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

Support subword_features and conv_depth params in Tok2Vec

This commit is contained in:
Matthew Honnibal 2018-08-27 01:50:48 +02:00
parent 9c33d4d1df
commit 8051136d70
1 changed files with 25 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

36
spacy/_ml.py
View File

@ -248,27 +248,39 @@ def link_vectors_to_models(vocab):
def Tok2Vec(width, embed_size, **kwargs): def Tok2Vec(width, embed_size, **kwargs):
pretrained_vectors = kwargs.get('pretrained_vectors', None) pretrained_vectors = kwargs.get('pretrained_vectors', None)
cnn_maxout_pieces = kwargs.get('cnn_maxout_pieces', 2) cnn_maxout_pieces = kwargs.get('cnn_maxout_pieces', 2)
subword_features = kwargs.get('subword_features', True)
conv_depth = kwargs.get('conv_depth', 4)
cols = [ID, NORM, PREFIX, SUFFIX, SHAPE, ORTH] cols = [ID, NORM, PREFIX, SUFFIX, SHAPE, ORTH]
with Model.define_operators({'>>': chain, '|': concatenate, '**': clone, with Model.define_operators({'>>': chain, '|': concatenate, '**': clone,
'+': add, '*': reapply}): '+': add, '*': reapply}):
norm = HashEmbed(width, embed_size, column=cols.index(NORM), norm = HashEmbed(width, embed_size, column=cols.index(NORM),
name='embed_norm') name='embed_norm')
prefix = HashEmbed(width, embed_size//2, column=cols.index(PREFIX), if subword_features:
name='embed_prefix') prefix = HashEmbed(width, embed_size//2, column=cols.index(PREFIX),
suffix = HashEmbed(width, embed_size//2, column=cols.index(SUFFIX), name='embed_prefix')
name='embed_suffix') suffix = HashEmbed(width, embed_size//2, column=cols.index(SUFFIX),
shape = HashEmbed(width, embed_size//2, column=cols.index(SHAPE), name='embed_suffix')
name='embed_shape') shape = HashEmbed(width, embed_size//2, column=cols.index(SHAPE),
name='embed_shape')
else:
prefix, suffix, shape = None
if pretrained_vectors is not None: if pretrained_vectors is not None:
glove = StaticVectors(pretrained_vectors, width, column=cols.index(ID)) glove = StaticVectors(pretrained_vectors, width, column=cols.index(ID))
embed = uniqued( if subword_features:
(glove | norm | prefix | suffix | shape) embed = uniqued(
>> LN(Maxout(width, width*5, pieces=3)), column=cols.index(ORTH)) (glove | norm | prefix | suffix | shape)
else: >> LN(Maxout(width, width*5, pieces=3)), column=cols.index(ORTH))
else:
embed = uniqued(
(glove | norm)
>> LN(Maxout(width, width*2, pieces=3)), column=cols.index(ORTH))
elif subword_features:
embed = uniqued( embed = uniqued(
(norm | prefix | suffix | shape) (norm | prefix | suffix | shape)
>> LN(Maxout(width, width*4, pieces=3)), column=cols.index(ORTH)) >> LN(Maxout(width, width*4, pieces=3)), column=cols.index(ORTH))
else:
embed = norm
convolution = Residual( convolution = Residual(
ExtractWindow(nW=1) ExtractWindow(nW=1)
@ -279,7 +291,7 @@ def Tok2Vec(width, embed_size, **kwargs):
FeatureExtracter(cols) FeatureExtracter(cols)
>> with_flatten( >> with_flatten(
embed embed
>> convolution ** 4, pad=4 >> convolution ** conv_depth, pad=conv_depth
) )
) )
# Work around thinc API limitations :(. TODO: Revise in Thinc 7 # Work around thinc API limitations :(. TODO: Revise in Thinc 7
@ -424,11 +436,13 @@ def build_tagger_model(nr_class, **cfg):
else: else:
token_vector_width = util.env_opt('token_vector_width', 128) token_vector_width = util.env_opt('token_vector_width', 128)
pretrained_vectors = cfg.get('pretrained_vectors') pretrained_vectors = cfg.get('pretrained_vectors')
subword_features = cfg.get('subword_features', True)
with Model.define_operators({'>>': chain, '+': add}): with Model.define_operators({'>>': chain, '+': add}):
if 'tok2vec' in cfg: if 'tok2vec' in cfg:
tok2vec = cfg['tok2vec'] tok2vec = cfg['tok2vec']
else: else:
tok2vec = Tok2Vec(token_vector_width, embed_size, tok2vec = Tok2Vec(token_vector_width, embed_size,
subword_features=subword_features,
pretrained_vectors=pretrained_vectors) pretrained_vectors=pretrained_vectors)
softmax = with_flatten(Softmax(nr_class, token_vector_width)) softmax = with_flatten(Softmax(nr_class, token_vector_width))
model = ( model = (