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

Bug fixes

This commit is contained in:
Matthew Honnibal 2017-05-05 20:09:50 +02:00
parent ef4fa594aa
commit efe9630e1c
1 changed files with 37 additions and 19 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

54
spacy/_ml.py
View File

@ -1,8 +1,7 @@
from thinc.api import layerize, chain, clone, concatenate, add from thinc.api import layerize, chain, clone
from thinc.neural import Model, Maxout, Softmax from thinc.neural import Model, Maxout, Softmax
from thinc.neural._classes.static_vectors import StaticVectors
from thinc.neural._classes.hash_embed import HashEmbed from thinc.neural._classes.hash_embed import HashEmbed
from thinc.neural._classes.convolution import ExtractWindow from .attrs import TAG, DEP
def get_col(idx): def get_col(idx):
@ -17,19 +16,19 @@ def build_model(state2vec, width, depth, nr_class):
return model return model
def build_parser_state2vec(tag_vectors, dep_vectors, **cfg): def build_parser_state2vec(width, nr_vector=1000, nF=1, nB=0, nS=1, nL=2, nR=2):
embed_tags = _reshape(chain(get_col(0), tag_vectors)) embed_tags = _reshape(chain(get_col(0), HashEmbed(width, nr_vector)))
embed_deps = _reshape(chain(get_col(1), dep_vectors)) embed_deps = _reshape(chain(get_col(1), HashEmbed(width, nr_vector)))
ops = embed_tags.ops
attr_names = ops.asarray([TAG, DEP], dtype='i') attr_names = ops.asarray([TAG, DEP], dtype='i')
extract = build_feature_extractor(attr_names, nF, nB, nS, nL, nR)
def forward(states, drop=0.): def forward(states, drop=0.):
n_tokens = state.nr_context_tokens(nF, nB, nS, nL, nR) tokens, attr_vals, tokvecs = extract(states)
for i, state in enumerate(states): tagvecs, bp_tagvecs = embed_deps.begin_update(attr_vals, drop=drop)
state.set_context_tokens(tokens[i], nF, nB, nS, nL, nR) depvecs, bp_depvecs = embed_tags.begin_update(attr_vals, drop=drop)
state.set_attributes(features[i], tokens[i], attr_names)
state.set_token_vectors(token_vectors[i], tokens[i])
tagvecs, bp_tag_vecs = embed_deps.begin_update(attr_vals, drop=drop) tokvecs = tokvecs.reshape((tokvecs.shape[0], tokvecs.shape[1] *
depvecs, bp_dep_vecs = embed_tags.begin_update(attr_vals, drop=drop) tokvecs.shape[2]))
vector = ops.concatenate((tagvecs, depvecs, tokvecs)) vector = ops.concatenate((tagvecs, depvecs, tokvecs))
@ -38,6 +37,7 @@ def build_parser_state2vec(tag_vectors, dep_vectors, **cfg):
d_depvecs, d_tagvecs, d_tokvecs = ops.backprop_concatenate(d_vector, shapes) d_depvecs, d_tagvecs, d_tokvecs = ops.backprop_concatenate(d_vector, shapes)
bp_tagvecs(d_tagvecs) bp_tagvecs(d_tagvecs)
bp_depvecs(d_depvecs) bp_depvecs(d_depvecs)
d_tokvecs = d_tokvecs.reshape((len(states), tokens.shape[1], tokvecs.shape[2]))
return (d_tokvecs, tokens) return (d_tokvecs, tokens)
return vector, backward return vector, backward
model = layerize(forward) model = layerize(forward)
@ -45,11 +45,31 @@ def build_parser_state2vec(tag_vectors, dep_vectors, **cfg):
return model return model
def build_feature_extractor(attr_names, nF, nB, nS, nL, nR):
def forward(states, drop=0.):
ops = model.ops
n_tokens = states[0].nr_context_tokens(nF, nB, nS, nL, nR)
vector_length = states[0].token_vector_length
tokens = ops.allocate((len(states), n_tokens), dtype='i')
features = ops.allocate((len(states), n_tokens, attr_names.shape[0]), dtype='i')
tokvecs = ops.allocate((len(states), n_tokens, vector_length), dtype='f')
for i, state in enumerate(states):
state.set_context_tokens(tokens[i], nF, nB, nS, nL, nR)
state.set_attributes(features[i], tokens[i], attr_names)
state.set_token_vectors(tokvecs[i], tokens[i])
def backward(d_features, sgd=None):
return d_features
return (tokens, features, tokvecs), backward
model = layerize(forward)
return model
def _reshape(layer): def _reshape(layer):
def forward(X, drop=0.): def forward(X, drop=0.):
Xh = X.reshape((X.shape[0] * X.shape[1], X.shape[2])) Xh = X.reshape((X.shape[0] * X.shape[1], X.shape[2]))
yh, bp_yh = layer.begin_update(Xh, drop=drop) yh, bp_yh = layer.begin_update(Xh, drop=drop)
n = X.shape[0] n = X.shape[0]
old_shape = X.shape
def backward(d_y, sgd=None): def backward(d_y, sgd=None):
d_yh = d_y.reshape((n, d_y.size / n)) d_yh = d_y.reshape((n, d_y.size / n))
d_Xh = bp_yh(d_yh, sgd) d_Xh = bp_yh(d_yh, sgd)
@ -59,7 +79,9 @@ def _reshape(layer):
model._layers.append(layer) model._layers.append(layer)
return model return model
#from thinc.api import layerize, chain, clone, concatenate, add
# from thinc.neural._classes.convolution import ExtractWindow
# from thinc.neural._classes.static_vectors import StaticVectors
#def build_tok2vec(lang, width, depth, embed_size, cols): #def build_tok2vec(lang, width, depth, embed_size, cols):
# with Model.define_operators({'>>': chain, '|': concatenate, '**': clone}): # with Model.define_operators({'>>': chain, '|': concatenate, '**': clone}):
@ -73,7 +95,3 @@ def _reshape(layer):
# >> (ExtractWindow(nW=1) >> Maxout(width, width*3)) ** depth # >> (ExtractWindow(nW=1) >> Maxout(width, width*3)) ** depth
# ) # )
# return tok2vec # return tok2vec
if __name__ == '__main__':
test_build_model()