def get_intermediate_beams(self, examples, verbose=False):
intermediates = [[] for _ in examples]
terminated = [set() for ex in examples]
particles = [[ParsePath.empty(ex.context)] for ex in examples]
distributions = [[1] for _ in xrange(len(examples))]
num_iterations = get_num_iterations(
self._config.iterations_per_utterance, examples)
if verbose:
iterations = verboserate(range(num_iterations),
desc='Performing randomized search')
else:
iterations = xrange(num_iterations)
for _ in iterations:
for ex_idx, (beam, terminated_set) in enumerate(
zip(particles, terminated)):
intermediates[ex_idx].append(Beam(
sorted(terminated_set, key=self._ranker, reverse=True) +
sorted(beam, key=self._ranker, reverse=True)))
terminated, particles, distributions = self.advance(
terminated, particles, distributions)
return [Beam(sorted(list(paths), key=self._ranker, reverse=True))
for paths in terminated], intermediates
def initial_beam(self, context):
"""Return the initial beam for the context.
Args:
context (Context)
Returns:
Beam
"""
return Beam([ParsePath.empty(context)])
def get_intermediate_beams(self, examples, verbose=False):
# Start with beam_size empty paths for each example
beams = [Beam([ParsePath.empty(ex.context)
for _ in xrange(self._config.beam_size)])
for ex in examples]
intermediates = [[] for _ in examples]
num_iterations = get_num_iterations(
self._config.iterations_per_utterance, examples)
if verbose:
iterations = verboserate(range(num_iterations),
desc='Batched REINFORCE')
else:
iterations = xrange(num_iterations)
for _ in iterations:
for ex_idx, beam in enumerate(beams):
intermediates[ex_idx].append(beam)
beams = self.advance(beams)
return beams, intermediates
def get_beams(self, examples, verbose=False):
terminated = [set() for _ in examples]
# initialize beams
beams = [[ParsePath.empty(ex.context)] for ex in examples]
# put all probability mass on the root
distributions = [[1] for _ in examples]
num_iterations = get_num_iterations(
self._config.iterations_per_utterance, examples)
iterations = xrange(num_iterations)
if verbose:
iterations = verboserate(
iterations, desc='Performing randomized search')
for _ in iterations:
terminated, beams, distributions = self.advance(
terminated, beams, distributions)
return [Beam(sorted(list(paths), key=self._ranker, reverse=True))
for paths in terminated]