ScientificDiscourseTagging/attention.py at master · jacklxc/ScientificDiscourseTagging · GitHub

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
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
import tensorflow as tf
import numpy as np
import keras.backend as K
from keras import activations, initializers, regularizers
from keras.engine.topology import Layer
from keras.layers import SimpleRNN
class TensorAttention(Layer):
    '''
    Attention layer that operates on tensors
    '''
    input_ndim = 4
    def __init__(self, att_input_shape, context='word', init='glorot_uniform', activation='tanh', weights=None, hard_k=0, proj_dim = None, rec_hid_dim = None, return_attention=False, **kwargs):
        self.supports_masking = True
        self.init = initializers.get(init)
        self.activation = activations.get(activation)
        self.context = context
        self.td1, self.td2, self.wd = att_input_shape # (c,w,d)
        self.return_attention = return_attention
        if proj_dim is not None:
            self.proj_dim = proj_dim
        else:
            self.proj_dim = int(self.wd/2) # p
        if rec_hid_dim is not None:
            self.rec_hid_dim = rec_hid_dim
        else:
            self.rec_hid_dim = int(self.proj_dim/2)
        self.initial_weights = weights
        self.hard = True if hard_k>0 else False
        self.k = hard_k
        super(TensorAttention, self).__init__(**kwargs)

    def build(self,input_shape):
        self.att_proj = self.add_weight(name='att_proj',shape=(self.wd, self.proj_dim),
                                        initializer=self.init, trainable=True) # P, (d,p)
        if self.context == 'word':
            self.att_scorer = self.add_weight(name='att_scorer',shape=(self.proj_dim,),initializer=self.init, trainable=True)
        elif self.context == 'clause':
            self.att_scorer = self.add_weight(name='att_scorer',shape=(self.rec_hid_dim,),initializer=self.init, trainable=True)
            self.recurrent_weight = self.add_weight(name='recurrent_weight',
                                                    shape=(self.rec_hid_dim,self.rec_hid_dim),
                                                    initializer=self.init, trainable=True)
            self.encoder_weight = self.add_weight(name='encoder_weight',
                                                    shape=(self.proj_dim,self.rec_hid_dim),
                                                    initializer=self.init, trainable=True)
        elif self.context == 'bidirectional_clause':
            self.att_scorer = self.add_weight(name='att_scorer',shape=(self.rec_hid_dim,),initializer=self.init, trainable=True)
            self.recurrent_weight_forward = self.add_weight(name='recurrent_weight_forward',
                                                    shape=(self.rec_hid_dim,self.rec_hid_dim),
                                                    initializer=self.init, trainable=True)
            self.encoder_weight_forward = self.add_weight(name='encoder_weight_forward',
                                                    shape=(self.proj_dim,self.rec_hid_dim),
                                                    initializer=self.init, trainable=True)
            self.recurrent_weight_backward = self.add_weight(name='recurrent_weight_backward',
                                                    shape=(self.rec_hid_dim,self.rec_hid_dim),
                                                    initializer=self.init, trainable=True)
            self.encoder_weight_backward = self.add_weight(name='encoder_weight_backward',
                                                    shape=(self.proj_dim,self.rec_hid_dim),
                                                    initializer=self.init, trainable=True)

        elif self.context == 'LSTM_clause':
            self.att_scorer = self.add_weight(name='att_scorer',shape=(self.rec_hid_dim,),initializer=self.init, trainable=True)
            self.kernel = self.add_weight(shape=(self.proj_dim, self.rec_hid_dim * 4),
                                name='kernel',initializer=self.init,trainable=True)
            self.recurrent_kernel = self.add_weight(shape=(self.rec_hid_dim, self.rec_hid_dim * 4),
                                name='recurrent_kernel',
                                initializer=self.init,trainable=True)
            self.bias = self.add_weight(shape=(self.rec_hid_dim * 4,),name='bias',
                                        initializer=self.init,trainable=True)
            self.kernel_i = self.kernel[:, :self.rec_hid_dim]
            self.kernel_f = self.kernel[:, self.rec_hid_dim: self.rec_hid_dim * 2]
            self.kernel_c = self.kernel[:, self.rec_hid_dim * 2: self.rec_hid_dim * 3]
            self.kernel_o = self.kernel[:, self.rec_hid_dim * 3:]

