SinaTools 0.1.40__py2.py3-none-any.whl → 1.0.1__py2.py3-none-any.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- {SinaTools-0.1.40.dist-info → SinaTools-1.0.1.dist-info}/METADATA +1 -1
- SinaTools-1.0.1.dist-info/RECORD +73 -0
- sinatools/VERSION +1 -1
- sinatools/ner/__init__.py +5 -7
- sinatools/ner/trainers/BertNestedTrainer.py +203 -203
- sinatools/ner/trainers/BertTrainer.py +163 -163
- sinatools/ner/trainers/__init__.py +2 -2
- SinaTools-0.1.40.dist-info/RECORD +0 -123
- sinatools/arabert/arabert/__init__.py +0 -14
- sinatools/arabert/arabert/create_classification_data.py +0 -260
- sinatools/arabert/arabert/create_pretraining_data.py +0 -534
- sinatools/arabert/arabert/extract_features.py +0 -444
- sinatools/arabert/arabert/lamb_optimizer.py +0 -158
- sinatools/arabert/arabert/modeling.py +0 -1027
- sinatools/arabert/arabert/optimization.py +0 -202
- sinatools/arabert/arabert/run_classifier.py +0 -1078
- sinatools/arabert/arabert/run_pretraining.py +0 -593
- sinatools/arabert/arabert/run_squad.py +0 -1440
- sinatools/arabert/arabert/tokenization.py +0 -414
- sinatools/arabert/araelectra/__init__.py +0 -1
- sinatools/arabert/araelectra/build_openwebtext_pretraining_dataset.py +0 -103
- sinatools/arabert/araelectra/build_pretraining_dataset.py +0 -230
- sinatools/arabert/araelectra/build_pretraining_dataset_single_file.py +0 -90
- sinatools/arabert/araelectra/configure_finetuning.py +0 -172
- sinatools/arabert/araelectra/configure_pretraining.py +0 -143
- sinatools/arabert/araelectra/finetune/__init__.py +0 -14
- sinatools/arabert/araelectra/finetune/feature_spec.py +0 -56
- sinatools/arabert/araelectra/finetune/preprocessing.py +0 -173
- sinatools/arabert/araelectra/finetune/scorer.py +0 -54
- sinatools/arabert/araelectra/finetune/task.py +0 -74
- sinatools/arabert/araelectra/finetune/task_builder.py +0 -70
- sinatools/arabert/araelectra/flops_computation.py +0 -215
- sinatools/arabert/araelectra/model/__init__.py +0 -14
- sinatools/arabert/araelectra/model/modeling.py +0 -1029
- sinatools/arabert/araelectra/model/optimization.py +0 -193
- sinatools/arabert/araelectra/model/tokenization.py +0 -355
- sinatools/arabert/araelectra/pretrain/__init__.py +0 -14
- sinatools/arabert/araelectra/pretrain/pretrain_data.py +0 -160
- sinatools/arabert/araelectra/pretrain/pretrain_helpers.py +0 -229
- sinatools/arabert/araelectra/run_finetuning.py +0 -323
- sinatools/arabert/araelectra/run_pretraining.py +0 -469
- sinatools/arabert/araelectra/util/__init__.py +0 -14
- sinatools/arabert/araelectra/util/training_utils.py +0 -112
- sinatools/arabert/araelectra/util/utils.py +0 -109
- sinatools/arabert/aragpt2/__init__.py +0 -2
- sinatools/arabert/aragpt2/create_pretraining_data.py +0 -95
- sinatools/arabert/aragpt2/gpt2/__init__.py +0 -2
- sinatools/arabert/aragpt2/gpt2/lamb_optimizer.py +0 -158
- sinatools/arabert/aragpt2/gpt2/optimization.py +0 -225
- sinatools/arabert/aragpt2/gpt2/run_pretraining.py +0 -397
- sinatools/arabert/aragpt2/grover/__init__.py +0 -0
- sinatools/arabert/aragpt2/grover/dataloader.py +0 -161
- sinatools/arabert/aragpt2/grover/modeling.py +0 -803
- sinatools/arabert/aragpt2/grover/modeling_gpt2.py +0 -1196
- sinatools/arabert/aragpt2/grover/optimization_adafactor.py +0 -234
- sinatools/arabert/aragpt2/grover/train_tpu.py +0 -187
- sinatools/arabert/aragpt2/grover/utils.py +0 -234
- sinatools/arabert/aragpt2/train_bpe_tokenizer.py +0 -59
- {SinaTools-0.1.40.data → SinaTools-1.0.1.data}/data/sinatools/environment.yml +0 -0
- {SinaTools-0.1.40.dist-info → SinaTools-1.0.1.dist-info}/AUTHORS.rst +0 -0
- {SinaTools-0.1.40.dist-info → SinaTools-1.0.1.dist-info}/LICENSE +0 -0
- {SinaTools-0.1.40.dist-info → SinaTools-1.0.1.dist-info}/WHEEL +0 -0
- {SinaTools-0.1.40.dist-info → SinaTools-1.0.1.dist-info}/entry_points.txt +0 -0
- {SinaTools-0.1.40.dist-info → SinaTools-1.0.1.dist-info}/top_level.txt +0 -0
@@ -1,1027 +0,0 @@
|
|
1
|
-
# coding=utf-8
|
2
|
-
# Copyright 2018 The Google AI Language Team Authors.
|
3
|
-
#
|
4
|
-
# Licensed under the Apache License, Version 2.0 (the "License");
|
5
|
-
# you may not use this file except in compliance with the License.
|
6
|
-
# You may obtain a copy of the License at
|
7
|
-
#
|
8
|
-
# http://www.apache.org/licenses/LICENSE-2.0
|
9
|
-
#
|
10
|
-
# Unless required by applicable law or agreed to in writing, software
|
11
|
-
# distributed under the License is distributed on an "AS IS" BASIS,
|
12
|
-
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
13
|
-
# See the License for the specific language governing permissions and
|
14
|
-
# limitations under the License.
|
15
|
-
"""The main BERT model and related functions."""
|
16
|
-
|
17
|
-
from __future__ import absolute_import
|
18
|
-
from __future__ import division
|
19
|
-
from __future__ import print_function
|
20
|
-
|
21
|
-
import collections
|
22
|
-
import copy
|
23
|
-
import json
|
24
|
-
import math
|
25
|
-
import re
|
26
|
-
import numpy as np
|
27
|
-
import six
|
28
|
-
import tensorflow as tf
|
29
|
-
|
30
|
-
|
31
|
-
class BertConfig(object):
|
32
|
-
"""Configuration for `BertModel`."""
|
33
|
-
|
34
|
-
def __init__(
|
35
|
-
self,
|
36
|
-
vocab_size,
|
37
|
-
hidden_size=768,
|
38
|
-
num_hidden_layers=12,
|
39
|
-
num_attention_heads=12,
|
40
|
-
intermediate_size=3072,
|
41
|
-
hidden_act="gelu",
|
42
|
-
hidden_dropout_prob=0.1,
|
43
|
-
attention_probs_dropout_prob=0.1,
|
44
|
-
max_position_embeddings=512,
|
45
|
-
type_vocab_size=16,
|
46
|
-
initializer_range=0.02,
|
47
|
-
):
|
48
|
-
"""Constructs BertConfig.
|
49
|
-
|
50
|
-
Args:
|
51
|
-
vocab_size: Vocabulary size of `inputs_ids` in `BertModel`.
