-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathdynamic_clip_attention.py
190 lines (140 loc) · 7.7 KB
/
dynamic_clip_attention.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
from keras.layers import Dense, Input, RepeatVector, Lambda, Permute, Multiply, Concatenate
from keras.layers.convolutional import Conv1D
from keras.models import Model
from keras.optimizers import RMSprop, Adam
from keras import backend as K
from keras.layers.wrappers import TimeDistributed
class _Attention(object):
def __init__(self, ques_length, answer_length, nr_hidden, dropout=0.0, L2=0.0, activation='relu'):
self.ques_length = ques_length
self.answer_length = answer_length
def __call__(self, sent1, sent2, reverse = False):
def _outer(AB):
att_ji = K.batch_dot(AB[1], K.permute_dimensions(AB[0], (0, 2, 1)))
return K.permute_dimensions(att_ji,(0, 2, 1))
if reverse:
return Lambda(_outer,
output_shape=(self.answer_length, self.ques_length))([sent2, sent1])
else:
return Lambda(_outer,
output_shape=(self.ques_length, self.answer_length))([sent1, sent2])
class _SoftAlignment(object):
def __init__(self, nr_hidden):
# self.max_length = max_length
self.nr_hidden = nr_hidden
def __call__(self, sentence, attention, ques_len, max_length, transpose=False):
def _normalize_attention(attmat):
att = attmat[0]
mat = attmat[1]
ques_len = attmat[2]
if transpose:
att = K.permute_dimensions(att,(0, 2, 1))
# 3d softmax
e = K.exp(att - K.max(att, axis=-1, keepdims=True))
g = e * ques_len
s = K.sum(g, axis=-1, keepdims=True) + K.epsilon()
sm_att = g / s
return K.batch_dot(sm_att, mat)
return Lambda(_normalize_attention,
output_shape=(max_length, self.nr_hidden))([attention, sentence, ques_len]) # Shape: (i, n)
def DynamicClipAttention(model_param, elmo_embedding):
hidden_dim = model_param["hidden_dim"]
question = Input(shape=(1, ), dtype="string", name='question_base_inner')
question_len = Input(shape=(model_param["enc_timesteps"],), dtype='float32', name='question_len')
answer_len = Input(shape=(model_param["dec_timesteps"],), dtype='float32', name='answer_len')
answer = Input(shape=(1, ), dtype="string", name='answer_good_base_inner')
embedding_layer_ques = Lambda(elmo_embedding)
embedding_layer_ans = Lambda(elmo_embedding)
question_emb = embedding_layer_ques(question)
answer_emb = embedding_layer_ans(answer)
ques_filter_repeat_len = RepeatVector(model_param["dec_timesteps"])(question_len)
ans_filter_repeat_len = RepeatVector(model_param["enc_timesteps"])(answer_len)
ans_repeat_len = RepeatVector(model_param["hidden_dim"])(answer_len)
ans_repear_vec = Permute((2,1))(ans_repeat_len)
ques_repeat_len = RepeatVector(model_param["hidden_dim"])(question_len)
ques_repear_vec = Permute((2,1))(ques_repeat_len)
SigmoidDense = Dense(hidden_dim,activation="sigmoid")
TanhDense = Dense(hidden_dim,activation="tanh")
QueTimeSigmoidDense = TimeDistributed(SigmoidDense,name="que_time_s")
QueTimeTanhDense = TimeDistributed(TanhDense,name="que_time_t")
AnsTimeSigmoidDense = TimeDistributed(SigmoidDense,name="ans_time_s")
AnsTimeTanhDense = TimeDistributed(TanhDense,name="ans_time_t")
question_sig = QueTimeSigmoidDense(question_emb)
question_tanh = QueTimeTanhDense(question_emb)
question_proj = Multiply()([question_sig,question_tanh])
answer_sig = AnsTimeSigmoidDense(answer_emb)
answer_tanh = AnsTimeTanhDense(answer_emb)
answer_proj = Multiply()([answer_sig,answer_tanh])
Attend = _Attention(model_param["enc_timesteps"], model_param["dec_timesteps"] , hidden_dim, dropout=0.2)
