-
Notifications
You must be signed in to change notification settings - Fork 40
/
data_utils.py
184 lines (168 loc) · 8.69 KB
/
data_utils.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
import numpy as np
import tensorflow as tf
from collections import defaultdict
from random import shuffle
import sys
import os
FLAGS = tf.app.flags.FLAGS
class Batcher(object):
def __init__(self, in_dir, batch_size, num_epochs=None):
self._batch_size = batch_size
self._epoch = 0
self._step = 0.
self._data = defaultdict(list)
self._starts = {}
self._ends = {}
self._bucket_probs = {}
self.sequence_batcher = SeqBatcher(in_dir, batch_size, 0, num_epochs=1)
def load_and_bucket_data(self, sess):
done = False
i = 0
while not done:
try:
batch = sess.run(self.sequence_batcher.next_batch_op)
self._data[batch[0].shape[1]].append(batch)
i += 1
except Exception as e:
done = True
# now flatten
for seq_len, batches in self._data.items():
self._data[seq_len] = [(label_batch[i], token_batch[i], shape_batch[i], char_batch[i], seq_len_batch[i], tok_len_batch[i])
for (label_batch, token_batch, shape_batch, char_batch, seq_len_batch, tok_len_batch) in batches
for i in range(label_batch.shape[0])]
self.reset_batch_pointer()
def next_batch(self):
if sum(self._bucket_probs.values()) == 0:
self.reset_batch_pointer()
# select bucket to create batch from
self._step += 1
bucket = self.select_bucket()
batch = self._data[bucket][self._starts[bucket]:self._ends[bucket]]
# update pointers
self._starts[bucket] = self._ends[bucket]
self._ends[bucket] = min(self._ends[bucket] + self._batch_size, len(self._data[bucket]))
self._bucket_probs[bucket] = max(0, self._ends[bucket] - self._starts[bucket])
_label_batch = np.array([b[0] for b in batch])
_token_batch = np.array([b[1] for b in batch])
_shape_batch = np.array([b[2] for b in batch])
_char_batch = np.array([b[3] for b in batch])
_seq_len_batch = np.array([b[4] for b in batch])
_tok_len_batch = np.array([b[5] for b in batch])
batch = (_label_batch, _token_batch, _shape_batch, _char_batch, _seq_len_batch, _tok_len_batch)
return batch
def reset_batch_pointer(self):
# shuffle each bucket
for bucket in self._data.values():
shuffle(bucket)
self._epoch += 1
self._step = 0.
# print('\nStarting epoch ' + str(self._epoch))
self._starts = {i: 0 for i in self._data.keys()}
self._ends = {i: min(self._batch_size, len(examples)) for i, examples in self._data.items()}
self._bucket_probs = {i: len(l) for (i, l) in self._data.items()}
def select_bucket(self):
buckets, weights = zip(*[(i, p) for i, p in self._bucket_probs.items() if p > 0])
total = float(sum(weights))
probs = [w / total for w in weights]
bucket = np.random.choice(buckets, p=probs)
return bucket
class SeqBatcher(object):
def __init__(self, in_pattern, batch_size, num_buckets=0, num_epochs=None):
self._batch_size = batch_size
self.num_buckets = num_buckets
self._epoch = 0
self._step = 1.
self.num_epochs = num_epochs
file_pattern = in_pattern + '/examples.proto' if os.path.isdir(in_pattern) else in_pattern
filenames = tf.matching_files(file_pattern)
# filenames = tf.Print(filenames, [filenames], message='filenames: ')
self.next_batch_op = self.input_pipeline(filenames, self._batch_size, self.num_buckets, self.num_epochs)
def example_parser(self, filename_queue):
reader = tf.TFRecordReader()
key, record_string = reader.read(filename_queue)
features = {
'labels': tf.FixedLenSequenceFeature([], tf.int64),
'tokens': tf.FixedLenSequenceFeature([], tf.int64),
'shapes': tf.FixedLenSequenceFeature([], tf.int64),
'chars': tf.FixedLenSequenceFeature([], tf.int64),
'seq_len': tf.FixedLenSequenceFeature([], tf.int64),
'tok_len': tf.FixedLenSequenceFeature([], tf.int64),
}
_, example = tf.parse_single_sequence_example(serialized=record_string, sequence_features=features)
labels = example['labels']
tokens = example['tokens']
shapes = example['shapes']
chars = example['chars']
seq_len = example['seq_len']
tok_len = example['tok_len']
# context = c['context']
return labels, tokens, shapes, chars, seq_len, tok_len
# return labels, tokens, labels, labels, labels
def input_pipeline(self, filenames, batch_size, num_buckets, num_epochs=None):
filename_queue = tf.train.string_input_producer(filenames, num_epochs=num_epochs, shuffle=True)
labels, tokens, shapes, chars, seq_len, tok_len = self.example_parser(filename_queue)