|
52
|
-
hidden_size: Size of the encoder layers and the pooler layer.
|
53
|
-
num_hidden_layers: Number of hidden layers in the Transformer encoder.
|
54
|
-
num_attention_heads: Number of attention heads for each attention layer in
|
55
|
-
the Transformer encoder.
|
56
|
-
intermediate_size: The size of the "intermediate" (i.e., feed-forward)
|
57
|
-
layer in the Transformer encoder.
|
58
|
-
hidden_act: The non-linear activation function (function or string) in the
|
59
|
-
encoder and pooler.
|
60
|
-
hidden_dropout_prob: The dropout probability for all fully connected
|
61
|
-
layers in the embeddings, encoder, and pooler.
|
62
|
-
attention_probs_dropout_prob: The dropout ratio for the attention
|
63
|
-
probabilities.
|
64
|
-
max_position_embeddings: The maximum sequence length that this model might
|
65
|
-
ever be used with. Typically set this to something large just in case
|
66
|
-
(e.g., 512 or 1024 or 2048).
|
67
|
-
type_vocab_size: The vocabulary size of the `token_type_ids` passed into
|
68
|
-
`BertModel`.
|
69
|
-
initializer_range: The stdev of the truncated_normal_initializer for
|
70
|
-
initializing all weight matrices.
|
71
|
-
"""
|
72
|
-
self.vocab_size = vocab_size
|
73
|
-
self.hidden_size = hidden_size
|
74
|
-
self.num_hidden_layers = num_hidden_layers
|
75
|
-
self.num_attention_heads = num_attention_heads
|
76
|
-
self.hidden_act = hidden_act
|
77
|
-
self.intermediate_size = intermediate_size
|
78
|
-
self.hidden_dropout_prob = hidden_dropout_prob
|
79
|
-
self.attention_probs_dropout_prob = attention_probs_dropout_prob
|
80
|
-
self.max_position_embeddings = max_position_embeddings
|
81
|
-
self.type_vocab_size = type_vocab_size
|
82
|
-
self.initializer_range = initializer_range
|
83
|
-
|
84
|
-
@classmethod
|
85
|
-
def from_dict(cls, json_object):
|
86
|
-
"""Constructs a `BertConfig` from a Python dictionary of parameters."""
|
87
|
-
config = BertConfig(vocab_size=None)
|
88
|
-
for (key, value) in six.iteritems(json_object):
|
89
|
-
config.__dict__[key] = value
|
90
|
-
return config
|
91
|
-
|
92
|
-
@classmethod
|
93
|
-
def from_json_file(cls, json_file):
|
94
|
-
"""Constructs a `BertConfig` from a json file of parameters."""
|
95
|
-
with tf.gfile.GFile(json_file, "r") as reader:
|
96
|
-
text = reader.read()
|
97
|
-
return cls.from_dict(json.loads(text))
|
98
|
-
|
99
|
-
def to_dict(self):
|
100
|
-
"""Serializes this instance to a Python dictionary."""
|
101
|
-
output = copy.deepcopy(self.__dict__)
|
102
|
-
return output
|
103
|
-
|
104
|
-
def to_json_string(self):
|
105
|
-
"""Serializes this instance to a JSON string."""
|
106
|
-
return json.dumps(self.to_dict(), indent=2, sort_keys=True) + "\n"
|
107
|
-
|
108
|
-
|
109
|
-
class BertModel(object):
|
110
|
-
"""BERT model ("Bidirectional Encoder Representations from Transformers").
|
111
|
-
|
112
|
-
Example usage:
|
113
|
-
|
114
|
-
```python
|
115
|
-
# Already been converted into WordPiece token ids
|
116
|
-
input_ids = tf.constant([[31, 51, 99], [15, 5, 0]])
|
117
|
-
input_mask = tf.constant([[1, 1, 1], [1, 1, 0]])
|
118
|
-
token_type_ids = tf.constant([[0, 0, 1], [0, 2, 0]])
|
119
|
-
|
120
|
-
config = modeling.BertConfig(vocab_size=32000, hidden_size=512,
|
121
|
-
num_hidden_layers=8, num_attention_heads=6, intermediate_size=1024)
|
122
|
-
|
123
|
-
model = modeling.BertModel(config=config, is_training=True,
|
124
|
-
input_ids=input_ids, input_mask=input_mask, token_type_ids=token_type_ids)
|
125
|
-
|
126
|
-
label_embeddings = tf.get_variable(...)
|
127
|
-
pooled_output = model.get_pooled_output()
|
128
|
-
logits = tf.matmul(pooled_output, label_embeddings)
|
129
|
-
...
|
130
|
-
```
|
131
|
-
"""
|
132
|
-
|
133
|
-
def __init__(
|
134
|
-
self,
|
135
|
-
config,
|
136
|
-
is_training,
|
137
|
-
input_ids,
|
138
|
-
input_mask=None,
|
139
|
-
token_type_ids=None,
|
140
|
-
use_one_hot_embeddings=False,
|
141
|
-
scope=None,
|
142
|
-
):
|
143
|
-
"""Constructor for BertModel.
|
144
|
-
|
145
|
-
Args:
|
146
|
-
config: `BertConfig` instance.
|
147
|
-
is_training: bool. true for training model, false for eval model. Controls
|
148
|
-
whether dropout will be applied.
|
149
|
-
input_ids: int32 Tensor of shape [batch_size, seq_length].
|
150
|
-
input_mask: (optional) int32 Tensor of shape [batch_size, seq_length].
|
151
|
-
token_type_ids: (optional) int32 Tensor of shape [batch_size, seq_length].
|
152
|
-
use_one_hot_embeddings: (optional) bool. Whether to use one-hot word
|
153
|
-
embeddings or tf.embedding_lookup() for the word embeddings.
|
154
|
-
scope: (optional) variable scope. Defaults to "bert".
|
155
|
-
|
156
|
-
Raises:
|
157
|
-
ValueError: The config is invalid or one of the input tensor shapes
|
158
|
-
is invalid.
|
159
|
-
"""
|
160
|
-
config = copy.deepcopy(config)
|
161
|
-
if not is_training:
|
162
|
-
config.hidden_dropout_prob = 0.0
|
163
|
-
config.attention_probs_dropout_prob = 0.0
|
164
|
-
|
165
|
-
input_shape = get_shape_list(input_ids, expected_rank=2)
|
166
|
-
batch_size = input_shape[0]
|
167
|
-
seq_length = input_shape[1]
|
168
|
-
|
169
|
-
if input_mask is None:
|
170
|
-
input_mask = tf.ones(shape=[batch_size, seq_length], dtype=tf.int32)
|
171
|
-
|
172
|
-
if token_type_ids is None:
|
173
|
-
token_type_ids = tf.zeros(shape=[batch_size, seq_length], dtype=tf.int32)
|
174
|
-
|
175
|
-
with tf.variable_scope(scope, default_name="bert"):
|
176
|
-
with tf.variable_scope("embeddings"):
|
177
|
-
# Perform embedding lookup on the word ids.
|
178
|
-
(self.embedding_output, self.embedding_table) = embedding_lookup(
|
179
|
-
input_ids=input_ids,
|
180
|
-
vocab_size=config.vocab_size,
|
181
|
-
embedding_size=config.hidden_size,
|
182
|
-
initializer_range=config.initializer_range,
|
183
|
-
word_embedding_name="word_embeddings",
|
184
|
-
use_one_hot_embeddings=use_one_hot_embeddings,
|
185
|
-
)
