explosion/spaCy
1
1
Fork 0
mirror of https://github.com/explosion/spaCy.git synced 2024-09-21 19:39:13 +03:00
Code Issues Packages Projects Releases Wiki Activity

Fix beam parsing

Browse Source
This commit is contained in:
Matthew Honnibal 2018-05-08 02:53:24 +02:00
parent 99649d114d
commit c49e44349a
3 changed files with 62 additions and 27 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
spacy/syntax/_beam_utils.pyx
View File

@ -59,7 +59,7 @@ cdef class ParserBeam(object):
cdef public object dones
def __init__(self, TransitionSystem moves, states, golds,
int width, float density):
int width, float density=0.):
self.moves = moves
self.states = states
self.golds = golds
@ -133,8 +133,12 @@ cdef class ParserBeam(object):
self.moves.set_costs(beam.is_valid[i], beam.costs[i],
state, gold)
if follow_gold:
min_cost = 0
for j in range(beam.nr_class):
if beam.costs[i][j] >= 1:
if beam.costs[i][j] < min_cost:
min_cost = beam.costs[i][j]
for j in range(beam.nr_class):
if beam.costs[i][j] > min_cost:
beam.is_valid[i][j] = 0

7
spacy/syntax/_parser_model.pyx
View File

@ -266,13 +266,14 @@ class ParserStepModel(Model):
def get_token_ids(self, batch):
states = _beam_utils.collect_states(batch)
cdef StateClass state
states = [state for state in states if not state.is_final()]
cdef np.ndarray ids = numpy.zeros((len(states), self.state2vec.nF),
dtype='i', order='C')
ids.fill(-1)
c_ids = <int*>ids.data
cdef StateClass state
for state in states:
if not state.c.is_final():
state.c.set_context_tokens(c_ids, ids.shape[1])
state.c.set_context_tokens(c_ids, ids.shape[1])
c_ids += ids.shape[1]
return ids

74
spacy/syntax/nn_parser.pyx
View File

@ -208,31 +208,62 @@ cdef class Parser:
for doc in batch_in_order:
yield doc
def predict(self, docs, beam_width=1):
def predict(self, docs, beam_width=1, drop=0.):
if isinstance(docs, Doc):
docs = [docs]
if beam_width < 2:
return self.greedy_parse(docs, drop=drop)
else:
return self.beam_parse(docs, beam_width=beam_width, drop=drop)
def greedy_parse(self, docs, drop=0.):
cdef vector[StateC*] states
cdef StateClass state
model = self.model(docs)
if beam_width == 1:
batch = self.moves.init_batch(docs)
weights = get_c_weights(model)
for state in batch:
if not state.is_final():
states.push_back(state.c)
sizes = get_c_sizes(model, states.size())
with nogil:
self._parseC(&states[0],
weights, sizes)
else:
batch = self.moves.init_beams(docs, beam_width)
unfinished = list(batch)
while unfinished:
scores = model.predict(unfinished)
unfinished = self.transition_beams(batch, scores)
batch = self.moves.init_batch(docs)
weights = get_c_weights(model)
for state in batch:
if not state.is_final():
states.push_back(state.c)
sizes = get_c_sizes(model, states.size())
with nogil:
self._parseC(&states[0],
weights, sizes)
return batch
def beam_parse(self, docs, int beam_width=3, float drop=0.):
cdef Beam beam
cdef Doc doc
cdef np.ndarray token_ids
model = self.model(docs)
beams = self.moves.init_beams(docs, beam_width)
token_ids = numpy.zeros((len(docs) * beam_width, self.nr_feature),
dtype='i', order='C')
cdef int* c_ids
cdef int nr_feature = self.nr_feature
cdef int n_states
model = self.model(docs)
todo = [beam for beam in beams if not beam.is_done]
while todo:
token_ids.fill(-1)
c_ids = <int*>token_ids.data
n_states = 0
for beam in todo:
for i in range(beam.size):
state = <StateC*>beam.at(i)
# This way we avoid having to score finalized states
# We do have to take care to keep indexes aligned, though
if not state.is_final():
state.set_context_tokens(c_ids, nr_feature)
c_ids += nr_feature
n_states += 1
if n_states == 0:
break
vectors = model.state2vec(token_ids[:n_states])
scores = model.vec2scores(vectors)
todo = self.transition_beams(todo, scores)
return beams
cdef void _parseC(self, StateC** states,
WeightsC weights, SizesC sizes) nogil:
cdef int i, j
@ -325,7 +356,7 @@ cdef class Parser:
beam_update_prob = 1-self.cfg.get('beam_update_prob', 0.5)
if self.cfg.get('beam_width', 1) >= 2 and numpy.random.random() >= beam_update_prob:
return self.update_beam(docs, golds,
self.cfg['beam_width'], self.cfg['beam_density'],
self.cfg['beam_width'],
drop=drop, sgd=sgd, losses=losses)
# Chop sequences into lengths of this many transitions, to make the
# batch uniform length.
@ -352,12 +383,11 @@ cdef class Parser:
def update_beam(self, docs, golds, width, drop=0., sgd=None, losses=None):
lengths = [len(d) for d in docs]
states = self.moves.init_batch(docs)
for gold in golds:
self.moves.preprocess_gold(gold)
cut_gold = numpy.random.choice(range(20, 100))
states, golds, max_steps = self._init_gold_batch(docs, golds, max_length=cut_gold)
model, finish_update = self.model.begin_update(docs, drop=drop)
states_d_scores, backprops, beams = _beam_utils.update_beam(
self.moves, self.nr_feature, 500, states, golds, model.state2vec,
self.moves, self.nr_feature, max_steps, states, golds, model.state2vec,
model.vec2scores, width, drop=drop, losses=losses)
for i, d_scores in enumerate(states_d_scores):
ids, bp_vectors, bp_scores = backprops[i]