Align = _SoftAlignment( hidden_dim)
ques_atten_metrics = Attend(question_proj,answer_proj)
ans_atten_metrics = Attend(question_proj,answer_proj,reverse = True)
answer_align = Align(question_proj,ques_atten_metrics,ques_filter_repeat_len,model_param["dec_timesteps"], transpose=True)
question_align = Align(answer_proj,ans_atten_metrics,ans_filter_repeat_len,model_param["enc_timesteps"],transpose=True)
ans_temp_sim_output = Multiply()([answer_proj,answer_align])
ques_temp_sim_output = Multiply()([question_proj,question_align])
ans_sim_output = Multiply()([ans_temp_sim_output,ans_repear_vec])
ques_sim_output = Multiply()([ques_temp_sim_output,ques_repear_vec])
cnns = [Conv1D(kernel_size=filter_length,
filters=hidden_dim,
activation='relu',
padding='same') for filter_length in [1,2,3,4,5]]
cnn_feature = Concatenate()([cnn(ans_sim_output) for cnn in cnns])
maxpool = Lambda(lambda x: K.max(x, axis=1, keepdims=False), output_shape=lambda x: (x[0], x[2]))
cnn_pool = maxpool(cnn_feature)
OutputDense = Dense(hidden_dim,activation="relu")
feature = OutputDense(cnn_pool)
cnns1 = [Conv1D(kernel_size=filter_length,
filters=hidden_dim,
activation='relu',
padding='same') for filter_length in [1,2,3,4,5]]
cnn1_feature = Concatenate()([cnn(ques_sim_output) for cnn in cnns1])
cnn1_pool = maxpool(cnn1_feature)
OutputDense1 = Dense(hidden_dim,activation="relu")
feature1 = OutputDense1(cnn1_pool)
feature_total = Concatenate()([feature,feature1],)
FinalDense = Dense(hidden_dim, activation="relu")
feature_all = FinalDense(feature_total)
ScoreDense = Dense(1)#, activation="relu")
score = ScoreDense(feature_all)
basic_model = Model(inputs=[question,answer,question_len,answer_len],outputs=[score])
questions = Input(
shape=(1,), dtype='string', name='question_base')
question_lens = Input(shape=(model_param["enc_timesteps"],), dtype='float32', name='question_len')
good_answer = Input(
shape=(1,), dtype='string', name='answer_base')
answers = Input(
shape=(model_param["random_size"],), dtype='string', name='answer_bad_base')
answers_length = Input(shape=(model_param["random_size"],model_param["dec_timesteps"],), dtype='float32', name='answers_length')
good_answer_length = Input(shape=(model_param["dec_timesteps"],),dtype='float32', name='good_answer_len')
good_similarity = basic_model([questions, good_answer, question_lens,good_answer_length])
sim_list = []
for i in range(model_param["random_size"]):
convert_layer = Lambda(lambda x:x[:,i],output_shape=(model_param["dec_timesteps"],))
temp_tensor = convert_layer(answers)
temp_length = convert_layer(answers_length)
temp_sim = basic_model([questions,temp_tensor,question_lens,temp_length])
sim_list.append(temp_sim)
total_sim = Concatenate()(sim_list)
total_prob = Lambda(lambda x: K.log(K.softmax(x + K.epsilon())),
output_shape = (model_param["random_size"], ))(total_sim)
prediction_model = Model(
inputs=[questions, good_answer,question_lens,good_answer_length],
outputs=good_similarity, name='prediction_model')
prediction_model.compile(
loss="binary_crossentropy",
optimizer = RMSprop(lr=model_param["lr"],
clipnorm=1.0,
clipvalue=0.5))
training_model = Model(
inputs=[questions, answers,question_lens,answers_length], outputs=total_prob, name='training_model')
training_model.compile(
loss=lambda y_true,y_pred: K.mean(y_true*(K.log(K.clip(y_true,0.00001,1)) - y_pred )) ,
optimizer=Adam(lr=model_param["lr"],
beta_1=0.9,
beta_2=0.999,
clipvalue=0.5,
clipnorm=1.0,
epsilon=1e-04
))
return training_model, prediction_model