|
186
|
-
|
187
|
-
# Add positional embeddings and token type embeddings, then layer
|
188
|
-
# normalize and perform dropout.
|
189
|
-
self.embedding_output = embedding_postprocessor(
|
190
|
-
input_tensor=self.embedding_output,
|
191
|
-
use_token_type=True,
|
192
|
-
token_type_ids=token_type_ids,
|
193
|
-
token_type_vocab_size=config.type_vocab_size,
|
194
|
-
token_type_embedding_name="token_type_embeddings",
|
195
|
-
use_position_embeddings=True,
|
196
|
-
position_embedding_name="position_embeddings",
|
197
|
-
initializer_range=config.initializer_range,
|
198
|
-
max_position_embeddings=config.max_position_embeddings,
|
199
|
-
dropout_prob=config.hidden_dropout_prob,
|
200
|
-
)
|
201
|
-
|
202
|
-
with tf.variable_scope("encoder"):
|
203
|
-
# This converts a 2D mask of shape [batch_size, seq_length] to a 3D
|
204
|
-
# mask of shape [batch_size, seq_length, seq_length] which is used
|
205
|
-
# for the attention scores.
|
206
|
-
attention_mask = create_attention_mask_from_input_mask(
|
207
|
-
input_ids, input_mask
|
208
|
-
)
|
209
|
-
|
210
|
-
# Run the stacked transformer.
|
211
|
-
# `sequence_output` shape = [batch_size, seq_length, hidden_size].
|
212
|
-
self.all_encoder_layers = transformer_model(
|
213
|
-
input_tensor=self.embedding_output,
|
214
|
-
attention_mask=attention_mask,
|
215
|
-
hidden_size=config.hidden_size,
|
216
|
-
num_hidden_layers=config.num_hidden_layers,
|
217
|
-
num_attention_heads=config.num_attention_heads,
|
218
|
-
intermediate_size=config.intermediate_size,
|
219
|
-
intermediate_act_fn=get_activation(config.hidden_act),
|
220
|
-
hidden_dropout_prob=config.hidden_dropout_prob,
|
221
|
-
attention_probs_dropout_prob=config.attention_probs_dropout_prob,
|
222
|
-
initializer_range=config.initializer_range,
|
223
|
-
do_return_all_layers=True,
|
224
|
-
)
|
225
|
-
|
226
|
-
self.sequence_output = self.all_encoder_layers[-1]
|
227
|
-
# The "pooler" converts the encoded sequence tensor of shape
|
228
|
-
# [batch_size, seq_length, hidden_size] to a tensor of shape
|
229
|
-
# [batch_size, hidden_size]. This is necessary for segment-level
|
230
|
-
# (or segment-pair-level) classification tasks where we need a fixed
|
231
|
-
# dimensional representation of the segment.
|
232
|
-
with tf.variable_scope("pooler"):
|
233
|
-
# We "pool" the model by simply taking the hidden state corresponding
|
234
|
-
# to the first token. We assume that this has been pre-trained
|
235
|
-
first_token_tensor = tf.squeeze(self.sequence_output[:, 0:1, :], axis=1)
|
236
|
-
self.pooled_output = tf.layers.dense(
|
237
|
-
first_token_tensor,
|
238
|
-
config.hidden_size,
|
239
|
-
activation=tf.tanh,
|
240
|
-
kernel_initializer=create_initializer(config.initializer_range),
|
241
|
-
)
|
242
|
-
|
243
|
-
def get_pooled_output(self):
|
244
|
-
return self.pooled_output
|
245
|
-
|
246
|
-
def get_sequence_output(self):
|
247
|
-
"""Gets final hidden layer of encoder.
|
248
|
-
|
249
|
-
Returns:
|
250
|
-
float Tensor of shape [batch_size, seq_length, hidden_size] corresponding
|
251
|
-
to the final hidden of the transformer encoder.
|
252
|
-
"""
|
253
|
-
return self.sequence_output
|
254
|
-
|
255
|
-
def get_all_encoder_layers(self):
|
256
|
-
return self.all_encoder_layers
|
257
|
-
|
258
|
-
def get_embedding_output(self):
|
259
|
-
"""Gets output of the embedding lookup (i.e., input to the transformer).
|
260
|
-
|
261
|
-
Returns:
|
262
|
-
float Tensor of shape [batch_size, seq_length, hidden_size] corresponding
|
263
|
-
to the output of the embedding layer, after summing the word
|
264
|
-
embeddings with the positional embeddings and the token type embeddings,
|
265
|
-
then performing layer normalization. This is the input to the transformer.
|
266
|
-
"""
|
267
|
-
return self.embedding_output
|
268
|
-
|
269
|
-
def get_embedding_table(self):
|
270
|
-
return self.embedding_table
|
271
|
-
|
272
|
-
|
273
|
-
def gelu(x):
|
274
|
-
"""Gaussian Error Linear Unit.
|
275
|
-
|
276
|
-
This is a smoother version of the RELU.
|
277
|
-
Original paper: https://arxiv.org/abs/1606.08415
|
278
|
-
Args:
|
279
|
-
x: float Tensor to perform activation.
|
280
|
-
|
281
|
-
Returns:
|
282
|
-
`x` with the GELU activation applied.
|
283
|
-
"""
|
284
|
-
cdf = 0.5 * (1.0 + tf.tanh((np.sqrt(2 / np.pi) * (x + 0.044715 * tf.pow(x, 3)))))
|
285
|
-
return x * cdf
|
286
|
-
|
287
|
-
|
288
|
-
def get_activation(activation_string):
|
289
|
-
"""Maps a string to a Python function, e.g., "relu" => `tf.nn.relu`.
|
290
|
-
|
291
|
-
Args:
|
292
|
-
activation_string: String name of the activation function.
|
293
|
-
|
294
|
-
Returns:
|
295
|
-
A Python function corresponding to the activation function. If
|
296
|
-
`activation_string` is None, empty, or "linear", this will return None.
|
297
|
-
If `activation_string` is not a string, it will return `activation_string`.
|
298
|
-
|
299
|
-
Raises:
|
300
|
-
ValueError: The `activation_string` does not correspond to a known
|
301
|
-
activation.
|
302
|
-
"""
|
303
|
-
|
304
|
-
# We assume that anything that"s not a string is already an activation
|
305
|
-
# function, so we just return it.
|
306
|
-
if not isinstance(activation_string, six.string_types):
|
307
|
-
return activation_string
|
308
|
-
|
309
|
-
if not activation_string:
|
310
|
-
return None
|
311
|
-
|
312
|
-
act = activation_string.lower()
|
313
|
-
if act == "linear":
|
314
|
-
return None
|
315
|
-
elif act == "relu":
|
316
|
-
return tf.nn.relu
|
317
|
-
elif act == "gelu":
|
318
|
-
return gelu
|
319
|
-
elif act == "tanh":
|
320
|
-
return tf.tanh
|
321
|
-
else:
|
322
|
-
raise ValueError("Unsupported activation: %s" % act)
|
323
|
-
|
324
|
-
|
325
|
-
def get_assignment_map_from_checkpoint(tvars, init_checkpoint):
|
326
|
-
"""Compute the union of the current variables and checkpoint variables."""
