Spaces:
Running
Running
File size: 21,798 Bytes
0b8359d |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 |
# Copyright 2019 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Tests for Keras-based transformer block layer."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import json
import numpy as np
import tensorflow as tf
from tensorflow.python.keras import keras_parameterized # pylint: disable=g-direct-tensorflow-import
from official.nlp.modeling.layers import attention
from official.nlp.modeling.layers import transformer_scaffold
# Test class that wraps a standard attention layer. If this layer is called
# at any point, the list passed to the config object will be filled with a
# boolean 'True'. We register this class as a Keras serializable so we can
# test serialization below.
@tf.keras.utils.register_keras_serializable(package='TestOnlyAttention')
class ValidatedAttentionLayer(attention.MultiHeadAttention):
def __init__(self, call_list, **kwargs):
super(ValidatedAttentionLayer, self).__init__(**kwargs)
self.list = call_list
def call(self, inputs, attention_mask=None):
self.list.append(True)
return super(ValidatedAttentionLayer, self).call(
inputs, attention_mask=attention_mask)
def get_config(self):
config = super(ValidatedAttentionLayer, self).get_config()
config['call_list'] = []
return config
# Test class implements a simple feedforward layer. If this layer is called
# at any point, the list passed to the config object will be filled with a
# boolean 'True'. We register this class as a Keras serializable so we can
# test serialization below.
@tf.keras.utils.register_keras_serializable(package='TestOnlyFeedforward')
class ValidatedFeedforwardLayer(tf.keras.layers.Layer):
def __init__(self, call_list, activation, **kwargs):
super(ValidatedFeedforwardLayer, self).__init__(**kwargs)
self.list = call_list
self.activation = activation
def build(self, input_shape):
hidden_size = input_shape.as_list()[-1]
self._feedforward_dense = tf.keras.layers.experimental.EinsumDense(
'...x,xy->...y',
output_shape=hidden_size,
bias_axes='y',
activation=self.activation,
name='feedforward')
def call(self, inputs):
self.list.append(True)
return self._feedforward_dense(inputs)
def get_config(self):
config = super(ValidatedFeedforwardLayer, self).get_config()
config['call_list'] = []
config['activation'] = self.activation
return config
# This decorator runs the test in V1, V2-Eager, and V2-Functional mode. It
# guarantees forward compatibility of this code for the V2 switchover.
@keras_parameterized.run_all_keras_modes
class TransformerLayerTest(keras_parameterized.TestCase):
def tearDown(self):
super(TransformerLayerTest, self).tearDown()
tf.keras.mixed_precision.experimental.set_policy('float32')
def test_layer_creation(self):
sequence_length = 21
width = 80
call_list = []
attention_layer_cfg = {
'num_heads': 10,
'key_size': 8,
'call_list': call_list,
}
test_layer = transformer_scaffold.TransformerScaffold(
attention_cls=ValidatedAttentionLayer,
attention_cfg=attention_layer_cfg,
num_attention_heads=10,
intermediate_size=2048,
intermediate_activation='relu')
# Create a 3-dimensional input (the first dimension is implicit).
data_tensor = tf.keras.Input(shape=(sequence_length, width))
output_tensor = test_layer(data_tensor)
# The default output of a transformer layer should be the same as the input.
self.assertEqual(data_tensor.shape.as_list(), output_tensor.shape.as_list())
# If call_list[0] exists and is True, the passed layer class was
# instantiated from the given config properly.
self.assertNotEmpty(call_list)
self.assertTrue(call_list[0], "The passed layer class wasn't instantiated.")