            self.recurrent_kernel_i = self.recurrent_kernel[:, :self.rec_hid_dim]
            self.recurrent_kernel_f = (
                self.recurrent_kernel[:, self.rec_hid_dim: self.rec_hid_dim * 2])
            self.recurrent_kernel_c = (
                self.recurrent_kernel[:, self.rec_hid_dim * 2: self.rec_hid_dim * 3])
            self.recurrent_kernel_o = self.recurrent_kernel[:, self.rec_hid_dim * 3:]

            self.bias_i = self.bias[:self.rec_hid_dim]
            self.bias_f = self.bias[self.rec_hid_dim: self.rec_hid_dim * 2]
            self.bias_c = self.bias[self.rec_hid_dim * 2: self.rec_hid_dim * 3]
            self.bias_o = self.bias[self.rec_hid_dim * 3:]

        elif self.context == 'biLSTM_clause':
            self.att_scorer = self.add_weight(name='att_scorer',shape=(self.rec_hid_dim*2,),initializer=self.init, trainable=True)
            self.kernel_forward = self.add_weight(shape=(self.proj_dim, self.rec_hid_dim * 4),
                                name='kernel_forward',initializer=self.init,trainable=True)
            self.recurrent_kernel_forward = self.add_weight(shape=(self.rec_hid_dim, self.rec_hid_dim * 4),
                                name='recurrent_kernel_forward',
                                initializer=self.init,trainable=True)
            self.bias_forward = self.add_weight(shape=(self.rec_hid_dim * 4,),name='bias_forward',
                                        initializer=self.init,trainable=True)
            self.kernel_i_forward = self.kernel_forward[:, :self.rec_hid_dim]
            self.kernel_f_forward = self.kernel_forward[:, self.rec_hid_dim: self.rec_hid_dim * 2]
            self.kernel_c_forward = self.kernel_forward[:, self.rec_hid_dim * 2: self.rec_hid_dim * 3]
            self.kernel_o_forward = self.kernel_forward[:, self.rec_hid_dim * 3:]

            self.recurrent_kernel_i_forward = self.recurrent_kernel_forward[:, :self.rec_hid_dim]
            self.recurrent_kernel_f_forward = (
                self.recurrent_kernel_forward[:, self.rec_hid_dim: self.rec_hid_dim * 2])
            self.recurrent_kernel_c_forward = (
                self.recurrent_kernel_forward[:, self.rec_hid_dim * 2: self.rec_hid_dim * 3])
            self.recurrent_kernel_o_forward = self.recurrent_kernel_forward[:, self.rec_hid_dim * 3:]

            self.bias_i_forward = self.bias_forward[:self.rec_hid_dim]
            self.bias_f_forward = self.bias_forward[self.rec_hid_dim: self.rec_hid_dim * 2]
            self.bias_c_forward = self.bias_forward[self.rec_hid_dim * 2: self.rec_hid_dim * 3]
            self.bias_o_forward = self.bias_forward[self.rec_hid_dim * 3:]

            self.kernel_backward = self.add_weight(shape=(self.proj_dim, self.rec_hid_dim * 4),
                                name='kernel_backward',initializer=self.init,trainable=True)
            self.recurrent_kernel_backward = self.add_weight(shape=(self.rec_hid_dim, self.rec_hid_dim * 4),
                                name='recurrent_kernel_backward',
                                initializer=self.init,trainable=True)
            self.bias_backward = self.add_weight(shape=(self.rec_hid_dim * 4,),name='bias_backward',
                                        initializer=self.init,trainable=True)
            self.kernel_i_backward = self.kernel_backward[:, :self.rec_hid_dim]
            self.kernel_f_backward = self.kernel_backward[:, self.rec_hid_dim: self.rec_hid_dim * 2]
            self.kernel_c_backward = self.kernel_backward[:, self.rec_hid_dim * 2: self.rec_hid_dim * 3]
            self.kernel_o_backward = self.kernel_backward[:, self.rec_hid_dim * 3:]