|
327
|
-
assignment_map = {}
|
328
|
-
initialized_variable_names = {}
|
329
|
-
|
330
|
-
name_to_variable = collections.OrderedDict()
|
331
|
-
for var in tvars:
|
332
|
-
name = var.name
|
333
|
-
m = re.match("^(.*):\\d+$", name)
|
334
|
-
if m is not None:
|
335
|
-
name = m.group(1)
|
336
|
-
name_to_variable[name] = var
|
337
|
-
|
338
|
-
init_vars = tf.train.list_variables(init_checkpoint)
|
339
|
-
|
340
|
-
assignment_map = collections.OrderedDict()
|
341
|
-
for x in init_vars:
|
342
|
-
(name, var) = (x[0], x[1])
|
343
|
-
if name not in name_to_variable:
|
344
|
-
continue
|
345
|
-
assignment_map[name] = name
|
346
|
-
initialized_variable_names[name] = 1
|
347
|
-
initialized_variable_names[name + ":0"] = 1
|
348
|
-
|
349
|
-
return (assignment_map, initialized_variable_names)
|
350
|
-
|
351
|
-
|
352
|
-
def dropout(input_tensor, dropout_prob):
|
353
|
-
"""Perform dropout.
|
354
|
-
|
355
|
-
Args:
|
356
|
-
input_tensor: float Tensor.
|
357
|
-
dropout_prob: Python float. The probability of dropping out a value (NOT of
|
358
|
-
*keeping* a dimension as in `tf.nn.dropout`).
|
359
|
-
|
360
|
-
Returns:
|
361
|
-
A version of `input_tensor` with dropout applied.
|
362
|
-
"""
|
363
|
-
if dropout_prob is None or dropout_prob == 0.0:
|
364
|
-
return input_tensor
|
365
|
-
|
366
|
-
output = tf.nn.dropout(input_tensor, 1.0 - dropout_prob)
|
367
|
-
return output
|
368
|
-
|
369
|
-
|
370
|
-
def layer_norm(input_tensor, name=None):
|
371
|
-
"""Run layer normalization on the last dimension of the tensor."""
|
372
|
-
return tf.contrib.layers.layer_norm(
|
373
|
-
inputs=input_tensor, begin_norm_axis=-1, begin_params_axis=-1, scope=name
|
374
|
-
)
|
375
|
-
|
376
|
-
|
377
|
-
def layer_norm_and_dropout(input_tensor, dropout_prob, name=None):
|
378
|
-
"""Runs layer normalization followed by dropout."""
|
379
|
-
output_tensor = layer_norm(input_tensor, name)
|
380
|
-
output_tensor = dropout(output_tensor, dropout_prob)
|
381
|
-
return output_tensor
|
382
|
-
|
383
|
-
|
384
|
-
def create_initializer(initializer_range=0.02):
|
385
|
-
"""Creates a `truncated_normal_initializer` with the given range."""
|
386
|
-
return tf.truncated_normal_initializer(stddev=initializer_range)
|
387
|
-
|
388
|
-
|
389
|
-
def embedding_lookup(
|
390
|
-
input_ids,
|
391
|
-
vocab_size,
|
392
|
-
embedding_size=128,
|
393
|
-
initializer_range=0.02,
|
394
|
-
word_embedding_name="word_embeddings",
|
395
|
-
use_one_hot_embeddings=False,
|
396
|
-
):
|
397
|
-
"""Looks up words embeddings for id tensor.
|
398
|
-
|
399
|
-
Args:
|
400
|
-
input_ids: int32 Tensor of shape [batch_size, seq_length] containing word
|
401
|
-
ids.
|
402
|
-
vocab_size: int. Size of the embedding vocabulary.
|
403
|
-
embedding_size: int. Width of the word embeddings.
|
404
|
-
initializer_range: float. Embedding initialization range.
|
405
|
-
word_embedding_name: string. Name of the embedding table.
|
406
|
-
use_one_hot_embeddings: bool. If True, use one-hot method for word
|
407
|
-
embeddings. If False, use `tf.gather()`.
|
408
|
-
|
409
|
-
Returns:
|
410
|
-
float Tensor of shape [batch_size, seq_length, embedding_size].
|
411
|
-
"""
|
412
|
-
# This function assumes that the input is of shape [batch_size, seq_length,
|
413
|
-
# num_inputs].
|
414
|
-
#
|
415
|
-
# If the input is a 2D tensor of shape [batch_size, seq_length], we
|
416
|
-
# reshape to [batch_size, seq_length, 1].
|
417
|
-
if input_ids.shape.ndims == 2:
|
418
|
-
input_ids = tf.expand_dims(input_ids, axis=[-1])
|
419
|
-
|
420
|
-
embedding_table = tf.get_variable(
|
421
|
-
name=word_embedding_name,
|
422
|
-
shape=[vocab_size, embedding_size],
|
423
|
-
initializer=create_initializer(initializer_range),
|
424
|
-
)
|
425
|
-
|
426
|
-
flat_input_ids = tf.reshape(input_ids, [-1])
|
427
|
-
if use_one_hot_embeddings:
|
428
|
-
one_hot_input_ids = tf.one_hot(flat_input_ids, depth=vocab_size)
|
429
|
-
output = tf.matmul(one_hot_input_ids, embedding_table)
|
430
|
-
else:
|
431
|
-
output = tf.gather(embedding_table, flat_input_ids)
|
432
|
-
|
433
|
-
input_shape = get_shape_list(input_ids)
|
434
|
-
|
435
|
-
output = tf.reshape(output, input_shape[0:-1] + [input_shape[-1] * embedding_size])
|
436
|
-
return (output, embedding_table)
|
437
|
-
|
438
|
-
|
439
|
-
def embedding_postprocessor(
|
440
|
-
input_tensor,
|
441
|
-
use_token_type=False,
|
442
|
-
token_type_ids=None,
|
443
|
-
token_type_vocab_size=16,
|
444
|
-
token_type_embedding_name="token_type_embeddings",
|
445
|
-
use_position_embeddings=True,
|
446
|
-
position_embedding_name="position_embeddings",
|
447
|
-
initializer_range=0.02,
|
448
|
-
max_position_embeddings=512,
|
449
|
-
dropout_prob=0.1,
|
450
|
-
):
|
451
|
-
"""Performs various post-processing on a word embedding tensor.
|
452
|
-
|
453
|
-
Args:
|
454
|
-
input_tensor: float Tensor of shape [batch_size, seq_length,
|
455
|
-
embedding_size].
|
456
|
-
use_token_type: bool. Whether to add embeddings for `token_type_ids`.
|
457
|
-
token_type_ids: (optional) int32 Tensor of shape [batch_size, seq_length].
|
458
|
-
Must be specified if `use_token_type` is True.
|
459
|
-
token_type_vocab_size: int. The vocabulary size of `token_type_ids`.
|
460
|
-
token_type_embedding_name: string. The name of the embedding table variable
|
461
|
-
for token type ids.
|
462
|
-
use_position_embeddings: bool. Whether to add position embeddings for the
|
463
|
-
position of each token in the sequence.
|
464
|
-
position_embedding_name: string. The name of the embedding table variable
|
465
|
-
for positional embeddings.
|
466
|
-
initializer_range: float. Range of the weight initialization.
|
467
|
-
max_position_embeddings: int. Maximum sequence length that might ever be
|
468
|
-
used with this model. This can be longer than the sequence length of
|
469
|
-
input_tensor, but cannot be shorter.
|
470
|
-
dropout_prob: float. Dropout probability applied to the final output tensor.
|
471
|
-
|
472
|
-
Returns:
|
473
|
-
float tensor with same shape as `input_tensor`.
