-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathmodels.py
194 lines (167 loc) · 9.35 KB
/
models.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
191
192
193
194
# from keras.layers import Dense, SimpleRNN, LSTM, Input, Flatten, Bidirectional, GRU, \
# TimeDistributed, Embedding, Conv1D, ConvLSTM2D, MaxPooling1D, Average, Concatenate, merge, average
# from keras.layers.merge import multiply, concatenate
# from keras.layers.core import *
# from keras.optimizers import Adam
# from keras.engine import Model
# from keras import backend as K
import tensorflow.keras.backend as K
from tensorflow.python.keras.engine.training import Model
from tensorflow.keras.layers import SimpleRNN, LSTM, Input, Bidirectional, TimeDistributed, Conv1D, MaxPooling1D, \
Concatenate, Permute, Reshape, Dense, Lambda, RepeatVector, multiply, Flatten, Dropout
def attention_3d_block(inputs, seq_length, single_attention_vector):
# inputs.shape = (batch_size, seq_length, input_dim)
input_dim = int(K.int_shape(inputs)[2])
a = Permute((2, 1))(inputs)
# Reshape has no purpose except making the code more explicit and clear:
a = Reshape((input_dim, seq_length))(a)
a = Dense(seq_length, activation='softmax')(a)
if single_attention_vector:
a = Lambda(lambda x: K.mean(x, axis=1), name='dim_reduction')(a)
a = RepeatVector(input_dim)(a)
# when you apply a Dense layer, it applies to the last dimension of your tensor.
# Permute is used to apply a Dense layer along the time axis (by default it's axis=1 in Keras)
a_probs = Permute((2, 1), name='attention_vec')(a)
output_attention_mul = multiply([inputs, a_probs], name='attention_mul')
return output_attention_mul
def context_model(seq_len, max_seq_length, emb_dim, classes, nodes=128, dropout=0.2,
single_attention_vector=False):
# Encode each time step # dummyModel(3, 20, 1024, 42)
in_sentence = Input(shape=(max_seq_length, emb_dim,)) # , dtype='int64')
# embedded_sentence = Embedding(len(word_index) + 1, emb_dim, trainable=True)(in_sentence)
lstm_sentence = LSTM(nodes)(in_sentence)
encoded_model = Model(in_sentence, lstm_sentence)
encoded_model.summary()
# Model contextual time steps
sequence_input = Input(shape=(seq_len, max_seq_length, emb_dim))
seq_encoded = TimeDistributed(encoded_model)(sequence_input)
seq_encoded = Dropout(dropout)(seq_encoded)
# Encode entire sentence
seq_encoded = LSTM(nodes, return_sequences=True)(seq_encoded)
# Apply attention layer
attention_mul = attention_3d_block(seq_encoded, seq_len, single_attention_vector)
attention_mul = Flatten()(attention_mul)
# Prediction
prediction = Dense(classes, activation='softmax')(attention_mul)
model = Model(sequence_input, prediction)
model.compile(loss='categorical_crossentropy', optimizer='rmsprop', metrics=['accuracy'])
print(model.summary())
return model
def context_model_att(seq_len, max_seq_length, emb_dim, classes, nodes=128, dropout=0.2,
single_attention_vector=False, train_with_mean=False):
# Encode each time step # dummyModel(3, 20, 1024, 42)
in_sentence = Input(shape=(max_seq_length, emb_dim,)) # , dtype='int64')
# embedded_sentence = Embedding(len(word_index) + 1, emb_dim, trainable=True)(in_sentence)
lstm_sentence = LSTM(nodes, return_sequences=True)(in_sentence)
# Apply attention layer
attention_mul = attention_3d_block(lstm_sentence, max_seq_length, single_attention_vector)
# encoded_sentence = Flatten()(attention_mul)
if train_with_mean:
mean_vectors_norms = Lambda(lambda x: K.mean(x, axis=1), name='dim_reduction1')(attention_mul)
mean_vectors = Lambda(lambda x: K.mean(x, axis=1), name='dim_reduction2')(in_sentence)
concatenate_sentence = Concatenate()([mean_vectors_norms, mean_vectors])
encoded_model = Model(in_sentence, concatenate_sentence)
else:
encoded_sentence = Flatten()(attention_mul)
encoded_model = Model(in_sentence, encoded_sentence)
encoded_model.summary()
# Model contextual time steps
sequence_input = Input(shape=(seq_len, max_seq_length, emb_dim))
seq_encoded = TimeDistributed(encoded_model)(sequence_input)
seq_encoded = Dropout(dropout)(seq_encoded)
# Encode entire sentence
seq_encoded = LSTM(nodes, return_sequences=True)(seq_encoded)
# Apply attention layer
attention_mul = attention_3d_block(seq_encoded, seq_len, single_attention_vector)
attention_mul = Flatten(name='flatten_attention')(attention_mul)
# Prediction
prediction = Dense(classes, activation='softmax')(attention_mul)
model = Model(sequence_input, prediction)