def test_layer_creation_with_feedforward_cls(self):
sequence_length = 21
width = 80
call_list = []
attention_layer_cfg = {
'num_heads': 10,
'key_size': 8,
'call_list': call_list,
}
feedforward_call_list = []
feedforward_layer_cfg = {
'activation': 'relu',
'call_list': feedforward_call_list,
}
test_layer = transformer_scaffold.TransformerScaffold(
attention_cls=ValidatedAttentionLayer,
attention_cfg=attention_layer_cfg,
feedforward_cls=ValidatedFeedforwardLayer,
feedforward_cfg=feedforward_layer_cfg,
num_attention_heads=10,
intermediate_size=None,
intermediate_activation=None)
# Create a 3-dimensional input (the first dimension is implicit).
data_tensor = tf.keras.Input(shape=(sequence_length, width))
output_tensor = test_layer(data_tensor)
# The default output of a transformer layer should be the same as the input.
self.assertEqual(data_tensor.shape.as_list(), output_tensor.shape.as_list())
# If call_list[0] exists and is True, the passed layer class was
# instantiated from the given config properly.
self.assertNotEmpty(call_list)
self.assertTrue(call_list[0], "The passed layer class wasn't instantiated.")
self.assertNotEmpty(feedforward_call_list)
self.assertTrue(feedforward_call_list[0],
"The passed layer class wasn't instantiated.")
def test_layer_creation_with_mask(self):
sequence_length = 21
width = 80
call_list = []
attention_layer_cfg = {
'num_heads': 10,
'key_size': 8,
'call_list': call_list,
}
test_layer = transformer_scaffold.TransformerScaffold(
attention_cls=ValidatedAttentionLayer,
attention_cfg=attention_layer_cfg,
num_attention_heads=10,
intermediate_size=2048,
intermediate_activation='relu')
# Create a 3-dimensional input (the first dimension is implicit).
data_tensor = tf.keras.Input(shape=(sequence_length, width))
# Create a 2-dimensional input (the first dimension is implicit).
mask_tensor = tf.keras.Input(shape=(sequence_length, sequence_length))
output_tensor = test_layer([data_tensor, mask_tensor])
# The default output of a transformer layer should be the same as the input.
self.assertEqual(data_tensor.shape.as_list(), output_tensor.shape.as_list())
# If call_list[0] exists and is True, the passed layer class was
# instantiated from the given config properly.
self.assertNotEmpty(call_list)
self.assertTrue(call_list[0], "The passed layer class wasn't instantiated.")
def test_layer_creation_with_incorrect_mask_fails(self):
sequence_length = 21
width = 80
call_list = []
attention_layer_cfg = {
'num_heads': 10,
'key_size': 8,
'call_list': call_list,
}
test_layer = transformer_scaffold.TransformerScaffold(
attention_cls=ValidatedAttentionLayer,
attention_cfg=attention_layer_cfg,
num_attention_heads=10,
intermediate_size=2048,
intermediate_activation='relu')
# Create a 3-dimensional input (the first dimension is implicit).
data_tensor = tf.keras.Input(shape=(sequence_length, width))
# Create a 2-dimensional input (the first dimension is implicit).
mask_tensor = tf.keras.Input(shape=(sequence_length, sequence_length - 3))
with self.assertRaisesRegex(ValueError, 'When passing a mask tensor.*'):
_ = test_layer([data_tensor, mask_tensor])
def test_layer_invocation(self):
sequence_length = 21
width = 80
call_list = []
attention_layer_cfg = {
'num_heads': 10,
'key_size': 8,
'call_list': call_list,
}
test_layer = transformer_scaffold.TransformerScaffold(
attention_cls=ValidatedAttentionLayer,
attention_cfg=attention_layer_cfg,
num_attention_heads=10,
intermediate_size=2048,
intermediate_activation='relu')
# Create a 3-dimensional input (the first dimension is implicit).
data_tensor = tf.keras.Input(shape=(sequence_length, width))
output_tensor = test_layer(data_tensor)
# Create a model from the test layer.
model = tf.keras.Model(data_tensor, output_tensor)
# Invoke the model on test data. We can't validate the output data itself
# (the NN is too complex) but this will rule out structural runtime errors.
batch_size = 6
input_data = 10 * np.random.random_sample(
(batch_size, sequence_length, width))
_ = model.predict(input_data)
# If call_list[0] exists and is True, the passed layer class was
# instantiated from the given config properly.
self.assertNotEmpty(call_list)
self.assertTrue(call_list[0], "The passed layer class wasn't instantiated.")
