File size: 10,621 Bytes
18ddfe2 |
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 |
# 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
from absl.testing import parameterized
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 transformer
# 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
@parameterized.named_parameters(('base', transformer.Transformer),
('xla', transformer.CompiledTransformer))
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, transformer_cls):
test_layer = transformer_cls(
num_attention_heads=10,
intermediate_size=2048,
intermediate_activation='relu')
sequence_length = 21
width = 80
# 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())
def test_layer_creation_with_mask(self, transformer_cls):
test_layer = transformer_cls(
num_attention_heads=10,
intermediate_size=2048,
intermediate_activation='relu')
sequence_length = 21
width = 80
# 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())
def test_layer_creation_with_incorrect_mask_fails(self, transformer_cls):
test_layer = transformer_cls(
num_attention_heads=10,
intermediate_size=2048,
intermediate_activation='relu')
sequence_length = 21
width = 80
# 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, transformer_cls):
test_layer = transformer_cls(
num_attention_heads=10,
intermediate_size=2048,
intermediate_activation='relu')
sequence_length = 21
width = 80
# 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)
def test_layer_invocation_with_mask(self, transformer_cls):
test_layer = transformer_cls(
num_attention_heads=10,
intermediate_size=2048,
intermediate_activation='relu')
sequence_length = 21
width = 80
# 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])
def test_layer_output_range(self, transformer_cls):
test_layer = transformer_cls(
num_attention_heads=10,
intermediate_size=2048,
intermediate_activation='relu')
sequence_length = 21
width = 80
batch_size = 6
input_data = 10 * np.random.random_sample(
(batch_size, sequence_length, width))
mask_data = np.random.randint(
2, size=(batch_size, sequence_length, sequence_length))
output_tensor = test_layer([input_data, mask_data])
# The layer only attends to the first token and outputs the first token
# embeeding.
new_layer = transformer_cls(
num_attention_heads=10,
intermediate_size=2048,
intermediate_activation='relu',
output_range=1)
_ = new_layer([input_data, mask_data])
new_layer.set_weights(test_layer.get_weights())
new_output_tensor = new_layer([input_data, mask_data])
self.assertAllClose(new_output_tensor, output_tensor[:, 0:1, :])
def test_layer_invocation_with_float16_dtype(self, transformer_cls):
tf.keras.mixed_precision.experimental.set_policy('mixed_float16')
test_layer = transformer_cls(
num_attention_heads=10,
intermediate_size=2048,
intermediate_activation='relu')
sequence_length = 21
width = 80
# 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])
def test_transform_with_initializer(self, transformer_cls):
test_layer = transformer_cls(
num_attention_heads=10,
intermediate_size=2048,
intermediate_activation='relu',
kernel_initializer=tf.keras.initializers.TruncatedNormal(stddev=0.02))
sequence_length = 21
width = 80
# 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())
def test_dynamic_layer_sequence(self, transformer_cls):
test_layer = transformer_cls(
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).
width = 30
input_tensor = tf.keras.Input(shape=(None, width))
output_tensor = test_layer(input_tensor)
model = tf.keras.Model(input_tensor, output_tensor)
input_length = 17
input_data = np.ones((1, input_length, width))
output_data = model.predict(input_data)
self.assertAllEqual([1, input_length, width], output_data.shape)
def _create_cache(batch_size, init_decode_length, num_heads, head_size):
return {
'key':
tf.zeros([batch_size, init_decode_length, num_heads, head_size],
dtype=tf.float32),
'value':
tf.zeros([batch_size, init_decode_length, num_heads, head_size],
dtype=tf.float32)
}
@keras_parameterized.run_all_keras_modes
class TransformerDecoderLayerTest(keras_parameterized.TestCase):
def test_decoder_block_with_cache(self):
num_attention_heads = 2
hidden_size = 16
decoder_block = transformer.TransformerDecoderLayer(
num_attention_heads=num_attention_heads,
intermediate_size=32,
intermediate_activation='relu',
dropout_rate=0.1,
attention_dropout_rate=0.1)
# Forward path.
dummy_tensor = tf.zeros([2, 4, 16], dtype=tf.float32)
dummy_mask = tf.zeros([2, 4, 4], dtype=tf.float32)
inputs = [dummy_tensor, dummy_tensor, dummy_mask, dummy_mask]
cache = _create_cache(2, 0, num_attention_heads,
hidden_size // num_attention_heads)
output, cache = decoder_block(inputs, cache)
self.assertEqual(output.shape, (2, 4, hidden_size))
self.assertEqual(cache['value'].shape, (2, 4, 2, 8))
if __name__ == '__main__':
tf.test.main()
|