-
Notifications
You must be signed in to change notification settings - Fork 2
/
train.py
187 lines (152 loc) · 6.66 KB
/
train.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
import numpy as np
import tensorflow as tf
from itertools import islice
import dataset as ds
import time
import utils
def processLine(embeddings, str, max_doc_length, max_que_length):
start_pos, end_pos, doc, que = str.split(';')
start_pos = int(start_pos)
end_pos = int(end_pos)
document = doc.split(' ')
question = que.split(' ')
doc_v = embeddings.sentence2Vectors(document, max_doc_length)
que_v = embeddings.sentence2Vectors(question, max_que_length)
return start_pos, end_pos, document, question, doc_v, que_v
def processLineBatch(file, embeddings, batch_size, max_sequence_length, max_question_length,
question_ph, document_ph, dropout_rate_ph,
doc_len_ph, que_len_ph, start_true_ph, end_true_ph, batch_size_ph, learning_rate_ph, dropout_rate, lrate):
next_n_lines = list(islice(file, batch_size))
if not next_n_lines or len(next_n_lines) != batch_size: return None
q = []
d = []
s = []
e = []
dl = []
ql = []
batch_size_fact = 0;
for line_ in next_n_lines:
start_pos, end_pos, document, question, doc_v, que_v = processLine(
embeddings, line_, max_sequence_length, max_question_length
)
if len(document) > max_sequence_length or len(question) >= max_question_length or start_pos < 0:
print("Wrong example. Skip", document[0])
continue;
q.append(que_v)
d.append(doc_v)
s.append(start_pos)
e.append(end_pos)
dl.append(len(document))
ql.append(len(question))
batch_size_fact += 1
feed_dict = {
question_ph: q,
document_ph: d,
start_true_ph: s,
end_true_ph: e,
doc_len_ph: dl,
que_len_ph: ql,
dropout_rate_ph: dropout_rate,
batch_size_ph: batch_size_fact,
learning_rate_ph: lrate
}
if batch_size_fact != batch_size: print("Batch Size Fact", batch_size_fact)
return feed_dict
def accuracy(sess, params, accuracy, pr_start_idx, pr_end_idx):
try:
acc, starts, ends = sess.run(
(accuracy, pr_start_idx, pr_end_idx),
params
)
return acc, starts, ends
#writer.add_summary(stat, step)
#print('AVG accuracy', acc)
except:
print("Test Error", params)
def trainStep(sess, feed_dict, writer,
train_step, summary_op,
step, profiling = False):
run_options = None
run_metadata = None
if profiling:
run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
start_time = time.time()
try:
_, stat = sess.run(
(train_step, summary_op),
feed_dict = feed_dict,
options=run_options, run_metadata=run_metadata
)
writer.add_summary(stat, step)
#writer.add_summary(stat_train, step)
except:
print("Train Error", step)
def loss_and_accuracy(start_true, end_true, batch_size, sum_start_scores, sum_end_scores, max_sequence_length):
# loss and train step
onehot_labels_start = tf.one_hot(start_true, max_sequence_length)
onehot_labels_end = tf.one_hot(end_true, max_sequence_length)
with tf.name_scope('Loss'):
loss_start = tf.nn.softmax_cross_entropy_with_logits(
labels = onehot_labels_start,
logits = sum_start_scores)
loss_start = tf.reduce_mean(loss_start)
loss_end = tf.nn.softmax_cross_entropy_with_logits(
labels = onehot_labels_end,
logits = sum_end_scores)
loss_end = tf.reduce_mean(loss_end)
sum_loss = loss_start + loss_end
with tf.name_scope('Accuracy'):
with tf.name_scope('Prediction'):
pr_start_idx = tf.to_int32(tf.argmax(sum_start_scores, 1))
pr_end_idx = tf.to_int32(tf.argmax(sum_end_scores, 1))
with tf.name_scope('Accuracy'):
#accuracy_avg = tf.py_func(utils.f1_score_int_avg, [pr_start_idx, pr_end_idx, start_true, end_true], tf.float64)
accuracy = tf.py_func(utils.f1_score_int_list, [pr_start_idx, pr_end_idx, start_true, end_true], tf.float64)
return sum_loss, accuracy, pr_start_idx, pr_end_idx
# iter_start_scores, iter_end_scores is shape of (B, D, number of iterations)
def loss_and_accuracy_v2(start_true, end_true, batch_size, iter_start_scores, iter_end_scores, max_sequence_length, iter_num):
# loss and train step
onehot_labels_start = tf.one_hot(start_true, max_sequence_length)
onehot_labels_end = tf.one_hot(end_true, max_sequence_length)
def getIterNum(vec):
a = vec - tf.pad(vec[0:-1], [[0,1]])
b = tf.pad(vec[1:], [[0,1]])
c = tf.add(a,b)
d = tf.abs(tf.subtract(vec, c))
return tf.argmin(d)
# m is (D, number of iterations)
#def sliceByEffectiveIter(m):
# n = tf.cast(getIterNum(tf.argmax(m, 0)), tf.int32)
# return m[0: , 0:n+1]
# m is (D, number of iterations)
def leaveOnlyLastEffectiveIter(m):
n = tf.cast(getIterNum(tf.argmax(m, 0)), tf.int32)
return m[0: , n]
onehot_labels_start = tf.tile(tf.expand_dims(onehot_labels_start, -1), [1, 1, iter_num])
onehot_labels_end = tf.tile(tf.expand_dims(onehot_labels_end, -1), [1, 1, iter_num])
print("iter_end_scores", iter_end_scores)
print(onehot_labels_start)
print(onehot_labels_end)
start_scores = tf.map_fn(lambda m: leaveOnlyLastEffectiveIter(m), iter_start_scores, dtype=tf.float32)
end_scores = tf.map_fn(lambda m: leaveOnlyLastEffectiveIter(m), iter_end_scores, dtype=tf.float32)
with tf.name_scope('Loss'):
loss_start = tf.nn.softmax_cross_entropy_with_logits(
labels = onehot_labels_start,
logits = iter_start_scores,
dim = 1)
loss_start = tf.reduce_mean(loss_start)
loss_end = tf.nn.softmax_cross_entropy_with_logits(
labels = onehot_labels_end,
logits = iter_end_scores,
dim = 1)
loss_end = tf.reduce_mean(loss_end)
sum_loss = loss_start + loss_end
with tf.name_scope('Accuracy'):
with tf.name_scope('Prediction'):
pr_start_idx = tf.to_int32(tf.argmax(start_scores, 1))
pr_end_idx = tf.to_int32(tf.argmax(end_scores, 1))
with tf.name_scope('Accuracy'):
#accuracy_avg = tf.py_func(utils.f1_score_int_avg, [pr_start_idx, pr_end_idx, start_true, end_true], tf.float64)
accuracy = tf.py_func(utils.f1_score_int_list, [pr_start_idx, pr_end_idx, start_true, end_true], tf.float64)
return sum_loss, accuracy, pr_start_idx, pr_end_idx