Rooba
/
spaCy
mirror of https://github.com/explosion/spaCy.git
1
1
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Pass tokvecs through as a list, instead of concatenated. Also fix padding

This commit is contained in:
Matthew Honnibal 2017-05-20 13:23:05 -05:00
parent d52b65aec2
commit 3b7c108246
4 changed files with 30 additions and 29 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
spacy/_ml.py
View File

@ -134,13 +134,14 @@ def Tok2Vec(width, embed_size, preprocess=None):
shape = get_col(cols.index(SHAPE)) >> HashEmbed(width, embed_size//2) shape = get_col(cols.index(SHAPE)) >> HashEmbed(width, embed_size//2)
tok2vec = ( tok2vec = (
flatten with_flatten(
>> (lower | prefix | suffix | shape ) (lower | prefix | suffix | shape )
>> Maxout(width, width*4, pieces=3) >> Maxout(width, width*4, pieces=3)
>> Residual(ExtractWindow(nW=1) >> Maxout(width, width*3)) >> Residual(ExtractWindow(nW=1) >> Maxout(width, width*3))
>> Residual(ExtractWindow(nW=1) >> Maxout(width, width*3)) >> Residual(ExtractWindow(nW=1) >> Maxout(width, width*3))
>> Residual(ExtractWindow(nW=1) >> Maxout(width, width*3)) >> Residual(ExtractWindow(nW=1) >> Maxout(width, width*3))
>> Residual(ExtractWindow(nW=1) >> Maxout(width, width*3)) >> Residual(ExtractWindow(nW=1) >> Maxout(width, width*3)),
pad=4, ndim=5)
) )
if preprocess not in (False, None): if preprocess not in (False, None):
tok2vec = preprocess >> tok2vec tok2vec = preprocess >> tok2vec

6
spacy/language.py
View File

@ -179,10 +179,10 @@ class Language(object):
tok2vec = self.pipeline[0] tok2vec = self.pipeline[0]
feats = tok2vec.doc2feats(docs) feats = tok2vec.doc2feats(docs)
for proc in self.pipeline[1:]: for proc in self.pipeline[1:]:
tokvecs, bp_tokvecs = tok2vec.model.begin_update(feats, drop=drop)
grads = {} grads = {}
d_tokvecs = proc.update((docs, tokvecs), golds, sgd=get_grads, drop=drop) tokvecses, bp_tokvecses = tok2vec.model.begin_update(feats, drop=drop)
bp_tokvecs(d_tokvecs, sgd=get_grads) d_tokvecses = proc.update((docs, tokvecses), golds, sgd=get_grads, drop=drop)
bp_tokvecses(d_tokvecses, sgd=get_grads)
if sgd is not None: if sgd is not None:
for key, (W, dW) in grads.items(): for key, (W, dW) in grads.items():
# TODO: Unhack this when thinc improves # TODO: Unhack this when thinc improves

