explosion/spaCy
1
1
Fork 0
mirror of https://github.com/explosion/spaCy.git synced 2025-01-26 09:14:32 +03:00
Code Issues Packages Projects Releases Wiki Activity

Unhack beam parsing, moving it under options instead of global flags

Browse Source
This commit is contained in:
Matthew Honnibal 2017-08-18 13:31:15 -05:00
parent 0209a06b4e
commit f75420ae79
2 changed files with 41 additions and 29 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

33
spacy/syntax/_beam_utils.pyx
View File

@ -49,7 +49,7 @@ cdef class ParserBeam(object):
cdef public object dones
def __init__(self, TransitionSystem moves, states, golds,
int width=4, float density=0.001):
int width, float density):
self.moves = moves
self.states = states
self.golds = golds
@ -89,7 +89,10 @@ cdef class ParserBeam(object):
self._set_scores(beam, scores[i])
if self.golds is not None:
self._set_costs(beam, self.golds[i], follow_gold=follow_gold)
beam.advance(_transition_state, _hash_state, <void*>self.moves.c)
if follow_gold:
beam.advance(_transition_state, NULL, <void*>self.moves.c)
else:
beam.advance(_transition_state, _hash_state, <void*>self.moves.c)
beam.check_done(_check_final_state, NULL)
if beam.is_done and self.golds is not None:
for j in range(beam.size):
@ -145,15 +148,16 @@ def get_token_ids(states, int n_tokens):
nr_update = 0
def update_beam(TransitionSystem moves, int nr_feature, int max_steps,
states, tokvecs, golds,
state2vec, vec2scores, drop=0., sgd=None,
losses=None, int width=4, float density=0.001):
state2vec, vec2scores,
int width, float density,
sgd=None, losses=None, drop=0.):
global nr_update
cdef MaxViolation violn
nr_update += 1
pbeam = ParserBeam(moves, states, golds,
width=width, density=density)
gbeam = ParserBeam(moves, states, golds,
width=width, density=density)
width=width, density=0.0)
cdef StateClass state
beam_maps = []
backprops = []
@ -194,13 +198,13 @@ def update_beam(TransitionSystem moves, int nr_feature, int max_steps,
violn.check_crf(pbeam[i], gbeam[i])
histories = []
losses = []
for i, violn in enumerate(violns):
if violn.cost < 1:
histories.append([])
losses.append([])
else:
for violn in violns:
if violn.p_hist:
histories.append(violn.p_hist + violn.g_hist)
losses.append(violn.p_probs + violn.g_probs)
else:
histories.append([])
losses.append([])
states_d_scores = get_gradient(moves.n_moves, beam_maps, histories, losses)
return states_d_scores, backprops[:len(states_d_scores)]
@ -215,10 +219,6 @@ def get_states(pbeams, gbeams, beam_map, nr_update):
for eg_id, (pbeam, gbeam) in enumerate(zip(pbeams, gbeams)):
p_indices.append([])
g_indices.append([])
if pbeam.loss > 0 and pbeam.min_score > (gbeam.score + numpy.sqrt(nr_update)):
pbeams.dones[eg_id] = True
gbeams.dones[eg_id] = True
continue
for i in range(pbeam.size):
state = <StateClass>pbeam.at(i)
if not state.is_final():
@ -269,9 +269,12 @@ def get_gradient(nr_class, beam_maps, histories, losses):
assert len(histories) == len(losses)
for eg_id, hists in enumerate(histories):
for loss, hist in zip(losses[eg_id], hists):
if abs(loss) == 0.0 or numpy.isnan(loss):
if loss == 0.0 or numpy.isnan(loss):
continue
key = tuple([eg_id])
# Adjust loss for length
avg_loss = loss / len(hist)
loss += avg_loss * (nr_step - len(hist))
for j, clas in enumerate(hist):
i = beam_maps[j][key]
# In step j, at state i action clas

