python test_pruning.py time mk

real	0m1,708s
user	0m1,969s
sys	0m1,316s

$ time python test_pruning.py time prune

real	0m16,054s
user	0m23,763s
sys	0m25,403s

This shows that first, the MK (which we may not use often) is of course the fastest, as we don't have to use the scipy function and the matrix computation. The vectorized version gives us a slight gain of about 2 second for 10000 operations. Since numpy is very slow in creating arrays and matrices, I wonder if one could somehow speed this whole procedure up by preventing matrices to be created.

[LinguList]

The dictionary that stores precomputed matrix exponentials yields a speed-up of 12 seconds on the test.

[PhyloStar]

The branch lengths are typically sampled in MCMC and vary with every iteration. We will have to see if storing branch lengths is going to speedup the performance. But I think we can live with up to a degree of rounding the branch lengths. Just thinking loud. Also, storing the branch length by branch name might also be needed.

[PhyloStar]

This shows that first, the MK (which we may not use often) is of course the fastest, as we don't have to use the scipy function and the matrix computation. The vectorized version gives us a slight gain of about 2 second for 10000 operations. Since numpy is very slow in creating arrays and matrices, I wonder if one could somehow speed this whole procedure up by preventing matrices to be created.

Cool test. We can always use F81 model that distinguishes rates between two pairs of states since it also does not need expm function and has a closed formula for matrix exponential.

https://github.com/pylogeny/bayes/blob/master/src/cybayes/commands/mcmc_gamma.pyx#L932

https://en.wikipedia.org/wiki/Models_of_DNA_evolution#F81_model_(Felsenstein_1981)

[PhyloStar]

The node_dict creation can be moved out of the prune function and scores can be passed to the prune function right?

[LinguList]

Yes, this is also possible. I suggest then that we add the F81 model as one more model for testing now.

[LinguList]

I am currently also looking into first tree move operations, and have a re-rooting function to present as a PR to the pylotree package soon, so I hope we can add the mcmc code also at some point sooner than later (but first we'll do some clear tests of the data).