def test_layer_invocation_with_feedforward_cls(self):
sequence_length = 21
width = 80
call_list = []
attention_layer_cfg = {
'num_heads': 10,
'key_size': 8,
'call_list': call_list,
}
feedforward_call_list = []
feedforward_layer_cfg = {
'activation': 'relu',
'call_list': feedforward_call_list,
}
feedforward_layer = ValidatedFeedforwardLayer(**feedforward_layer_cfg)
test_layer = transformer_scaffold.TransformerScaffold(
attention_cls=ValidatedAttentionLayer,
attention_cfg=attention_layer_cfg,
feedforward_cls=feedforward_layer,
num_attention_heads=10,
intermediate_size=None,
intermediate_activation=None)
# Create a 3-dimensional input (the first dimension is implicit).
data_tensor = tf.keras.Input(shape=(sequence_length, width))
# Create a 2-dimensional input (the first dimension is implicit).
mask_tensor = tf.keras.Input(shape=(sequence_length, sequence_length))
output_tensor = test_layer([data_tensor, mask_tensor])
# Create a model from the test layer.
model = tf.keras.Model([data_tensor, mask_tensor], output_tensor)
# Invoke the model on test data. We can't validate the output data itself
# (the NN is too complex) but this will rule out structural runtime errors.
batch_size = 6
input_data = 10 * np.random.random_sample(
(batch_size, sequence_length, width))
# The attention mask should be of shape (batch, from_seq_len, to_seq_len),
# which here is (batch, sequence_length, sequence_length)
mask_data = np.random.randint(
2, size=(batch_size, sequence_length, sequence_length))
_ = model.predict([input_data, mask_data])
# If call_list[0] exists and is True, the passed layer class was
# instantiated from the given config properly.
self.assertNotEmpty(call_list)
self.assertTrue(call_list[0], "The passed layer class wasn't instantiated.")
self.assertNotEmpty(feedforward_call_list)
self.assertTrue(feedforward_call_list[0],
"The passed layer class wasn't instantiated.")
def test_layer_invocation_with_mask(self):
sequence_length = 21
width = 80
call_list = []
attention_layer_cfg = {
'num_heads': 10,
'key_size': 8,
'call_list': call_list,
}
test_layer = transformer_scaffold.TransformerScaffold(
attention_cls=ValidatedAttentionLayer,
attention_cfg=attention_layer_cfg,
num_attention_heads=10,
intermediate_size=2048,
intermediate_activation='relu')
# Create a 3-dimensional input (the first dimension is implicit).
data_tensor = tf.keras.Input(shape=(sequence_length, width))
# Create a 2-dimensional input (the first dimension is implicit).
mask_tensor = tf.keras.Input(shape=(sequence_length, sequence_length))
output_tensor = test_layer([data_tensor, mask_tensor])
# Create a model from the test layer.
model = tf.keras.Model([data_tensor, mask_tensor], output_tensor)
# Invoke the model on test data. We can't validate the output data itself
# (the NN is too complex) but this will rule out structural runtime errors.
batch_size = 6
input_data = 10 * np.random.random_sample(
(batch_size, sequence_length, width))
# The attention mask should be of shape (batch, from_seq_len, to_seq_len),
# which here is (batch, sequence_length, sequence_length)
mask_data = np.random.randint(
2, size=(batch_size, sequence_length, sequence_length))
_ = model.predict([input_data, mask_data])
# If call_list[0] exists and is True, the passed layer class was
# instantiated from the given config properly.
self.assertNotEmpty(call_list)
self.assertTrue(call_list[0], "The passed layer class wasn't instantiated.")