            self.recurrent_kernel_i_backward = self.recurrent_kernel_backward[:, :self.rec_hid_dim]
            self.recurrent_kernel_f_backward = (
                self.recurrent_kernel_backward[:, self.rec_hid_dim: self.rec_hid_dim * 2])
            self.recurrent_kernel_c_backward = (
                self.recurrent_kernel_backward[:, self.rec_hid_dim * 2: self.rec_hid_dim * 3])
            self.recurrent_kernel_o_backward = self.recurrent_kernel_backward[:, self.rec_hid_dim * 3:]

            self.bias_i_backward = self.bias_backward[:self.rec_hid_dim]
            self.bias_f_backward = self.bias_backward[self.rec_hid_dim: self.rec_hid_dim * 2]
            self.bias_c_backward = self.bias_backward[self.rec_hid_dim * 2: self.rec_hid_dim * 3]
            self.bias_o_backward = self.bias_backward[self.rec_hid_dim * 3:]

        elif self.context == 'para':
            self.att_scorer = self.add_weight(name='att_scorer',shape=(self.td1, self.td2, self.proj_dim),
                                              initializer=self.init, trainable=True) # (c,w,p)
        if self.initial_weights is not None:
            self.set_weights(self.initial_weights)
            del self.initial_weights
        super(TensorAttention, self).build(input_shape)

    def compute_output_shape(self, input_shape):
        if self.return_attention:
            return [(input_shape[0], input_shape[1], input_shape[3]),(input_shape[0], input_shape[1], input_shape[2])]
        else:
            return (input_shape[0], input_shape[1], input_shape[3])

    def compute_mask(self, input, input_mask=None):
        if input_mask is not None:
            if self.return_attention:
                return [input_mask[:,:,-1], input_mask]
            else:
                return input_mask[:,:,-1]
        else:
            return None

    def call(self, X, mask=None):
        # input: D (sample,c,w,d)
        proj_input = self.activation(tf.tensordot(X, self.att_proj, axes=[[3],[0]])) # tanh(dot(D,P))=Dl,（sample,c,w,p）
        if self.context == 'word':
            raw_att_scores = tf.tensordot(proj_input, self.att_scorer, axes=[[3],[0]]) # (sample,c,w)
        elif self.context == 'clause':
            def step(X, states):
                new_state = activations.tanh(tf.tensordot(X,self.encoder_weight, axes=[[2],[0]]) \
                    + tf.tensordot(states[0],self.recurrent_weight, axes=[[2],[0]]))
                return new_state, [new_state]
            # Make all-zero initial state.
            # Directly obtaining the first input dimension is not allowed, so this is the work-aronud.
            initial_state = tf.tensordot(K.max(proj_input*0,axis=2),K.zeros((self.proj_dim, self.rec_hid_dim)), axes = [[2],[0]])
            proj_input_permute = K.permute_dimensions(proj_input,(0,2,1,3))
            _,all_rnn_out,_ = K.rnn(step,proj_input_permute,[initial_state])
            raw_att_scores = tf.tensordot(K.permute_dimensions(all_rnn_out,(0,2,1,3)),
                                                self.att_scorer, axes=[[3],[0]])

        elif self.context == 'bidirectional_clause':
            def step_forward(X, states):
                new_state = activations.tanh(tf.tensordot(X,self.encoder_weight_forward, axes=[[2],[0]]) \
                    + tf.tensordot(states[0],self.recurrent_weight_forward, axes=[[2],[0]]))
                return new_state, [new_state]
            def step_backward(X, states):
                new_state = activations.tanh(tf.tensordot(X,self.encoder_weight_backward, axes=[[2],[0]]) \
                    + tf.tensordot(states[0],self.recurrent_weight_backward, axes=[[2],[0]]))
                return new_state, [new_state]
            # Make all-zero initial state.
            # Directly obtaining the first input dimension is not allowed, so this is the work-aronud.
            initial_state = tf.tensordot(K.max(proj_input*0,axis=2),K.zeros((self.proj_dim, self.rec_hid_dim)), axes = [[2],[0]])
            proj_input_permute = K.permute_dimensions(proj_input,(0,2,1,3))
            proj_input_permute_backward = K.reverse(proj_input_permute, 1)
            _,all_rnn_out_forward,_ = K.rnn(step_forward,proj_input_permute,[initial_state])
            _,all_rnn_out_backward,_ = K.rnn(step_backward,proj_input_permute,[initial_state])
            all_rnn_out = all_rnn_out_forward+all_rnn_out_backward
            raw_att_scores = tf.tensordot(K.permute_dimensions(all_rnn_out,(0,2,1,3)),
                                                self.att_scorer, axes=[[3],[0]])