# min_after_dequeue defines how big a buffer we will randomly sample
# from -- bigger means better shuffling but slower start up and more
# memory used.
# capacity must be larger than min_after_dequeue and the amount larger
# determines the maximum we will prefetch. Recommendation:
# min_after_dequeue + (num_threads + a small safety margin) * batch_size
min_after_dequeue = 10000
capacity = min_after_dequeue + 12 * batch_size
# next_batch = tf.train.batch([labels, tokens, shapes, chars, seq_len], batch_size=batch_size, capacity=capacity,
# dynamic_pad=True, allow_smaller_final_batch=True)
if num_buckets == 0:
next_batch = tf.train.batch([labels, tokens, shapes, chars, seq_len, tok_len], batch_size=batch_size, capacity=capacity,
dynamic_pad=True, allow_smaller_final_batch=True)
else:
bucket, next_batch = tf.contrib.training.bucket([labels, tokens, shapes, chars, seq_len, tok_len], np.random.randint(num_buckets),
batch_size, num_buckets, num_threads=1, capacity=capacity,
dynamic_pad=True, allow_smaller_final_batch=False)
return next_batch
# class NodeBatcher(object):
# def __init__(self, in_dir, max_seq, max_word, batch_size, num_epochs=None):
# self._batch_size = batch_size
# self._max_seq = max_seq
# self._max_word = max_word
# self._epoch = 0
# self._step = 1.
# self.num_epochs = num_epochs
# in_file = [in_dir + '/examples.proto']
# self.next_batch_op = self.input_pipeline(in_file, self._max_seq, self._max_word, self._batch_size, self.num_epochs)
#
# def example_parser(self, filename_queue, max_seq, max_word):
# reader = tf.TFRecordReader()
# key, record_string = reader.read(filename_queue)
# features = {
# 'labels': tf.FixedLenFeature([max_seq], tf.int64),
# 'tokens': tf.FixedLenFeature([max_seq], tf.int64),
# 'shapes': tf.FixedLenFeature([max_seq], tf.int64),
# 'chars': tf.FixedLenFeature([], tf.int64),
# 'seq_len': tf.FixedLenFeature([], tf.int64),
# 'tok_len': tf.FixedLenFeature([], tf.int64)
# }
#
# example = tf.parse_single_example(record_string, features)
# labels = example['labels']
# tokens = example['tokens']
# shapes = example['shapes']
# chars = example['chars'][0]
# seq_len = example['seq_len'][0]
# tok_len = example['tok_len'][0]
# return labels, tokens, shapes, chars, seq_len, tok_len
#
# def input_pipeline(self, filenames, max_seq, max_word, batch_size, num_epochs=None):
# filename_queue = tf.train.string_input_producer(filenames, num_epochs=num_epochs, shuffle=True)
# labels, tokens, shapes, chars, seq_len, tok_len = self.example_parser(filename_queue, max_seq, max_word)
# # min_after_dequeue defines how big a buffer we will randomly sample
# # from -- bigger means better shuffling but slower start up and more
# # memory used.
# # capacity must be larger than min_after_dequeue and the amount larger
# # determines the maximum we will prefetch. Recommendation:
# # min_after_dequeue + (num_threads + a small safety margin) * batch_size
# min_after_dequeue = 10000
# capacity = min_after_dequeue + 12 * batch_size
# next_batch = tf.train.shuffle_batch(
# [labels, tokens, shapes, chars, seq_len, tok_len], batch_size=batch_size, capacity=capacity,
# min_after_dequeue=min_after_dequeue, allow_smaller_final_batch=True)
# return next_batch