37
spacy/syntax/nn_parser.pyx
View File

@ -67,7 +67,6 @@ from ..attrs cimport ID, TAG, DEP, ORTH, NORM, PREFIX, SUFFIX, TAG
from . import _beam_utils
USE_FINE_TUNE = True
BEAM_PARSE = True
def get_templates(*args, **kwargs):
return []
@ -299,6 +298,10 @@ cdef class Parser:
self.moves = self.TransitionSystem(self.vocab.strings, {})
else:
self.moves = moves
if 'beam_width' not in cfg:
cfg['beam_width'] = util.env_opt('beam_width', 1)
if 'beam_density' not in cfg:
cfg['beam_density'] = util.env_opt('beam_density', 0.0)
self.cfg = cfg
if 'actions' in self.cfg:
for action, labels in self.cfg.get('actions', {}).items():
@ -321,9 +324,7 @@ cdef class Parser:
if beam_width is None:
beam_width = self.cfg.get('beam_width', 1)
if beam_density is None:
beam_density = self.cfg.get('beam_density', 0.001)
if BEAM_PARSE:
beam_width = 16
beam_density = self.cfg.get('beam_density', 0.0)
cdef Beam beam
if beam_width == 1:
states = self.parse_batch([doc], [doc.tensor])
@ -339,7 +340,7 @@ cdef class Parser:
return output
def pipe(self, docs, int batch_size=1000, int n_threads=2,
beam_width=1, beam_density=0.001):
beam_width=None, beam_density=None):
"""
Process a stream of documents.
@ -351,8 +352,10 @@ cdef class Parser:
The number of threads with which to work on the buffer in parallel.
Yields (Doc): Documents, in order.
"""
if BEAM_PARSE:
beam_width = 16
if beam_width is None:
beam_width = self.cfg.get('beam_width', 1)
if beam_density is None:
beam_density = self.cfg.get('beam_density', 0.0)
cdef Doc doc
cdef Beam beam
for docs in cytoolz.partition_all(batch_size, docs):
@ -430,7 +433,7 @@ cdef class Parser:
next_step.push_back(st)
return states
def beam_parse(self, docs, tokvecses, int beam_width=16, float beam_density=0.001):
def beam_parse(self, docs, tokvecses, int beam_width=3, float beam_density=0.001):
cdef Beam beam
cdef np.ndarray scores
cdef Doc doc
@ -480,9 +483,10 @@ cdef class Parser:
return beams
def update(self, docs_tokvecs, golds, drop=0., sgd=None, losses=None):
if BEAM_PARSE and numpy.random.random() >= 0.5:
return self.update_beam(docs_tokvecs, golds, drop=drop, sgd=sgd,
losses=losses)
if self.cfg.get('beam_width', 1) >= 2 and numpy.random.random() >= 0.5:
return self.update_beam(docs_tokvecs, golds,
self.cfg['beam_width'], self.cfg['beam_density'],
drop=drop, sgd=sgd, losses=losses)
if losses is not None and self.name not in losses:
losses[self.name] = 0.
docs, tokvec_lists = docs_tokvecs
@ -548,7 +552,12 @@ cdef class Parser:
bp_my_tokvecs(d_tokvecs, sgd=sgd)
return d_tokvecs
def update_beam(self, docs_tokvecs, golds, drop=0., sgd=None, losses=None):
def update_beam(self, docs_tokvecs, golds, width=None, density=None,
drop=0., sgd=None, losses=None):
if width is None:
width = self.cfg.get('beam_width', 2)
if density is None:
density = self.cfg.get('beam_density', 0.0)
if losses is not None and self.name not in losses:
losses[self.name] = 0.
docs, tokvecs = docs_tokvecs
@ -570,8 +579,8 @@ cdef class Parser:
states_d_scores, backprops = _beam_utils.update_beam(self.moves, self.nr_feature, 500,
states, tokvecs, golds,
state2vec, vec2scores,
drop, sgd, losses,
width=16)
width, density,
sgd=sgd, drop=drop, losses=losses)
backprop_lower = []
for i, d_scores in enumerate(states_d_scores):
if losses is not None: