def __init__(self, start, rules, max_program_depth=4):
self.start = start
self.rules = rules
self.max_program_depth = max_program_depth
self.hash = hash(format(rules))
self.remove_non_productive(max_program_depth)
self.remove_non_reachable(max_program_depth)
for S in self.rules:
s = sum([self.rules[S][P][1] for P in self.rules[S]])
for P in self.rules[S]:
args_P, w = self.rules[S][P]
self.rules[S][P] = (args_P, w / s)
self.hash_table_programs = {}
self.max_probability = {}
self.compute_max_probability()
self.list_derivations = {}
self.vose_samplers = {}
for S in self.rules:
self.list_derivations[S] = sorted(
self.rules[S], key=lambda P: self.rules[S][P][1]
)
self.vose_samplers[S] = vose.Sampler(
np.array([self.rules[S][P][1] for P in self.list_derivations[S]])
)
def __setstate__(self, d):
self.__dict__ = d
self.vose_samplers = {
S: vose.Sampler(
np.array([self.rules[S][P][1] for P in self.list_derivations[S]])
)
for S in self.rules
}
def sample(self):
"""Sample a row at random.
The `sample` method of a Series is very slow. Additionally, it is not designed to be used
repetitively and requires O(n) steps every time it is called. Instead, we use a Cython
implemention of Vose's alias method that takes O(n) time to build and O(1) time to query.
"""
if self.sampler is None:
self.sampler = vose.Sampler(
weights=self.series.to_numpy(dtype=float),
seed=np.random.randint(2**16))
idx = self.sampler.sample()
return self.series.index[idx]