def score_batches(coord: Coordinator, pp: ParallelParse):
t_sys = pp.transition_system
while True:
batch = pp.batches.get()
# Score states and move to next states, based on the scoring
batch_scores = pp.score_batch(batch)
pp.advance_batch(batch, batch_scores)
# Put non-finished sentences back in scoring queue
queue_again = [sent_id for sent_id in batch.ids
if not t_sys.is_final(pp.states[sent_id])]
pp.num_left -= (len(batch.ids) - len(queue_again))
for sent_id in queue_again:
pp.needs_scoring.put(sent_id)
# Wake the batch prepare thread up and give it a chance to enqueue
# a smaller batch if need be
if len(queue_again) == 0 and pp.num_left > 0:
pp.needs_scoring.put(None)
# Our work here is done. Stop the other thread.
if pp.num_left == 0:
coord.request_stop()
pp.needs_scoring.put(None)
break
def _run_threads(self):
# Setup threads
coord = Coordinator()
t_prepare_batch = Thread(target=prepare_batches, args=[coord, self])
t_score_batch = Thread(target=score_batches, args=[coord, self])
threads = [t_prepare_batch, t_score_batch]
for t in threads:
t.start()
coord.join(threads)
def prepare_batches(coord: Coordinator, pp: ParallelParse):
def build_batch(sent_ids):
sents = [pp.sentences[sent_id] for sent_id in sent_ids]
states = [pp.states[sent_id] for sent_id in sent_ids]
allowed_list = pp.allowed_batch(states)
return SentenceBatch(ids=sent_ids,
sents=sents,
states=states,
feed_dict=pp.prepare_feed(sents, states),
action_costs_list=pp.action_costs(states, sents, allowed_list),
allowed_list=allowed_list
)
sent_ids = []
while True:
item = pp.needs_scoring.get()
if item is None:
if coord.should_stop():
break
else:
if len(sent_ids) >= pp.num_left:
pp.batches.put(build_batch(sent_ids))
sent_ids = []
else:
sent_ids.append(item)
if len(sent_ids) >= min(pp.batch_size, pp.num_left):
pp.batches.put(build_batch(sent_ids))
sent_ids = []
