-
Notifications
You must be signed in to change notification settings - Fork 2
/
decoder.py
127 lines (100 loc) · 5.82 KB
/
decoder.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
import numpy as np
import tensorflow as tf
import highway_maxout as hmn
'''
doc_length is shape of [Batch]
'''
def decoderBatch(U, lstm_dec, dropout_rate, batch_size, doc_length, FLAGS):
max_sequence_length = FLAGS.max_sequence_length
max_decoder_iterations = FLAGS.max_decoder_iterations
lstm_size = FLAGS.lstm_size
# returns (batch, D)
def _maskToPreventIndexesOutOfLengthOfDoc(max_sequence_length, doc_length):
def createVector(max_sequence_length, l):
v = tf.ones(shape=[1,max_sequence_length - l], dtype=tf.float32) * (10 ** -30)
return tf.pad(v, [[0, 0], [l, 0]])
mask = tf.map_fn(lambda l: createVector(max_sequence_length, l), doc_length, dtype=tf.float32)
return tf.squeeze(mask)
mask = _maskToPreventIndexesOutOfLengthOfDoc(max_sequence_length, doc_length)
with tf.name_scope('DYNAMIC_POINTING_DECODER'):
lstm_dec_state = lstm_dec.zero_state(batch_size, tf.float32)
start_pos = tf.zeros(shape=[batch_size], dtype=tf.int32)
end_pos = tf.zeros(shape=[batch_size], dtype=tf.int32)
sum_start_scores = tf.zeros([batch_size, max_sequence_length, 1])
sum_end_scores = tf.zeros([batch_size, max_sequence_length, 1])
for i_ in range(max_decoder_iterations):
scores_start, scores_end, start_pos, end_pos, lstm_dec_state = decoderIteration(U,
lstm_dec_state,
start_pos,
end_pos,
lstm_dec,
dropout_rate,
max_sequence_length,
lstm_size,
FLAGS,
batch_size,
mask,
i_)
#sum_start_scores = tf.add(sum_start_scores, scores_start)
#sum_end_scores = tf.add(sum_start_scores, scores_end)
sum_start_scores = tf.concat([sum_start_scores, tf.expand_dims(scores_start, -1)], axis=2)
sum_end_scores = tf.concat([sum_end_scores, tf.expand_dims(scores_end, -1)], axis=2)
# returns (B, D, number of iterations)
return sum_start_scores[0:, 0:, 1: ], sum_end_scores[0:, 0:, 1: ]
'''
U (batch, D, 2L)
scores_start [batch, D]
scores_end [batch, D]
start_pos [batch]
end_pos [batch]
new_lstm_state [batch, L, 1]
# returns batched tuple (scores_start, scores_end, start_pos, end_pos, new_lstm_state)
'''
def decoderIteration(U, lstm_state, start_pos, end_pos, lstm_dec, dropout_rate, max_sequence_length, lstm_size, FLAGS, batch_size, mask, iter_number):
# returns (batch, 2L)
def _getPos(U, start_positions, rows_size, vec_size):
def createMask(pos, size, vector):
return tf.pad(vector, [[pos, size - pos - 1], [0, 0]])
ones = tf.ones([1, vec_size], tf.float32)
# (batch, D, 2L)
mask_matrix = tf.map_fn(lambda pos: createMask(pos, rows_size, ones), start_positions, dtype=tf.float32)
positions = tf.multiply(U, mask_matrix)
# (batch, 2L)
res = tf.reduce_sum(positions, 1)
res.set_shape([None, vec_size])
return res;
with tf.name_scope('Decoder_Iteration'):
with tf.name_scope('Next_Start'):
scores_start = hmn.HMN_Batch(U,
lstm_state.h,
_getPos(U, start_pos, max_sequence_length, lstm_size * 2),
_getPos(U, end_pos, max_sequence_length, lstm_size * 2),
batch_size,
'start',
FLAGS,
dropout_rate,
iter_number)
scores_start = tf.add(scores_start, mask)
new_start_pos = tf.to_int32(tf.argmax(scores_start, 1))
with tf.name_scope('Next_End'):
scores_end = hmn.HMN_Batch(U,
lstm_state.h,
_getPos(U, new_start_pos, max_sequence_length, lstm_size * 2),
_getPos(U, end_pos, max_sequence_length, lstm_size * 2),
batch_size,
'end',
FLAGS,
dropout_rate,
iter_number)
scores_end = tf.add(scores_end, mask)
new_end_pos = tf.to_int32(tf.argmax(scores_end, 1))
with tf.name_scope('LSTM_State_Update'):
lstm_input = tf.concat(
[
_getPos(U, new_start_pos, max_sequence_length, lstm_size * 2),
_getPos(U, new_end_pos, max_sequence_length, lstm_size * 2)
],
axis = 1
)
output, new_lstm_state = lstm_dec(lstm_input, lstm_state)
return scores_start, scores_end, new_start_pos , new_end_pos, new_lstm_state