        elif self.context == 'LSTM_clause':
            def step(inputs, states):
                h_tm1 = states[0]  # previous memory state
                c_tm1 = states[1]  # previous carry state

                x_i = tf.tensordot(inputs, self.kernel_i,axes=[[2],[0]])
                x_f = tf.tensordot(inputs, self.kernel_f,axes=[[2],[0]])
                x_c = tf.tensordot(inputs, self.kernel_c,axes=[[2],[0]])
                x_o = tf.tensordot(inputs, self.kernel_o,axes=[[2],[0]])
                x_i = K.bias_add(x_i, self.bias_i)
                x_f = K.bias_add(x_f, self.bias_f)
                x_c = K.bias_add(x_c, self.bias_c)
                x_o = K.bias_add(x_o, self.bias_o)
                i = activations.hard_sigmoid(x_i + tf.tensordot(h_tm1,
                                                          self.recurrent_kernel_i,axes=[[2],[0]]))
                f = activations.hard_sigmoid(x_f + tf.tensordot(h_tm1,
                                                          self.recurrent_kernel_f,axes=[[2],[0]]))
                c = f * c_tm1 + i * activations.tanh(x_c + tf.tensordot(h_tm1,
                                                                self.recurrent_kernel_c,axes=[[2],[0]]))
                o = activations.hard_sigmoid(x_o + tf.tensordot(h_tm1,
                                                          self.recurrent_kernel_o,axes=[[2],[0]]))
                h = o * activations.tanh(c)

                return h, [h, c]
            # Make all-zero initial state.
            # Directly obtaining the first input dimension is not allowed, so this is the work-aronud.
            initial_state = tf.tensordot(K.max(proj_input*0,axis=2),K.zeros((self.proj_dim, self.rec_hid_dim)), axes = [[2],[0]])
            proj_input_permute = K.permute_dimensions(proj_input,(0,2,1,3))
            _,all_rnn_out,_ = K.rnn(step,proj_input_permute,[initial_state,initial_state])
            raw_att_scores = tf.tensordot(K.permute_dimensions(all_rnn_out,(0,2,1,3)),
                                                self.att_scorer, axes=[[3],[0]])
        elif self.context == 'biLSTM_clause':
            def step_forward(inputs, states):
                h_tm1 = states[0]  # previous memory state
                c_tm1 = states[1]  # previous carry state

                x_i = tf.tensordot(inputs, self.kernel_i_forward,axes=[[2],[0]])
                x_f = tf.tensordot(inputs, self.kernel_f_forward,axes=[[2],[0]])
                x_c = tf.tensordot(inputs, self.kernel_c_forward,axes=[[2],[0]])
                x_o = tf.tensordot(inputs, self.kernel_o_forward,axes=[[2],[0]])
                x_i = K.bias_add(x_i, self.bias_i_forward)
                x_f = K.bias_add(x_f, self.bias_f_forward)
                x_c = K.bias_add(x_c, self.bias_c_forward)
                x_o = K.bias_add(x_o, self.bias_o_forward)
                i = activations.hard_sigmoid(x_i + tf.tensordot(h_tm1,
                                                          self.recurrent_kernel_i_forward,axes=[[2],[0]]))
                f = activations.hard_sigmoid(x_f + tf.tensordot(h_tm1,
                                                          self.recurrent_kernel_f_forward,axes=[[2],[0]]))
                c = f * c_tm1 + i * activations.tanh(x_c + tf.tensordot(h_tm1,
                                                                self.recurrent_kernel_c_forward,axes=[[2],[0]]))
                o = activations.hard_sigmoid(x_o + tf.tensordot(h_tm1,
                                                          self.recurrent_kernel_o_forward,axes=[[2],[0]]))
                h = o * activations.tanh(c)