|
474
|
-
|
475
|
-
Raises:
|
476
|
-
ValueError: One of the tensor shapes or input values is invalid.
|
477
|
-
"""
|
478
|
-
input_shape = get_shape_list(input_tensor, expected_rank=3)
|
479
|
-
batch_size = input_shape[0]
|
480
|
-
seq_length = input_shape[1]
|
481
|
-
width = input_shape[2]
|
482
|
-
|
483
|
-
output = input_tensor
|
484
|
-
|
485
|
-
if use_token_type:
|
486
|
-
if token_type_ids is None:
|
487
|
-
raise ValueError(
|
488
|
-
"`token_type_ids` must be specified if" "`use_token_type` is True."
|
489
|
-
)
|
490
|
-
token_type_table = tf.get_variable(
|
491
|
-
name=token_type_embedding_name,
|
492
|
-
shape=[token_type_vocab_size, width],
|
493
|
-
initializer=create_initializer(initializer_range),
|
494
|
-
)
|
495
|
-
# This vocab will be small so we always do one-hot here, since it is always
|
496
|
-
# faster for a small vocabulary.
|
497
|
-
flat_token_type_ids = tf.reshape(token_type_ids, [-1])
|
498
|
-
one_hot_ids = tf.one_hot(flat_token_type_ids, depth=token_type_vocab_size)
|
499
|
-
token_type_embeddings = tf.matmul(one_hot_ids, token_type_table)
|
500
|
-
token_type_embeddings = tf.reshape(
|
501
|
-
token_type_embeddings, [batch_size, seq_length, width]
|
502
|
-
)
|
503
|
-
output += token_type_embeddings
|
504
|
-
|
505
|
-
if use_position_embeddings:
|
506
|
-
assert_op = tf.assert_less_equal(seq_length, max_position_embeddings)
|
507
|
-
with tf.control_dependencies([assert_op]):
|
508
|
-
full_position_embeddings = tf.get_variable(
|
509
|
-
name=position_embedding_name,
|
510
|
-
shape=[max_position_embeddings, width],
|
511
|
-
initializer=create_initializer(initializer_range),
|
512
|
-
)
|
513
|
-
# Since the position embedding table is a learned variable, we create it
|
514
|
-
# using a (long) sequence length `max_position_embeddings`. The actual
|
515
|
-
# sequence length might be shorter than this, for faster training of
|
516
|
-
# tasks that do not have long sequences.
|
517
|
-
#
|
518
|
-
# So `full_position_embeddings` is effectively an embedding table
|
519
|
-
# for position [0, 1, 2, ..., max_position_embeddings-1], and the current
|
520
|
-
# sequence has positions [0, 1, 2, ... seq_length-1], so we can just
|
521
|
-
# perform a slice.
|
522
|
-
position_embeddings = tf.slice(
|
523
|
-
full_position_embeddings, [0, 0], [seq_length, -1]
|
524
|
-
)
|
525
|
-
num_dims = len(output.shape.as_list())
|
526
|
-
|
527
|
-
# Only the last two dimensions are relevant (`seq_length` and `width`), so
|
528
|
-
# we broadcast among the first dimensions, which is typically just
|
529
|
-
# the batch size.
|
530
|
-
position_broadcast_shape = []
|
531
|
-
for _ in range(num_dims - 2):
|
532
|
-
position_broadcast_shape.append(1)
|
533
|
-
position_broadcast_shape.extend([seq_length, width])
|
534
|
-
position_embeddings = tf.reshape(
|
535
|
-
position_embeddings, position_broadcast_shape
|
536
|
-
)
|
537
|
-
output += position_embeddings
|
538
|
-
|
539
|
-
output = layer_norm_and_dropout(output, dropout_prob)
|
540
|
-
return output
|
541
|
-
|
542
|
-
|
543
|
-
def create_attention_mask_from_input_mask(from_tensor, to_mask):
|
544
|
-
"""Create 3D attention mask from a 2D tensor mask.
|
545
|
-
|
546
|
-
Args:
|
547
|
-
from_tensor: 2D or 3D Tensor of shape [batch_size, from_seq_length, ...].
|
548
|
-
to_mask: int32 Tensor of shape [batch_size, to_seq_length].
|
549
|
-
|
550
|
-
Returns:
|
551
|
-
float Tensor of shape [batch_size, from_seq_length, to_seq_length].
|
552
|
-
"""
|
553
|
-
from_shape = get_shape_list(from_tensor, expected_rank=[2, 3])
|
554
|
-
batch_size = from_shape[0]
|
555
|
-
from_seq_length = from_shape[1]
|
556
|
-
|
557
|
-
to_shape = get_shape_list(to_mask, expected_rank=2)
|
558
|
-
to_seq_length = to_shape[1]
|
559
|
-
|
560
|
-
to_mask = tf.cast(tf.reshape(to_mask, [batch_size, 1, to_seq_length]), tf.float32)
|
561
|
-
|
562
|
-
# We don't assume that `from_tensor` is a mask (although it could be). We
|
563
|
-
# don't actually care if we attend *from* padding tokens (only *to* padding)
|
564
|
-
# tokens so we create a tensor of all ones.
|
565
|
-
#
|
566
|
-
# `broadcast_ones` = [batch_size, from_seq_length, 1]
|
567
|
-
broadcast_ones = tf.ones(shape=[batch_size, from_seq_length, 1], dtype=tf.float32)
|
568
|
-
|
569
|
-
# Here we broadcast along two dimensions to create the mask.
|
570
|
-
mask = broadcast_ones * to_mask
|
571
|
-
|
572
|
-
return mask
|
573
|
-
|
574
|
-
|
575
|
-
def attention_layer(
|
576
|
-
from_tensor,
|
577
|
-
to_tensor,
|
578
|
-
attention_mask=None,
|
579
|
-
num_attention_heads=1,
|
580
|
-
size_per_head=512,
|
581
|
-
query_act=None,
|
582
|
-
key_act=None,
|
583
|
-
value_act=None,
|
584
|
-
attention_probs_dropout_prob=0.0,
|
585
|
-
initializer_range=0.02,
|
586
|
-
do_return_2d_tensor=False,
|
587
|
-
batch_size=None,
|
588
|
-
from_seq_length=None,
|
589
|
-
to_seq_length=None,
|
590
|
-
):
|
591
|
-
"""Performs multi-headed attention from `from_tensor` to `to_tensor`.
|
592
|
-
|
593
|
-
This is an implementation of multi-headed attention based on "Attention
|
594
|
-
is all you Need". If `from_tensor` and `to_tensor` are the same, then
|
595
|
-
this is self-attention. Each timestep in `from_tensor` attends to the
|
596
|
-
corresponding sequence in `to_tensor`, and returns a fixed-with vector.
|
597
|
-
|
598
|
-
This function first projects `from_tensor` into a "query" tensor and
|
599
|
-
`to_tensor` into "key" and "value" tensors. These are (effectively) a list
|
600
|
-
of tensors of length `num_attention_heads`, where each tensor is of shape
|
601
|
-
[batch_size, seq_length, size_per_head].
|
602
|
-
|
603
|
-
Then, the query and key tensors are dot-producted and scaled. These are
|
604
|
-
softmaxed to obtain attention probabilities. The value tensors are then
|
605
|
-
interpolated by these probabilities, then concatenated back to a single
|
606
|
-
tensor and returned.
|
607
|
-
|
608
|
-
In practice, the multi-headed attention are done with transposes and
|
609
|
-
reshapes rather than actual separate tensors.