def test_layer_invocation_with_float16_dtype(self):
tf.keras.mixed_precision.experimental.set_policy('mixed_float16')
sequence_length = 21
width = 80
call_list = []
attention_layer_cfg = {
'num_heads': 10,
'key_size': 8,
'call_list': call_list,
}
test_layer = transformer_scaffold.TransformerScaffold(
attention_cls=ValidatedAttentionLayer,
attention_cfg=attention_layer_cfg,
num_attention_heads=10,
intermediate_size=2048,
intermediate_activation='relu')
# Create a 3-dimensional input (the first dimension is implicit).
data_tensor = tf.keras.Input(shape=(sequence_length, width))
# Create a 2-dimensional input (the first dimension is implicit).
mask_tensor = tf.keras.Input(shape=(sequence_length, sequence_length))
output_tensor = test_layer([data_tensor, mask_tensor])
# Create a model from the test layer.
model = tf.keras.Model([data_tensor, mask_tensor], output_tensor)
# Invoke the model on test data. We can't validate the output data itself
# (the NN is too complex) but this will rule out structural runtime errors.
batch_size = 6
input_data = (10 * np.random.random_sample(
(batch_size, sequence_length, width)))
# The attention mask should be of shape (batch, from_seq_len, to_seq_len),
# which here is (batch, sequence_length, sequence_length)
mask_data = np.random.randint(
2, size=(batch_size, sequence_length, sequence_length))
_ = model.predict([input_data, mask_data])
# If call_list[0] exists and is True, the passed layer class was
# instantiated from the given config properly.
self.assertNotEmpty(call_list)
self.assertTrue(call_list[0], "The passed layer class wasn't instantiated.")
def test_transform_with_initializer(self):
sequence_length = 21
width = 80
call_list = []
attention_layer_cfg = {
'num_heads': 10,
'key_size': 8,
'call_list': call_list,
}
test_layer = transformer_scaffold.TransformerScaffold(
attention_cls=ValidatedAttentionLayer,
attention_cfg=attention_layer_cfg,
num_attention_heads=10,
intermediate_size=2048,
intermediate_activation='relu',
kernel_initializer=tf.keras.initializers.TruncatedNormal(stddev=0.02))
# Create a 3-dimensional input (the first dimension is implicit).
data_tensor = tf.keras.Input(shape=(sequence_length, width))
output = test_layer(data_tensor)
# The default output of a transformer layer should be the same as the input.
self.assertEqual(data_tensor.shape.as_list(), output.shape.as_list())
# If call_list[0] exists and is True, the passed layer class was
# instantiated from the given config properly.
self.assertNotEmpty(call_list)
self.assertTrue(call_list[0])
def test_layer_restoration_from_config(self):
sequence_length = 21
width = 80
call_list = []
attention_layer_cfg = {
'num_heads': 10,
'key_size': 8,
'call_list': call_list,
'name': 'test_layer',
}
test_layer = transformer_scaffold.TransformerScaffold(
attention_cls=ValidatedAttentionLayer,
attention_cfg=attention_layer_cfg,
num_attention_heads=10,
intermediate_size=2048,
intermediate_activation='relu')
# Create a 3-dimensional input (the first dimension is implicit).
data_tensor = tf.keras.Input(shape=(sequence_length, width))
# Create a 2-dimensional input (the first dimension is implicit).
mask_tensor = tf.keras.Input(shape=(sequence_length, sequence_length))
output_tensor = test_layer([data_tensor, mask_tensor])
# Create a model from the test layer.
model = tf.keras.Model([data_tensor, mask_tensor], output_tensor)
