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

Fix gradient bug in parser

This commit is contained in:
Matthew Honnibal 2017-08-30 17:31:53 -05:00
parent 173089a45a
commit 4ceebde523
1 changed files with 10 additions and 12 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

20
spacy/syntax/nn_parser.pyx
View File

@ -36,7 +36,7 @@ from murmurhash.mrmr cimport hash64
from preshed.maps cimport MapStruct from preshed.maps cimport MapStruct
from preshed.maps cimport map_get from preshed.maps cimport map_get
from thinc.api import layerize, chain, noop, clone from thinc.api import layerize, chain, noop, clone, with_flatten
from thinc.neural import Model, Affine, ReLu, Maxout from thinc.neural import Model, Affine, ReLu, Maxout
from thinc.neural._classes.batchnorm import BatchNorm as BN from thinc.neural._classes.batchnorm import BatchNorm as BN
from thinc.neural._classes.selu import SELU from thinc.neural._classes.selu import SELU
@ -393,7 +393,7 @@ cdef class Parser:
tokvecs = self.model[0].ops.flatten(tokvecses) tokvecs = self.model[0].ops.flatten(tokvecses)
if USE_FINE_TUNE: if USE_FINE_TUNE:
tokvecs += self.model[0].ops.flatten(self.model[0]((docs, tokvecses))) tokvecs = self.model[0].ops.flatten(self.model[0]((docs, tokvecses)))
nr_state = len(docs) nr_state = len(docs)
nr_class = self.moves.n_moves nr_class = self.moves.n_moves
@ -453,7 +453,7 @@ cdef class Parser:
cdef StateClass stcls, output cdef StateClass stcls, output
tokvecs = self.model[0].ops.flatten(tokvecses) tokvecs = self.model[0].ops.flatten(tokvecses)
if USE_FINE_TUNE: if USE_FINE_TUNE:
tokvecs += self.model[0].ops.flatten(self.model[0]((docs, tokvecses))) tokvecs = self.model[0].ops.flatten(self.model[0]((docs, tokvecses)))
cuda_stream = get_cuda_stream() cuda_stream = get_cuda_stream()
state2vec, vec2scores = self.get_batch_model(len(docs), tokvecs, state2vec, vec2scores = self.get_batch_model(len(docs), tokvecs,
cuda_stream, 0.0) cuda_stream, 0.0)
@ -531,9 +531,8 @@ cdef class Parser:
docs = [docs] docs = [docs]
golds = [golds] golds = [golds]
if USE_FINE_TUNE: if USE_FINE_TUNE:
my_tokvecs, bp_my_tokvecs = self.model[0].begin_update(docs_tokvecs, drop=drop) tokvecs, bp_my_tokvecs = self.model[0].begin_update(docs_tokvecs, drop=drop)
my_tokvecs = self.model[0].ops.flatten(my_tokvecs) tokvecs = self.model[0].ops.flatten(tokvecs)
tokvecs += my_tokvecs
cuda_stream = get_cuda_stream() cuda_stream = get_cuda_stream()
@ -586,7 +585,7 @@ cdef class Parser:
backprops, sgd, cuda_stream) backprops, sgd, cuda_stream)
d_tokvecs = self.model[0].ops.unflatten(d_tokvecs, [len(d) for d in docs]) d_tokvecs = self.model[0].ops.unflatten(d_tokvecs, [len(d) for d in docs])
if USE_FINE_TUNE: if USE_FINE_TUNE:
bp_my_tokvecs(d_tokvecs, sgd=sgd) d_tokvecs = bp_my_tokvecs(d_tokvecs, sgd=sgd)
return d_tokvecs return d_tokvecs
def update_beam(self, docs_tokvecs, golds, width=None, density=None, def update_beam(self, docs_tokvecs, golds, width=None, density=None,
@ -606,9 +605,8 @@ cdef class Parser:
assert min(lengths) >= 1 assert min(lengths) >= 1
tokvecs = self.model[0].ops.flatten(tokvecs) tokvecs = self.model[0].ops.flatten(tokvecs)
if USE_FINE_TUNE: if USE_FINE_TUNE:
my_tokvecs, bp_my_tokvecs = self.model[0].begin_update(docs_tokvecs, drop=drop) tokvecs, bp_my_tokvecs = self.model[0].begin_update(docs_tokvecs, drop=drop)
my_tokvecs = self.model[0].ops.flatten(my_tokvecs) tokvecs = self.model[0].ops.flatten(tokvecs)
tokvecs += my_tokvecs
states = self.moves.init_batch(docs) states = self.moves.init_batch(docs)
for gold in golds: for gold in golds:
@ -642,7 +640,7 @@ cdef class Parser:
self._make_updates(d_tokvecs, backprop_lower, sgd, cuda_stream) self._make_updates(d_tokvecs, backprop_lower, sgd, cuda_stream)
d_tokvecs = self.model[0].ops.unflatten(d_tokvecs, lengths) d_tokvecs = self.model[0].ops.unflatten(d_tokvecs, lengths)
if USE_FINE_TUNE: if USE_FINE_TUNE:
bp_my_tokvecs(d_tokvecs, sgd=sgd) d_tokvecs = bp_my_tokvecs(d_tokvecs, sgd=sgd)
return d_tokvecs return d_tokvecs
def _init_gold_batch(self, whole_docs, whole_golds): def _init_gold_batch(self, whole_docs, whole_golds):