Likelihood ideas

[jhellewell14]

Currently we use full Felsenstein likelihood for the tree (for comparison against phyML). We could try to implement some of calculation saving tricks used in MAPLE or we could also think about how to further integrate PoPPUNK and this tree program.

[jhellewell14]

Profiling suggests nearly all computation time is spent in the log and exp functions in the child_likelihood here:

trees_rs/src/genetic_data.rs

Line 134 in f1fdbdf

pub fn child_likelihood_i(i: usize, ll: ndarray::ArrayBase<ndarray::ViewRepr<&f64>, ndarray::Dim<[usize; 1]>>, p: &na::Matrix4<f64>) -> f64 {

[jhellewell14]

I wonder if you could parallelise the section that iterates over nodes that need their likelihoods updating:

trees_rs/src/treestate.rs

Lines 149 to 166 in f1fdbdf

	for node in nodes {
	let (lchild, rchild) = (node.get_lchild().unwrap(), node.get_rchild().unwrap());
	let seql = match temp_likelihoods.contains_key(&lchild) {
	true => {temp_likelihoods.get(&lchild).unwrap().slice(s![.., ..])},
	false => {slice_data(lchild, self.gen_data)},
	};
	let seqr = match temp_likelihoods.contains_key(&rchild) {
	true => {temp_likelihoods.get(&rchild).unwrap().slice(s![.., ..])},
	false => {slice_data(rchild, self.gen_data)},
	};
	let node_ll = node_likelihood(seql, seqr,
	&matrix_exp(&rate_mat, new_ts.top.nodes[lchild].get_branchlen()),
	&matrix_exp(&rate_mat, new_ts.top.nodes[rchild].get_branchlen()));
	temp_likelihoods.insert(node.get_id(), node_ll);
	}