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

remove theano dependency, using keras backend functions

This commit is contained in:
Chen Buskilla 2016-11-13 15:06:01 +02:00
parent a592075720
commit 738f38e8d6
1 changed files with 7 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
examples/keras_parikh_entailment/keras_decomposable_attention.py
View File

@ -6,7 +6,6 @@ from keras.layers import InputSpec, Layer, Input, Dense, merge
from keras.layers import Activation, Dropout, Embedding, TimeDistributed from keras.layers import Activation, Dropout, Embedding, TimeDistributed
from keras.layers import Bidirectional, GRU from keras.layers import Bidirectional, GRU
import keras.backend as K import keras.backend as K
import theano.tensor as T
from keras.models import Sequential, Model, model_from_json from keras.models import Sequential, Model, model_from_json
from keras.regularizers import l2 from keras.regularizers import l2
from keras.optimizers import Adam from keras.optimizers import Adam
@ -106,8 +105,8 @@ class _Attention(object):
def __call__(self, sent1, sent2): def __call__(self, sent1, sent2):
def _outer(AB): def _outer(AB):
att_ji = T.batched_dot(AB[1], AB[0].dimshuffle((0, 2, 1))) att_ji = K.batch_dot(AB[1], K.permute_dimensions(AB[0], (0, 2, 1)))
return att_ji.dimshuffle((0, 2, 1)) return K.permute_dimensions(att_ji,(0, 2, 1))
return merge( return merge(
@ -126,12 +125,12 @@ class _SoftAlignment(object):
att = attmat[0] att = attmat[0]
mat = attmat[1] mat = attmat[1]
if transpose: if transpose:
att = att.dimshuffle((0, 2, 1)) att = K.permute_dimensions(att,(0, 2, 1))
# 3d softmax # 3d softmax
e = K.exp(att - K.max(att, axis=-1, keepdims=True)) e = K.exp(att - K.max(att, axis=-1, keepdims=True))
s = K.sum(e, axis=-1, keepdims=True) s = K.sum(e, axis=-1, keepdims=True)
sm_att = e / s sm_att = e / s
return T.batched_dot(sm_att, mat) return K.batch_dot(sm_att, mat)
return merge([attention, sentence], mode=_normalize_attention, return merge([attention, sentence], mode=_normalize_attention,
output_shape=(self.max_length, self.nr_hidden)) # Shape: (i, n) output_shape=(self.max_length, self.nr_hidden)) # Shape: (i, n)
@ -191,7 +190,7 @@ class _GlobalSumPooling1D(Layer):
def call(self, x, mask=None): def call(self, x, mask=None):
if mask is not None: if mask is not None:
return K.sum(x * T.clip(mask, 0, 1), axis=1) return K.sum(x * K.clip(mask, 0, 1), axis=1)
else: else:
return K.sum(x, axis=1) return K.sum(x, axis=1)
@ -204,6 +203,7 @@ def test_build_model():
def test_fit_model(): def test_fit_model():
def _generate_X(nr_example, length, nr_vector): def _generate_X(nr_example, length, nr_vector):
X1 = numpy.ndarray((nr_example, length), dtype='int32') X1 = numpy.ndarray((nr_example, length), dtype='int32')
X1 *= X1 < nr_vector X1 *= X1 < nr_vector
@ -212,6 +212,7 @@ def test_fit_model():
X2 *= X2 < nr_vector X2 *= X2 < nr_vector
X2 *= 0 <= X2 X2 *= 0 <= X2
return [X1, X2] return [X1, X2]
def _generate_Y(nr_example, nr_class): def _generate_Y(nr_example, nr_class):
ys = numpy.zeros((nr_example, nr_class), dtype='int32') ys = numpy.zeros((nr_example, nr_class), dtype='int32')
for i in range(nr_example): for i in range(nr_example):