def __init__(self, filename=None, interval=None, restore=True):
"""
Set the checkpointing filename and time interval.
Parameters
----------
filename
name of the file to which data will be written. If None, no
checkpointing will be done.
interval
time expressed in seconds
restore
flag to restore from this filename or not. will be set to 0 after
restoration
"""
self.checkpointer = checkpointing.Checkpointer(filename, interval)
self.restore = restore
def trex(
self,
a=8,
k=1000,
start=None,
order=None,
return_all=False,
filename=None,
interval=None,
ui=None,
):
"""TrexML policy for tree sampling - all trees up to size 'a' and
then keep no more than 'k' best trees at each tree size.
'order' is an optional list of tip names.
'start' is an optional list of initial trees. Each of the trees must
contain the same tips.
'filename' and 'interval' control checkpointing.
Advanced step-wise addition algorithm
M. J. Wolf, S. Easteal, M. Kahn, B. D. McKay, and L. S. Jermiin.
Trexml: a maximum-likelihood approach for extensive tree-space
exploration.
Bioinformatics, 16(4):383 94, 2000."""
checkpointer = checkpointing.Checkpointer(filename, interval)
if checkpointer.available():
(init_tree_size, fixed_names, trees) = checkpointer.load()
names = self._consistentNameOrder(fixed_names, order)
elif start is not None:
if not isinstance(start, list):
start = [start]
fixed_names = start[0].get_tip_names()
names = self._consistentNameOrder(fixed_names, order)
trees = []
for tree in start:
# check the start tree represents a subset of tips
assert set(tree.get_tip_names()) < set(
self.names
), "Starting tree names not a subset of the sequence names"
(ancestry, fixed_names2,
lengths) = tree2ancestry(tree, order=fixed_names)
assert fixed_names2 == fixed_names
trees.append((None, None, ancestry))
init_tree_size = len(fixed_names)
else:
trees = [(None, None, numpy.identity(3, int))]
names = self._consistentNameOrder([], order)
init_tree_size = 3
tree_size = len(names)
assert tree_size > 3
if a > tree_size:
a = tree_size
if a < 4:
a = 4
# All trees of size a-1, no need to compare them
for n in range(init_tree_size + 1, a):
trees2 = []
for (err2, lengths2, ancestry) in trees:
for split_edge in range(len(ancestry)):
ancestry2 = grown(ancestry, split_edge)
trees2.append((None, None, ancestry2))
trees = trees2
init_tree_size = n
# Pre calculate how much work is to be done, for progress display
tree_count = len(trees)
total_work = 0
work_done = [0] * (init_tree_size + 1)
for n in range(init_tree_size + 1, tree_size + 1):
evals = tree_count * (n * 2 - 5)
total_work += evals * n
tree_count = min(k, evals)
work_done.append(total_work)
# For each tree size, grow at each edge of each tree. Keep best k.
for n in range(init_tree_size + 1, tree_size + 1):
evaluate = self.make_tree_scorer(names[:n])
def grown_tree(spec):
(tree_ordinal, tree, split_edge) = spec
(old_err, old_lengths, old_ancestry) = tree
ancestry = grown(old_ancestry, split_edge)
(err, lengths) = evaluate(ancestry)
return (err, tree_ordinal, split_edge, lengths, ancestry)
specs = [(i, tree, edge) for (i, tree) in enumerate(trees)
for edge in range(n * 2 - 5)]
candidates = ui.imap(
grown_tree,
specs,
noun=("%s leaf tree" % n),
start=work_done[n - 1] / total_work,
end=work_done[n] / total_work,
)
best = ismallest(candidates, k)
trees = [(err, lengths, ancestry)
for (err, parent_ordinal, split_edge, lengths,
ancestry) in best]
checkpointer.record((n, names[:n], trees))
results = (self.result2output(err, ancestry, lengths, names)
for (err, lengths, ancestry) in trees)
if return_all:
result = self.results2output(results)
else:
result = next(results)
return result