30
spacy/pipeline.pyx
View File

@ -10,7 +10,7 @@ cimport numpy as np
import cytoolz import cytoolz
import util import util
from thinc.api import add, layerize, chain, clone, concatenate from thinc.api import add, layerize, chain, clone, concatenate, with_flatten
from thinc.neural import Model, Maxout, Softmax, Affine from thinc.neural import Model, Maxout, Softmax, Affine
from thinc.neural._classes.hash_embed import HashEmbed from thinc.neural._classes.hash_embed import HashEmbed
from thinc.neural.util import to_categorical from thinc.neural.util import to_categorical
@ -52,16 +52,16 @@ class TokenVectorEncoder(object):
self.doc2feats = doc2feats() self.doc2feats = doc2feats()
self.model = model self.model = model
def __call__(self, docs, state=None): def __call__(self, docs):
if isinstance(docs, Doc): if isinstance(docs, Doc):
docs = [docs] docs = [docs]
tokvecs = self.predict(docs) tokvecses = self.predict(docs)
self.set_annotations(docs, tokvecs) self.set_annotations(docs, tokvecses)
def pipe(self, stream, batch_size=128, n_threads=-1): def pipe(self, stream, batch_size=128, n_threads=-1):
for docs in cytoolz.partition_all(batch_size, stream): for docs in cytoolz.partition_all(batch_size, stream):
tokvecs = self.predict(docs) tokvecses = self.predict(docs)
self.set_annotations(docs, tokvecs) self.set_annotations(docs, tokvecses)
yield from docs yield from docs
def predict(self, docs): def predict(self, docs):
@ -69,11 +69,9 @@ class TokenVectorEncoder(object):
tokvecs = self.model(feats) tokvecs = self.model(feats)
return tokvecs return tokvecs
def set_annotations(self, docs, tokvecs): def set_annotations(self, docs, tokvecses):
start = 0 for doc, tokvecs in zip(docs, tokvecses):
for doc in docs: doc.tensor = tokvecs
doc.tensor = tokvecs[start : start + len(doc)]
start += len(doc)
def begin_update(self, docs, drop=0.): def begin_update(self, docs, drop=0.):
if isinstance(docs, Doc): if isinstance(docs, Doc):
@ -136,7 +134,7 @@ class NeuralTagger(object):
docs, tokvecs = docs_tokvecs docs, tokvecs = docs_tokvecs
if self.model.nI is None: if self.model.nI is None:
self.model.nI = tokvecs.shape[1] self.model.nI = tokvecs[0].shape[1]
tag_scores, bp_tag_scores = self.model.begin_update(tokvecs, drop=drop) tag_scores, bp_tag_scores = self.model.begin_update(tokvecs, drop=drop)
loss, d_tag_scores = self.get_loss(docs, golds, tag_scores) loss, d_tag_scores = self.get_loss(docs, golds, tag_scores)
@ -146,6 +144,7 @@ class NeuralTagger(object):
return d_tokvecs return d_tokvecs
def get_loss(self, docs, golds, scores): def get_loss(self, docs, golds, scores):
scores = self.model.ops.flatten(scores)
tag_index = {tag: i for i, tag in enumerate(self.vocab.morphology.tag_names)} tag_index = {tag: i for i, tag in enumerate(self.vocab.morphology.tag_names)}
cdef int idx = 0 cdef int idx = 0
@ -161,7 +160,7 @@ class NeuralTagger(object):
correct = self.model.ops.xp.array(correct, dtype='i') correct = self.model.ops.xp.array(correct, dtype='i')
d_scores = scores - to_categorical(correct, nb_classes=scores.shape[1]) d_scores = scores - to_categorical(correct, nb_classes=scores.shape[1])
loss = (d_scores**2).sum() loss = (d_scores**2).sum()
d_scores = self.model.ops.asarray(d_scores, dtype='f') d_scores = self.model.ops.unflatten(d_scores, [len(d) for d in docs])
return float(loss), d_scores return float(loss), d_scores
def begin_training(self, gold_tuples, pipeline=None): def begin_training(self, gold_tuples, pipeline=None):
@ -179,9 +178,8 @@ class NeuralTagger(object):
vocab.morphology = Morphology(vocab.strings, new_tag_map, vocab.morphology = Morphology(vocab.strings, new_tag_map,
vocab.morphology.lemmatizer) vocab.morphology.lemmatizer)
token_vector_width = pipeline[0].model.nO token_vector_width = pipeline[0].model.nO
self.model = rebatch(1024, Softmax(self.vocab.morphology.n_tags, self.model = with_flatten(
token_vector_width)) Softmax(self.vocab.morphology.n_tags, token_vector_width))
#self.model = Softmax(self.vocab.morphology.n_tags)
def use_params(self, params): def use_params(self, params):
with self.model.use_params(params): with self.model.use_params(params):

8
spacy/syntax/nn_parser.pyx
View File

@ -311,7 +311,8 @@ cdef class Parser:
return states return states
def update(self, docs_tokvecs, golds, drop=0., sgd=None): def update(self, docs_tokvecs, golds, drop=0., sgd=None):
docs, tokvecs = docs_tokvecs docs, tokvec_lists = docs_tokvecs
tokvecs = self.model[0].ops.flatten(tokvec_lists)
if isinstance(docs, Doc) and isinstance(golds, GoldParse): if isinstance(docs, Doc) and isinstance(golds, GoldParse):
docs = [docs] docs = [docs]
golds = [golds] golds = [golds]
@ -324,7 +325,8 @@ cdef class Parser:
state2vec, vec2scores = self.get_batch_model(len(states), tokvecs, cuda_stream, state2vec, vec2scores = self.get_batch_model(len(states), tokvecs, cuda_stream,
drop) drop)
todo = [(s, g) for s, g in zip(states, golds) if not s.is_final()] todo = [(s, g) for (s, g) in zip(states, golds)
if not s.is_final()]
backprops = [] backprops = []
cdef float loss = 0. cdef float loss = 0.
@ -365,7 +367,7 @@ cdef class Parser:
else: else:
xp.add.at(d_tokvecs, xp.add.at(d_tokvecs,
token_ids, d_state_features * active_feats) token_ids, d_state_features * active_feats)
return d_tokvecs return self.model[0].ops.unflatten(d_tokvecs, [len(d) for d in docs])
def get_batch_model(self, batch_size, tokvecs, stream, dropout): def get_batch_model(self, batch_size, tokvecs, stream, dropout):
lower, upper = self.model lower, upper = self.model