# Invoke the model on test data. We can't validate the output data itself
# (the NN is too complex) but this will rule out structural runtime errors.
batch_size = 6
input_data = 10 * np.random.random_sample(
(batch_size, sequence_length, width))
# The attention mask should be of shape (batch, from_seq_len, to_seq_len),
# which here is (batch, sequence_length, sequence_length)
mask_data = np.random.randint(
2, size=(batch_size, sequence_length, sequence_length))
pre_serialization_output = model.predict([input_data, mask_data])
# Serialize the model config. Pass the serialized data through json to
# ensure that we can serialize this layer to disk.
serialized_data = json.dumps(model.get_config())
post_string_serialized_data = json.loads(serialized_data)
# Create a new model from the old config, and copy the weights. These models
# should have identical outputs.
new_model = tf.keras.Model.from_config(post_string_serialized_data)
new_model.set_weights(model.get_weights())
output = new_model.predict([input_data, mask_data])
self.assertAllClose(pre_serialization_output, output)
# If the layer was configured correctly, it should have a list attribute
# (since it should have the custom class and config passed to it).
new_model.summary()
new_call_list = new_model.get_layer(
name='transformer_scaffold')._attention_layer.list
self.assertNotEmpty(new_call_list)
self.assertTrue(new_call_list[0],
"The passed layer class wasn't instantiated.")
def test_layer_with_feedforward_cls_restoration_from_config(self):
sequence_length = 21
width = 80
call_list = []
attention_layer_cfg = {
'num_heads': 10,
'key_size': 8,
'call_list': call_list,
'name': 'test_layer',
}
feedforward_call_list = []
feedforward_layer_cfg = {
'activation': 'relu',
'call_list': feedforward_call_list,
}
test_layer = transformer_scaffold.TransformerScaffold(
attention_cls=ValidatedAttentionLayer,
attention_cfg=attention_layer_cfg,
feedforward_cls=ValidatedFeedforwardLayer,
feedforward_cfg=feedforward_layer_cfg,
num_attention_heads=10,
intermediate_size=None,
intermediate_activation=None)
# Create a 3-dimensional input (the first dimension is implicit).
data_tensor = tf.keras.Input(shape=(sequence_length, width))
# Create a 2-dimensional input (the first dimension is implicit).
mask_tensor = tf.keras.Input(shape=(sequence_length, sequence_length))
output_tensor = test_layer([data_tensor, mask_tensor])
# Create a model from the test layer.
model = tf.keras.Model([data_tensor, mask_tensor], output_tensor)
# Invoke the model on test data. We can't validate the output data itself
# (the NN is too complex) but this will rule out structural runtime errors.
batch_size = 6
input_data = 10 * np.random.random_sample(
(batch_size, sequence_length, width))
# The attention mask should be of shape (batch, from_seq_len, to_seq_len),
# which here is (batch, sequence_length, sequence_length)
mask_data = np.random.randint(
2, size=(batch_size, sequence_length, sequence_length))
pre_serialization_output = model.predict([input_data, mask_data])
# Serialize the model config. Pass the serialized data through json to
# ensure that we can serialize this layer to disk.
serialized_data = json.dumps(model.get_config())
post_string_serialized_data = json.loads(serialized_data)
# Create a new model from the old config, and copy the weights. These models
# should have identical outputs.
new_model = tf.keras.Model.from_config(post_string_serialized_data)
new_model.set_weights(model.get_weights())
output = new_model.predict([input_data, mask_data])
self.assertAllClose(pre_serialization_output, output)
# If the layer was configured correctly, it should have a list attribute
# (since it should have the custom class and config passed to it).
new_model.summary()
new_call_list = new_model.get_layer(
name='transformer_scaffold')._attention_layer.list
self.assertNotEmpty(new_call_list)
self.assertTrue(new_call_list[0],
"The passed layer class wasn't instantiated.")
new_feedforward_call_list = new_model.get_layer(
name='transformer_scaffold')._feedforward_block.list
self.assertNotEmpty(new_feedforward_call_list)
self.assertTrue(new_feedforward_call_list[0],
"The passed layer class wasn't instantiated.")
if __name__ == '__main__':
tf.test.main()
|