                return h, [h, c]

            def step_backward(inputs, states):
                h_tm1 = states[0]  # previous memory state
                c_tm1 = states[1]  # previous carry state

                x_i = tf.tensordot(inputs, self.kernel_i_backward,axes=[[2],[0]])
                x_f = tf.tensordot(inputs, self.kernel_f_backward,axes=[[2],[0]])
                x_c = tf.tensordot(inputs, self.kernel_c_backward,axes=[[2],[0]])
                x_o = tf.tensordot(inputs, self.kernel_o_backward,axes=[[2],[0]])
                x_i = K.bias_add(x_i, self.bias_i_backward)
                x_f = K.bias_add(x_f, self.bias_f_backward)
                x_c = K.bias_add(x_c, self.bias_c_backward)
                x_o = K.bias_add(x_o, self.bias_o_backward)
                i = activations.hard_sigmoid(x_i + tf.tensordot(h_tm1,
                                                          self.recurrent_kernel_i_backward,axes=[[2],[0]]))
                f = activations.hard_sigmoid(x_f + tf.tensordot(h_tm1,
                                                          self.recurrent_kernel_f_backward,axes=[[2],[0]]))
                c = f * c_tm1 + i * activations.tanh(x_c + tf.tensordot(h_tm1,
                                                                self.recurrent_kernel_c_backward,axes=[[2],[0]]))
                o = activations.hard_sigmoid(x_o + tf.tensordot(h_tm1,
                                                          self.recurrent_kernel_o_backward,axes=[[2],[0]]))
                h = o * activations.tanh(c)

                return h, [h, c]

            # Make all-zero initial state.
            # Directly obtaining the first input dimension is not allowed, so this is the work-aronud.
            initial_state = tf.tensordot(K.max(proj_input*0,axis=2),K.zeros((self.proj_dim, self.rec_hid_dim)), axes = [[2],[0]])
            proj_input_permute = K.permute_dimensions(proj_input,(0,2,1,3))
            proj_input_permute_backward = K.reverse(proj_input_permute, 1)
            _,all_rnn_out_forward,_ = K.rnn(step_forward,proj_input_permute,[initial_state,initial_state])
            _,all_rnn_out_backward,_ = K.rnn(step_backward,proj_input_permute_backward,[initial_state,initial_state])
            all_rnn_out = K.concatenate([all_rnn_out_forward,all_rnn_out_backward],axis=-1)
            raw_att_scores = tf.tensordot(K.permute_dimensions(all_rnn_out,(0,2,1,3)),
                                                self.att_scorer, axes=[[3],[0]])


        elif self.context == 'para':
            raw_att_scores = K.sum(tf.tensordot(proj_input, self.att_scorer, axes=[[3],[2]]), axis = [1, 2]) # (sample,c,w)

        if self.hard: # Hard attention
            rep_att_score = K.repeat_elements(K.expand_dims(raw_att_scores),rep=self.wd,axis=-1)
            top = tf.nn.top_k(K.permute_dimensions(rep_att_score,(0,1,3,2)),k=self.k).indices
            permute_X = K.permute_dimensions(X,(0,1,3,2))
            reduced_X = K.permute_dimensions(tf.batch_gather(permute_X, top),(0,1,3,2))
            new_att_scores = K.softmax(tf.nn.top_k(raw_att_scores,k=self.k).values,axis=2)
            result = K.batch_dot(new_att_scores,reduced_X,axes=[2,2])
        else:
            att_scores = K.softmax(raw_att_scores, axis=2)
            result = K.batch_dot(att_scores,X,axes=[2,2]) # (sample,c,d)
        if self.return_attention:
            return [result, raw_att_scores]
        else:
            return result

    def get_config(self):
        return {'name': 'TensorAttention',
                'att_input_shape': (self.td1, self.td2, self.wd),
                'proj_dim': self.proj_dim,
                'rec_hid_dim': self.rec_hid_dim,
                'hard_k': self.k,
                'context': self.context,
                'trainable': True}