forked from 4kssoft/CUTIE
-
Notifications
You must be signed in to change notification settings - Fork 1
/
bert_embedding.py
186 lines (155 loc) · 7.74 KB
/
bert_embedding.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
# written by Xiaohui Zhao
# 2018-01
import tensorflow as tf
import numpy as np
import argparse
import os
os.environ["CUDA_VISIBLE_DEVICES"] = "1"
from model_cutie import CUTIE
from model_cutie_res import CUTIERes
from model_cutie_unet8 import CUTIEUNet
from data_loader_json import DataLoader
parser = argparse.ArgumentParser(description='CUTIE parameters')
# data
parser.add_argument('--doc_path', type=str, default='data/taxi_small')
parser.add_argument('--save_prefix', type=str, default='taxi', help='prefix for ckpt') # TBD: save log/models with prefix
# dict
parser.add_argument('--dict_path', type=str, default='dict/119547') # not used if load_dict is True
parser.add_argument('--load_dict', type=bool, default=True, help='True to work based on an existing dict')
parser.add_argument('--load_dict_from_path', type=str, default='dict/119547')
parser.add_argument('--large_dict', type=bool, default=True, help='True to use large dict for future ext')
# ckpt
parser.add_argument('--restore_ckpt', type=bool, default=True)
parser.add_argument('--restore_bertembedding_only', type=bool, default=True)
parser.add_argument('--embedding_file', type=str, default='../graph/bert/multi_cased_L-12_H-768_A-12/bert_model.ckpt')
parser.add_argument('--bert_dict_file', type=str, default='dict/vocab.txt')
# training
parser.add_argument('--embedding_size', type=int, default=768)
params = parser.parse_args()
class BertEmbedding(object):
def __init__(self, vocab_size=119547, hidden_size=768, initializer_range=0.02, use_one_hot_embeddings=False, trainable=False):
self.trainable = trainable
self.word_embedding_name = "word_embeddings"
self.input_ids = tf.placeholder(tf.int32, shape=[None, None, None, 1], name='grid_table')
with tf.variable_scope("bert"):
with tf.variable_scope("embeddings"):
# Perform embedding lookup on the word ids.
(self.embedding_output, self.embedding_table) = self.embedding_lookup(
input_ids=self.input_ids,
vocab_size=vocab_size,
embedding_size=hidden_size,
initializer_range=initializer_range,
word_embedding_name=self.word_embedding_name,
use_one_hot_embeddings=use_one_hot_embeddings)
def embedding_lookup(self, input_ids,
vocab_size,
embedding_size=128,
initializer_range=0.02,
word_embedding_name="word_embeddings",
use_one_hot_embeddings=False):
"""Looks up words embeddings for id tensor.
Args:
input_ids: int32 Tensor of shape [batch_size, seq_length] containing word
ids.
vocab_size: int. Size of the embedding vocabulary.
embedding_size: int. Width of the word embeddings.
initializer_range: float. Embedding initialization range.
word_embedding_name: string. Name of the embedding table.
use_one_hot_embeddings: bool. If True, use one-hot method for word
embeddings. If False, use `tf.nn.embedding_lookup()`. One hot is better
for TPUs.
Returns:
float Tensor of shape [batch_size, seq_length, embedding_size].
"""
# This function assumes that the input is of shape [batch_size, seq_length,
# num_inputs].
#
# If the input is a 2D tensor of shape [batch_size, seq_length], we
# reshape to [batch_size, seq_length, 1].
if input_ids.shape.ndims == 2:
input_ids = tf.expand_dims(input_ids, axis=[-1])
embedding_table = tf.get_variable(
name=word_embedding_name,
shape=[vocab_size, embedding_size],
initializer=self.create_initializer(initializer_range),
trainable=self.trainable)
if use_one_hot_embeddings:
flat_input_ids = tf.reshape(input_ids, [-1])
one_hot_input_ids = tf.one_hot(flat_input_ids, depth=vocab_size)
output = tf.matmul(one_hot_input_ids, embedding_table)
else:
output = tf.nn.embedding_lookup(embedding_table, input_ids)
input_shape = self.get_shape_list(input_ids)
output = tf.reshape(output,
input_shape[0:-1] + [input_shape[-1] * embedding_size])
return (output, embedding_table)
def create_initializer(self, initializer_range=0.02):
"""Creates a `truncated_normal_initializer` with the given range."""
return tf.truncated_normal_initializer(stddev=initializer_range)
def get_shape_list(self, tensor, expected_rank=None, name=None):
"""Returns a list of the shape of tensor, preferring static dimensions.
Args:
tensor: A tf.Tensor object to find the shape of.
expected_rank: (optional) int. The expected rank of `tensor`. If this is
specified and the `tensor` has a different rank, and exception will be
thrown.
name: Optional name of the tensor for the error message.
Returns:
A list of dimensions of the shape of tensor. All static dimensions will
be returned as python integers, and dynamic dimensions will be returned
as tf.Tensor scalars.
"""
if name is None:
name = tensor.name
if expected_rank is not None:
assert_rank(tensor, expected_rank, name)
shape = tensor.shape.as_list()
non_static_indexes = []
for (index, dim) in enumerate(shape):
if dim is None:
non_static_indexes.append(index)
if not non_static_indexes:
return shape
dyn_shape = tf.shape(tensor)
for index in non_static_indexes:
shape[index] = dyn_shape[index]
return shape
if __name__ == '__main__':
# data
data_loader = DataLoader(params, update_dict=False, load_dictionary=params.load_dict)
# save bert dictionary
with open(params.bert_dict_file, encoding='utf-8') as f:
vocabs = f.read().split('\n')
num_words = len(vocabs)
dictionary = {vocab:0 for vocab in vocabs}
word_to_index = dict(list(zip(dictionary.keys(), list(range(num_words)))))
index_to_word = dict(list(zip(list(range(num_words)), dictionary.keys())))
np.save(params.dict_path + '_dictionary.npy', dictionary)
np.save(params.dict_path + '_word_to_index.npy', word_to_index)
np.save(params.dict_path + '_index_to_word.npy', index_to_word)
# model
bert = BertEmbedding()
config = tf.ConfigProto(allow_soft_placement=False)
with tf.Session(config=config) as sess:
sess.run(tf.global_variables_initializer())
if params.restore_ckpt:
if params.restore_bertembedding_only:
try:
load_variable = {"bert/embeddings/word_embeddings": bert.embedding_table}
ckpt_saver = tf.train.Saver(load_variable, max_to_keep=50)
ckpt_path = params.embedding_file
ckpt = tf.train.get_checkpoint_state(ckpt_path)
print('Restoring from {}...'.format(ckpt_path))
ckpt_saver.restore(sess, ckpt_path)
print('Restored from {}'.format(ckpt_path))
except:
raise('Check your path {:s}'.format(ckpt_path))
data = data_loader.next_batch()
# one step training
feed_dict = {
bert.input_ids: data['grid_table'],
}
fetches = [bert.embedding_output, bert.embedding_table]
(output, table) = sess.run(fetches=fetches, feed_dict=feed_dict)
print(output[0][0][0])