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

Make parser consistent if maxout==1

This commit is contained in:
Matthew Honnibal 2017-10-20 16:24:16 +02:00
parent 3faf9189a2
commit 1036798155
1 changed files with 6 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
spacy/syntax/nn_parser.pyx
View File

@ -136,7 +136,8 @@ cdef class precompute_hiddens:
return self.begin_update(X)[0] return self.begin_update(X)[0]
def begin_update(self, token_ids, drop=0.): def begin_update(self, token_ids, drop=0.):
cdef np.ndarray state_vector = numpy.zeros((token_ids.shape[0], self.nO*self.nP), dtype='f') cdef np.ndarray state_vector = numpy.zeros(
(token_ids.shape[0], self.nO, self.nP), dtype='f')
# This is tricky, but (assuming GPU available); # This is tricky, but (assuming GPU available);
# - Input to forward on CPU # - Input to forward on CPU
# - Output from forward on CPU # - Output from forward on CPU
@ -166,16 +167,13 @@ cdef class precompute_hiddens:
mask = state_vector >= 0. mask = state_vector >= 0.
state_vector *= mask state_vector *= mask
else: else:
state_vector = state_vector.reshape(
(state_vector.shape[0], self.nO, self.nP))
state_vector, mask = self.ops.maxout(state_vector) state_vector, mask = self.ops.maxout(state_vector)
def backprop_nonlinearity(d_best, sgd=None): def backprop_nonlinearity(d_best, sgd=None):
if self.nP == 1: if self.nP == 1:
return d_best * mask return d_best * mask
else: else:
d_vector = self.ops.backprop_maxout(d_best, mask, self.nP) return self.ops.backprop_maxout(d_best, mask, self.nP)
return d_vector.reshape((d_vector.shape[0], self.nO*self.nP))
return state_vector, backprop_nonlinearity return state_vector, backprop_nonlinearity
@ -266,8 +264,9 @@ cdef class Parser:
tok2vec = Tok2Vec(token_vector_width, embed_size, tok2vec = Tok2Vec(token_vector_width, embed_size,
pretrained_dims=cfg.get('pretrained_dims', 0)) pretrained_dims=cfg.get('pretrained_dims', 0))
tok2vec = chain(tok2vec, flatten) tok2vec = chain(tok2vec, flatten)
lower = PrecomputableAffine(hidden_width * parser_maxout_pieces, lower = PrecomputableAffine(hidden_width,
nF=cls.nr_feature, nI=token_vector_width) nF=cls.nr_feature, nI=token_vector_width,
nP=parser_maxout_pieces)
lower.nP = parser_maxout_pieces lower.nP = parser_maxout_pieces
with Model.use_device('cpu'): with Model.use_device('cpu'):