|
610
|
-
|
611
|
-
Args:
|
612
|
-
from_tensor: float Tensor of shape [batch_size, from_seq_length,
|
613
|
-
from_width].
|
614
|
-
to_tensor: float Tensor of shape [batch_size, to_seq_length, to_width].
|
615
|
-
attention_mask: (optional) int32 Tensor of shape [batch_size,
|
616
|
-
from_seq_length, to_seq_length]. The values should be 1 or 0. The
|
617
|
-
attention scores will effectively be set to -infinity for any positions in
|
618
|
-
the mask that are 0, and will be unchanged for positions that are 1.
|
619
|
-
num_attention_heads: int. Number of attention heads.
|
620
|
-
size_per_head: int. Size of each attention head.
|
621
|
-
query_act: (optional) Activation function for the query transform.
|
622
|
-
key_act: (optional) Activation function for the key transform.
|
623
|
-
value_act: (optional) Activation function for the value transform.
|
624
|
-
attention_probs_dropout_prob: (optional) float. Dropout probability of the
|
625
|
-
attention probabilities.
|
626
|
-
initializer_range: float. Range of the weight initializer.
|
627
|
-
do_return_2d_tensor: bool. If True, the output will be of shape [batch_size
|
628
|
-
* from_seq_length, num_attention_heads * size_per_head]. If False, the
|
629
|
-
output will be of shape [batch_size, from_seq_length, num_attention_heads
|
630
|
-
* size_per_head].
|
631
|
-
batch_size: (Optional) int. If the input is 2D, this might be the batch size
|
632
|
-
of the 3D version of the `from_tensor` and `to_tensor`.
|
633
|
-
from_seq_length: (Optional) If the input is 2D, this might be the seq length
|
634
|
-
of the 3D version of the `from_tensor`.
|
635
|
-
to_seq_length: (Optional) If the input is 2D, this might be the seq length
|
636
|
-
of the 3D version of the `to_tensor`.
|
637
|
-
|
638
|
-
Returns:
|
639
|
-
float Tensor of shape [batch_size, from_seq_length,
|
640
|
-
num_attention_heads * size_per_head]. (If `do_return_2d_tensor` is
|
641
|
-
true, this will be of shape [batch_size * from_seq_length,
|
642
|
-
num_attention_heads * size_per_head]).
|
643
|
-
|
644
|
-
Raises:
|
645
|
-
ValueError: Any of the arguments or tensor shapes are invalid.
|
646
|
-
"""
|
647
|
-
|
648
|
-
def transpose_for_scores(
|
649
|
-
input_tensor, batch_size, num_attention_heads, seq_length, width
|
650
|
-
):
|
651
|
-
output_tensor = tf.reshape(
|
652
|
-
input_tensor, [batch_size, seq_length, num_attention_heads, width]
|
653
|
-
)
|
654
|
-
|
655
|
-
output_tensor = tf.transpose(output_tensor, [0, 2, 1, 3])
|
656
|
-
return output_tensor
|
657
|
-
|
658
|
-
from_shape = get_shape_list(from_tensor, expected_rank=[2, 3])
|
659
|
-
to_shape = get_shape_list(to_tensor, expected_rank=[2, 3])
|
660
|
-
|
661
|
-
if len(from_shape) != len(to_shape):
|
662
|
-
raise ValueError(
|
663
|
-
"The rank of `from_tensor` must match the rank of `to_tensor`."
|
664
|
-
)
|
665
|
-
|
666
|
-
if len(from_shape) == 3:
|
667
|
-
batch_size = from_shape[0]
|
668
|
-
from_seq_length = from_shape[1]
|
669
|
-
to_seq_length = to_shape[1]
|
670
|
-
elif len(from_shape) == 2:
|
671
|
-
if batch_size is None or from_seq_length is None or to_seq_length is None:
|
672
|
-
raise ValueError(
|
673
|
-
"When passing in rank 2 tensors to attention_layer, the values "
|
674
|
-
"for `batch_size`, `from_seq_length`, and `to_seq_length` "
|
675
|
-
"must all be specified."
|
676
|
-
)
|
677
|
-
|
678
|
-
# Scalar dimensions referenced here:
|
679
|
-
# B = batch size (number of sequences)
|
680
|
-
# F = `from_tensor` sequence length
|
681
|
-
# T = `to_tensor` sequence length
|
682
|
-
# N = `num_attention_heads`
|
683
|
-
# H = `size_per_head`
|
684
|
-
|
685
|
-
from_tensor_2d = reshape_to_matrix(from_tensor)
|
686
|
-
to_tensor_2d = reshape_to_matrix(to_tensor)
|
687
|
-
|
688
|
-
# `query_layer` = [B*F, N*H]
|
689
|
-
query_layer = tf.layers.dense(
|
690
|
-
from_tensor_2d,
|
691
|
-
num_attention_heads * size_per_head,
|
692
|
-
activation=query_act,
|
693
|
-
name="query",
|
694
|
-
kernel_initializer=create_initializer(initializer_range),
|
695
|
-
)
|
696
|
-
|
697
|
-
# `key_layer` = [B*T, N*H]
|
698
|
-
key_layer = tf.layers.dense(
|
699
|
-
to_tensor_2d,
|
700
|
-
num_attention_heads * size_per_head,
|
701
|
-
activation=key_act,
|
702
|
-
name="key",
|
703
|
-
kernel_initializer=create_initializer(initializer_range),
|
704
|
-
)
|
705
|
-
|
706
|
-
# `value_layer` = [B*T, N*H]
|
707
|
-
value_layer = tf.layers.dense(
|
708
|
-
to_tensor_2d,
|
709
|
-
num_attention_heads * size_per_head,
|
710
|
-
activation=value_act,
|
711
|
-
name="value",
|
712
|
-
kernel_initializer=create_initializer(initializer_range),
|
713
|
-
)
|
714
|
-
|
715
|
-
# `query_layer` = [B, N, F, H]
|
716
|
-
query_layer = transpose_for_scores(
|
717
|
-
query_layer, batch_size, num_attention_heads, from_seq_length, size_per_head
|
718
|
-
)
|
719
|
-
|
720
|
-
# `key_layer` = [B, N, T, H]
|
721
|
-
key_layer = transpose_for_scores(
|
722
|
-
key_layer, batch_size, num_attention_heads, to_seq_length, size_per_head
|
723
|
-
)
|
724
|
-
|
725
|
-
# Take the dot product between "query" and "key" to get the raw
|
726
|
-
# attention scores.
|
727
|
-
# `attention_scores` = [B, N, F, T]
|
728
|
-
attention_scores = tf.matmul(query_layer, key_layer, transpose_b=True)
|
729
|
-
attention_scores = tf.multiply(
|
730
|
-
attention_scores, 1.0 / math.sqrt(float(size_per_head))
|
731
|
-
)
|
732
|
-
|
733
|
-
if attention_mask is not None:
|
734
|
-
# `attention_mask` = [B, 1, F, T]
|
735
|
-
attention_mask = tf.expand_dims(attention_mask, axis=[1])