model.compile(loss='categorical_crossentropy', optimizer='rmsprop', metrics=['accuracy'])
print(model.summary())
return model
# context_model(3, 20, 1024, 42)
def model_attention_applied_after_lstm(seq_length, emb_dim, num_classes, single_attention_vector=False):
inputs = Input(shape=(seq_length, emb_dim,))
lstm_units = 64
# lstm_out = (SimpleRNN(lstm_units, return_sequences=True))(inputs)
lstm_out = LSTM(lstm_units, return_sequences=True)(inputs)
attention_mul = attention_3d_block(lstm_out, seq_length, single_attention_vector)
attention_mul = Flatten()(attention_mul)
# inter_rep = Dense(100)(attention_mul)
output = Dense(num_classes, activation='softmax')(attention_mul)
model = Model(inputs=[inputs], outputs=output)
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
print(model.summary())
return model
# model_attention_applied_after_lstm(3, 1024, 42, True)
def model_attention_applied_after_bilstm(seq_length, emb_dim, num_classes, single_attention_vector=False):
inputs = Input(shape=(seq_length, emb_dim,))
lstm_units = 64
# lstm_out = (SimpleRNN(lstm_units, return_sequences=True))(inputs)
lstm_out = Bidirectional(LSTM(lstm_units, return_sequences=True))(inputs)
attention_mul = attention_3d_block(lstm_out, seq_length, single_attention_vector)
attention_mul = Flatten(name='flatten_attention')(attention_mul)
# inter_rep = Dense(100)(attention_mul)
output = Dense(num_classes, activation='softmax')(attention_mul)
model = Model(inputs=[inputs], outputs=output)
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
print(model.summary())
return model
def model_attention_applied_after_bisrnn(seq_length, emb_dim, num_classes, single_attention_vector=False):
inputs = Input(shape=(seq_length, emb_dim,))
lstm_units = 64
# lstm_out = (SimpleRNN(lstm_units, return_sequences=True))(inputs)
lstm_out = Bidirectional(SimpleRNN(lstm_units, return_sequences=True))(inputs)
attention_mul = attention_3d_block(lstm_out, seq_length, single_attention_vector)
attention_mul = Flatten()(attention_mul)
# inter_rep = Dense(100)(attention_mul)
output = Dense(num_classes, activation='softmax')(attention_mul)
model = Model(inputs=[inputs], outputs=output)
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
print(model.summary())
return model
def model_for_utterance_level(seq_length, emb_dim, num_classes, single_attention_vector=False):
inputs = Input(shape=(seq_length, emb_dim,))
lstm_units = 64
# lstm_out = (SimpleRNN(lstm_units, return_sequences=True))(inputs)
lstm_out = Bidirectional(SimpleRNN(lstm_units, return_sequences=True))(inputs)
attention_mul = attention_3d_block(lstm_out, seq_length, single_attention_vector)
attention_mul = Flatten()(attention_mul)
# inter_rep = Dense(100)(attention_mul)
output = Dense(num_classes, activation='softmax')(attention_mul)
model = Model(inputs=[inputs], outputs=output)
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
print(model.summary())
return model
def non_context_model_for_utterance_level(emb_dim, num_classes):
inputs = Input(shape=(emb_dim,))
units = 128
reshape_features = Reshape((32, 32))(inputs)
hidden_out = Conv1D(units, kernel_size=3, activation='relu')(reshape_features)
hidden_out = Conv1D(units, kernel_size=3, activation='relu')(hidden_out)
hidden_out_pooling = MaxPooling1D(pool_size=2)(hidden_out)
inter_rep = Flatten()(hidden_out_pooling)
inter_rep = Dense(100)(inter_rep)
output = Dense(num_classes, activation='softmax')(inter_rep)
model = Model(inputs=[inputs], outputs=output)
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
print(model.summary())
return model
def context_model_att_with_pt_encoder(seq_len, max_seq_length, emb_dim, classes, encoded_model,
nodes=128, dropout=0.2, single_attention_vector=False):
# Model contextual time steps
sequence_input = Input(shape=(seq_len, max_seq_length, emb_dim))
seq_encoded = TimeDistributed(encoded_model)(sequence_input)
seq_encoded = Dropout(dropout)(seq_encoded)
# Encode entire sentence
seq_encoded = LSTM(nodes, return_sequences=True)(seq_encoded)
# Apply attention layer
attention_mul = attention_3d_block(seq_encoded, seq_len, single_attention_vector)
attention_mul = Flatten(name='flatten_attention')(attention_mul)
# Prediction
prediction = Dense(classes, activation='softmax')(attention_mul)
model = Model(sequence_input, prediction)
model.compile(loss='categorical_crossentropy', optimizer='rmsprop', metrics=['accuracy'])
print(model.summary())
return model