|
736
|
-
|
737
|
-
# Since attention_mask is 1.0 for positions we want to attend and 0.0 for
|
738
|
-
# masked positions, this operation will create a tensor which is 0.0 for
|
739
|
-
# positions we want to attend and -10000.0 for masked positions.
|
740
|
-
adder = (1.0 - tf.cast(attention_mask, tf.float32)) * -10000.0
|
741
|
-
|
742
|
-
# Since we are adding it to the raw scores before the softmax, this is
|
743
|
-
# effectively the same as removing these entirely.
|
744
|
-
attention_scores += adder
|
745
|
-
|
746
|
-
# Normalize the attention scores to probabilities.
|
747
|
-
# `attention_probs` = [B, N, F, T]
|
748
|
-
attention_probs = tf.nn.softmax(attention_scores)
|
749
|
-
|
750
|
-
# This is actually dropping out entire tokens to attend to, which might
|
751
|
-
# seem a bit unusual, but is taken from the original Transformer paper.
|
752
|
-
attention_probs = dropout(attention_probs, attention_probs_dropout_prob)
|
753
|
-
|
754
|
-
# `value_layer` = [B, T, N, H]
|
755
|
-
value_layer = tf.reshape(
|
756
|
-
value_layer, [batch_size, to_seq_length, num_attention_heads, size_per_head]
|
757
|
-
)
|
758
|
-
|
759
|
-
# `value_layer` = [B, N, T, H]
|
760
|
-
value_layer = tf.transpose(value_layer, [0, 2, 1, 3])
|
761
|
-
|
762
|
-
# `context_layer` = [B, N, F, H]
|
763
|
-
context_layer = tf.matmul(attention_probs, value_layer)
|
764
|
-
|
765
|
-
# `context_layer` = [B, F, N, H]
|
766
|
-
context_layer = tf.transpose(context_layer, [0, 2, 1, 3])
|
767
|
-
|
768
|
-
if do_return_2d_tensor:
|
769
|
-
# `context_layer` = [B*F, N*H]
|
770
|
-
context_layer = tf.reshape(
|
771
|
-
context_layer,
|
772
|
-
[batch_size * from_seq_length, num_attention_heads * size_per_head],
|
773
|
-
)
|
774
|
-
else:
|
775
|
-
# `context_layer` = [B, F, N*H]
|
776
|
-
context_layer = tf.reshape(
|
777
|
-
context_layer,
|
778
|
-
[batch_size, from_seq_length, num_attention_heads * size_per_head],
|
779
|
-
)
|
780
|
-
|
781
|
-
return context_layer
|
782
|
-
|
783
|
-
|
784
|
-
def transformer_model(
|
785
|
-
input_tensor,
|
786
|
-
attention_mask=None,
|
787
|
-
hidden_size=768,
|
788
|
-
num_hidden_layers=12,
|
789
|
-
num_attention_heads=12,
|
790
|
-
intermediate_size=3072,
|
791
|
-
intermediate_act_fn=gelu,
|
792
|
-
hidden_dropout_prob=0.1,
|
793
|
-
attention_probs_dropout_prob=0.1,
|
794
|
-
initializer_range=0.02,
|
795
|
-
do_return_all_layers=False,
|
796
|
-
):
|
797
|
-
"""Multi-headed, multi-layer Transformer from "Attention is All You Need".
|
798
|
-
|
799
|
-
This is almost an exact implementation of the original Transformer encoder.
|
800
|
-
|
801
|
-
See the original paper:
|
802
|
-
https://arxiv.org/abs/1706.03762
|
803
|
-
|
804
|
-
Also see:
|
805
|
-
https://github.com/tensorflow/tensor2tensor/blob/master/tensor2tensor/models/transformer.py
|
806
|
-
|
807
|
-
Args:
|
808
|
-
input_tensor: float Tensor of shape [batch_size, seq_length, hidden_size].
|
809
|
-
attention_mask: (optional) int32 Tensor of shape [batch_size, seq_length,
|
810
|
-
seq_length], with 1 for positions that can be attended to and 0 in
|
811
|
-
positions that should not be.
|
812
|
-
hidden_size: int. Hidden size of the Transformer.
|
813
|
-
num_hidden_layers: int. Number of layers (blocks) in the Transformer.
|
814
|
-
num_attention_heads: int. Number of attention heads in the Transformer.
|
815
|
-
intermediate_size: int. The size of the "intermediate" (a.k.a., feed
|
816
|
-
forward) layer.
|
817
|
-
intermediate_act_fn: function. The non-linear activation function to apply
|
818
|
-
to the output of the intermediate/feed-forward layer.
|
819
|
-
hidden_dropout_prob: float. Dropout probability for the hidden layers.
|
820
|
-
attention_probs_dropout_prob: float. Dropout probability of the attention
|
821
|
-
probabilities.
|
822
|
-
initializer_range: float. Range of the initializer (stddev of truncated
|
823
|
-
normal).
|
824
|
-
do_return_all_layers: Whether to also return all layers or just the final
|
825
|
-
layer.
|
826
|
-
|
827
|
-
Returns:
|
828
|
-
float Tensor of shape [batch_size, seq_length, hidden_size], the final
|
829
|
-
hidden layer of the Transformer.
|
830
|
-
|
831
|
-
Raises:
|
832
|
-
ValueError: A Tensor shape or parameter is invalid.
|
833
|
-
"""
|
834
|
-
if hidden_size % num_attention_heads != 0:
|
835
|
-
raise ValueError(
|
836
|
-
"The hidden size (%d) is not a multiple of the number of attention "
|
837
|
-
"heads (%d)" % (hidden_size, num_attention_heads)
|
838
|
-
)
|
839
|
-
|
840
|
-
attention_head_size = int(hidden_size / num_attention_heads)
|
841
|
-
input_shape = get_shape_list(input_tensor, expected_rank=3)
|
842
|
-
batch_size = input_shape[0]
|
843
|
-
seq_length = input_shape[1]
|
844
|
-
input_width = input_shape[2]
|
845
|
-
|
846
|
-
# The Transformer performs sum residuals on all layers so the input needs
|
847
|
-
# to be the same as the hidden size.
|
848
|
-
if input_width != hidden_size:
|
849
|
-
raise ValueError(
|
850
|
-
"The width of the input tensor (%d) != hidden size (%d)"
|
851
|
-
% (input_width, hidden_size)
|
852
|
-
)
|
853
|
-
|
854
|
-
# We keep the representation as a 2D tensor to avoid re-shaping it back and
|
855
|
-
# forth from a 3D tensor to a 2D tensor. Re-shapes are normally free on
|
856
|
-
# the GPU/CPU but may not be free on the TPU, so we want to minimize them to
|
857
|
-
# help the optimizer.
|
858
|
-
prev_output = reshape_to_matrix(input_tensor)
|
859
|
-
|
860
|
-
all_layer_outputs = []
|
861
|
-
for layer_idx in range(num_hidden_layers):
|
862
|
-
with tf.variable_scope("layer_%d" % layer_idx):
|
863
|
-
layer_input = prev_output
|
864
|
-
|
865
|
-
with tf.variable_scope("attention"):
|
866
|
-
attention_heads = []
|
867
|
-
with tf.variable_scope("self"):
|
868
|
-
attention_head = attention_layer(
|
869
|
-
from_tensor=layer_input,
|
870
|
-
to_tensor=layer_input,
|
871
|
-
attention_mask=attention_mask,
|
872
|
-
num_attention_heads=num_attention_heads,
|
873
|
-
size_per_head=attention_head_size,
|
874
|
-
attention_probs_dropout_prob=attention_probs_dropout_prob,
|
875
|
-
initializer_range=initializer_range,
|
876
|
-
do_return_2d_tensor=True,
|
877
|
-
batch_size=batch_size,
|
878
|
-
from_seq_length=seq_length,
|
879
|
-
to_seq_length=seq_length,
|
880
|
-
)
|
881
|
-
attention_heads.append(attention_head)
|
882
|
-
|
883
|
-
attention_output = None
|
884
|
-
if len(attention_heads) == 1:
|
885
|
-
attention_output = attention_heads[0]
|
886
|
-
else:
|
887
|
-
# In the case where we have other sequences, we just concatenate
|
888
|
-
# them to the self-attention head before the projection.
|
889
|
-
attention_output = tf.concat(attention_heads, axis=-1)
|
890
|
-
|
891
|
-
# Run a linear projection of `hidden_size` then add a residual
|
892
|
-
# with `layer_input`.
|
893
|
-
with tf.variable_scope("output"):
|
894
|
-
attention_output = tf.layers.dense(
|
895
|
-
attention_output,
|
896
|
-
hidden_size,
|
897
|
-
kernel_initializer=create_initializer(initializer_range),
|
898
|
-
)
|
899
|
-
attention_output = dropout(attention_output, hidden_dropout_prob)
|
900
|
-
attention_output = layer_norm(attention_output + layer_input)
|
901
|
-
|
902
|
-
# The activation is only applied to the "intermediate" hidden layer.
|
903
|
-
with tf.variable_scope("intermediate"):
|
904
|
-
intermediate_output = tf.layers.dense(
|
905
|
-
attention_output,
|
906
|
-
intermediate_size,
|
907
|
-
activation=intermediate_act_fn,
|
908
|
-
kernel_initializer=create_initializer(initializer_range),
|
909
|
-
)
|
910
|
-
|
911
|
-
# Down-project back to `hidden_size` then add the residual.
|
912
|
-
with tf.variable_scope("output"):
|
913
|
-
layer_output = tf.layers.dense(
|
914
|
-
intermediate_output,
|
915
|
-
hidden_size,
|
916
|
-
kernel_initializer=create_initializer(initializer_range),
|
917
|
-
)
|
918
|
-
layer_output = dropout(layer_output, hidden_dropout_prob)
|
919
|
-
layer_output = layer_norm(layer_output + attention_output)
|
920
|
-
prev_output = layer_output
|
921
|
-
all_layer_outputs.append(layer_output)
|
922
|
-
|
923
|
-
if do_return_all_layers:
|
924
|
-
final_outputs = []
|
925
|
-
for layer_output in all_layer_outputs:
|
926
|
-
final_output = reshape_from_matrix(layer_output, input_shape)
|
927
|
-
final_outputs.append(final_output)
|
928
|
-
return final_outputs
|
929
|
-
else:
|
930
|
-
final_output = reshape_from_matrix(prev_output, input_shape)
|
931
|
-
return final_output
|
932
|
-
|
933
|
-
|
934
|
-
def get_shape_list(tensor, expected_rank=None, name=None):
|
935
|
-
"""Returns a list of the shape of tensor, preferring static dimensions.
|
936
|
-
|
937
|
-
Args:
|
938
|
-
tensor: A tf.Tensor object to find the shape of.
|
939
|
-
expected_rank: (optional) int. The expected rank of `tensor`. If this is
|
940
|
-
specified and the `tensor` has a different rank, and exception will be
|
941
|
-
thrown.
|
942
|
-
name: Optional name of the tensor for the error message.
|
943
|
-
|
944
|
-
Returns:
|
945
|
-
A list of dimensions of the shape of tensor. All static dimensions will
|
946
|
-
be returned as python integers, and dynamic dimensions will be returned
|
947
|
-
as tf.Tensor scalars.
|
948
|
-
"""
|
949
|
-
if name is None:
|
950
|
-
name = tensor.name
|
951
|
-
|
952
|
-
if expected_rank is not None:
|
953
|
-
assert_rank(tensor, expected_rank, name)
|
954
|
-
|
955
|
-
shape = tensor.shape.as_list()
|
956
|
-
|
957
|
-
non_static_indexes = []
|
958
|
-
for (index, dim) in enumerate(shape):
|
959
|
-
if dim is None:
|
960
|
-
non_static_indexes.append(index)
|
961
|
-
|
962
|
-
if not non_static_indexes:
|
963
|
-
return shape
|
964
|
-
|
965
|
-
dyn_shape = tf.shape(tensor)
|
966
|
-
for index in non_static_indexes:
|
967
|
-
shape[index] = dyn_shape[index]
|
968
|
-
return shape
|
969
|
-
|
970
|
-
|
971
|
-
def reshape_to_matrix(input_tensor):
|
972
|
-
"""Reshapes a >= rank 2 tensor to a rank 2 tensor (i.e., a matrix)."""
|
973
|
-
ndims = input_tensor.shape.ndims
|
974
|
-
if ndims < 2:
|
975
|
-
raise ValueError(
|
976
|
-
"Input tensor must have at least rank 2. Shape = %s" % (input_tensor.shape)
|
977
|
-
)
|
978
|
-
if ndims == 2:
|
979
|
-
return input_tensor
|
980
|
-
|
981
|
-
width = input_tensor.shape[-1]
|
982
|
-
output_tensor = tf.reshape(input_tensor, [-1, width])
|
983
|
-
return output_tensor
|
984
|
-
|
985
|
-
|
986
|
-
def reshape_from_matrix(output_tensor, orig_shape_list):
|
987
|
-
"""Reshapes a rank 2 tensor back to its original rank >= 2 tensor."""
|
988
|
-
if len(orig_shape_list) == 2:
|
989
|
-
return output_tensor
|
990
|
-
|
991
|
-
output_shape = get_shape_list(output_tensor)
|
992
|
-
|
993
|
-
orig_dims = orig_shape_list[0:-1]
|
994
|
-
width = output_shape[-1]
|
995
|
-
|
996
|
-
return tf.reshape(output_tensor, orig_dims + [width])
|
997
|
-
|
998
|
-
|
999
|
-
def assert_rank(tensor, expected_rank, name=None):
|
1000
|
-
"""Raises an exception if the tensor rank is not of the expected rank.
|
1001
|
-
|
1002
|
-
Args:
|
1003
|
-
tensor: A tf.Tensor to check the rank of.
|
1004
|
-
expected_rank: Python integer or list of integers, expected rank.
|
1005
|
-
name: Optional name of the tensor for the error message.
|
1006
|
-
|
1007
|
-
Raises:
|
1008
|
-
ValueError: If the expected shape doesn't match the actual shape.
|
1009
|
-
"""
|
1010
|
-
if name is None:
|
1011
|
-
name = tensor.name
|
1012
|
-
|
1013
|
-
expected_rank_dict = {}
|
1014
|
-
if isinstance(expected_rank, six.integer_types):
|
1015
|
-
expected_rank_dict[expected_rank] = True
|
1016
|
-
else:
|
1017
|
-
for x in expected_rank:
|
1018
|
-
expected_rank_dict[x] = True
|
1019
|
-
|
1020
|
-
actual_rank = tensor.shape.ndims
|
1021
|
-
if actual_rank not in expected_rank_dict:
|
1022
|
-
scope_name = tf.get_variable_scope().name
|
1023
|
-
raise ValueError(
|
1024
|
-
"For the tensor `%s` in scope `%s`, the actual rank "
|
1025
|
-
"`%d` (shape = %s) is not equal to the expected rank `%s`"
|
1026
|
-
% (name, scope_name, actual_rank, str(tensor.shape), str(expected_rank))
|
1027
|
-
)
|