• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1# Copyright 2015 The TensorFlow Authors. All Rights Reserved.
2#
3# Licensed under the Apache License, Version 2.0 (the "License");
4# you may not use this file except in compliance with the License.
5# You may obtain a copy of the License at
6#
7#     http://www.apache.org/licenses/LICENSE-2.0
8#
9# Unless required by applicable law or agreed to in writing, software
10# distributed under the License is distributed on an "AS IS" BASIS,
11# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12# See the License for the specific language governing permissions and
13# limitations under the License.
14# ==============================================================================
15"""Layers that operate regularization via the addition of noise.
16"""
17from __future__ import absolute_import
18from __future__ import division
19from __future__ import print_function
20
21import numpy as np
22
23from tensorflow.python.keras import backend as K
24from tensorflow.python.keras.engine.base_layer import Layer
25from tensorflow.python.keras.utils import tf_utils
26from tensorflow.python.ops import array_ops
27from tensorflow.python.ops import math_ops
28from tensorflow.python.util.tf_export import keras_export
29
30
31@keras_export('keras.layers.GaussianNoise')
32class GaussianNoise(Layer):
33  """Apply additive zero-centered Gaussian noise.
34
35  This is useful to mitigate overfitting
36  (you could see it as a form of random data augmentation).
37  Gaussian Noise (GS) is a natural choice as corruption process
38  for real valued inputs.
39
40  As it is a regularization layer, it is only active at training time.
41
42  Arguments:
43    stddev: Float, standard deviation of the noise distribution.
44
45  Call arguments:
46    inputs: Input tensor (of any rank).
47    training: Python boolean indicating whether the layer should behave in
48      training mode (adding noise) or in inference mode (doing nothing).
49
50  Input shape:
51    Arbitrary. Use the keyword argument `input_shape`
52    (tuple of integers, does not include the samples axis)
53    when using this layer as the first layer in a model.
54
55  Output shape:
56    Same shape as input.
57  """
58
59  def __init__(self, stddev, **kwargs):
60    super(GaussianNoise, self).__init__(**kwargs)
61    self.supports_masking = True
62    self.stddev = stddev
63
64  def call(self, inputs, training=None):
65
66    def noised():
67      return inputs + K.random_normal(
68          shape=array_ops.shape(inputs), mean=0., stddev=self.stddev)
69
70    return K.in_train_phase(noised, inputs, training=training)
71
72  def get_config(self):
73    config = {'stddev': self.stddev}
74    base_config = super(GaussianNoise, self).get_config()
75    return dict(list(base_config.items()) + list(config.items()))
76
77  @tf_utils.shape_type_conversion
78  def compute_output_shape(self, input_shape):
79    return input_shape
80
81
82@keras_export('keras.layers.GaussianDropout')
83class GaussianDropout(Layer):
84  """Apply multiplicative 1-centered Gaussian noise.
85
86  As it is a regularization layer, it is only active at training time.
87
88  Arguments:
89    rate: Float, drop probability (as with `Dropout`).
90      The multiplicative noise will have
91      standard deviation `sqrt(rate / (1 - rate))`.
92
93  Call arguments:
94    inputs: Input tensor (of any rank).
95    training: Python boolean indicating whether the layer should behave in
96      training mode (adding dropout) or in inference mode (doing nothing).
97
98  Input shape:
99    Arbitrary. Use the keyword argument `input_shape`
100    (tuple of integers, does not include the samples axis)
101    when using this layer as the first layer in a model.
102
103  Output shape:
104    Same shape as input.
105  """
106
107  def __init__(self, rate, **kwargs):
108    super(GaussianDropout, self).__init__(**kwargs)
109    self.supports_masking = True
110    self.rate = rate
111
112  def call(self, inputs, training=None):
113    if 0 < self.rate < 1:
114
115      def noised():
116        stddev = np.sqrt(self.rate / (1.0 - self.rate))
117        return inputs * K.random_normal(
118            shape=array_ops.shape(inputs), mean=1.0, stddev=stddev)
119
120      return K.in_train_phase(noised, inputs, training=training)
121    return inputs
122
123  def get_config(self):
124    config = {'rate': self.rate}
125    base_config = super(GaussianDropout, self).get_config()
126    return dict(list(base_config.items()) + list(config.items()))
127
128  @tf_utils.shape_type_conversion
129  def compute_output_shape(self, input_shape):
130    return input_shape
131
132
133@keras_export('keras.layers.AlphaDropout')
134class AlphaDropout(Layer):
135  """Applies Alpha Dropout to the input.
136
137  Alpha Dropout is a `Dropout` that keeps mean and variance of inputs
138  to their original values, in order to ensure the self-normalizing property
139  even after this dropout.
140  Alpha Dropout fits well to Scaled Exponential Linear Units
141  by randomly setting activations to the negative saturation value.
142
143  Arguments:
144    rate: float, drop probability (as with `Dropout`).
145      The multiplicative noise will have
146      standard deviation `sqrt(rate / (1 - rate))`.
147    seed: A Python integer to use as random seed.
148
149  Call arguments:
150    inputs: Input tensor (of any rank).
151    training: Python boolean indicating whether the layer should behave in
152      training mode (adding dropout) or in inference mode (doing nothing).
153
154  Input shape:
155    Arbitrary. Use the keyword argument `input_shape`
156    (tuple of integers, does not include the samples axis)
157    when using this layer as the first layer in a model.
158
159  Output shape:
160    Same shape as input.
161  """
162
163  def __init__(self, rate, noise_shape=None, seed=None, **kwargs):
164    super(AlphaDropout, self).__init__(**kwargs)
165    self.rate = rate
166    self.noise_shape = noise_shape
167    self.seed = seed
168    self.supports_masking = True
169
170  def _get_noise_shape(self, inputs):
171    return self.noise_shape if self.noise_shape else array_ops.shape(inputs)
172
173  def call(self, inputs, training=None):
174    if 0. < self.rate < 1.:
175      noise_shape = self._get_noise_shape(inputs)
176
177      def dropped_inputs(inputs=inputs, rate=self.rate, seed=self.seed):  # pylint: disable=missing-docstring
178        alpha = 1.6732632423543772848170429916717
179        scale = 1.0507009873554804934193349852946
180        alpha_p = -alpha * scale
181
182        kept_idx = math_ops.greater_equal(
183            K.random_uniform(noise_shape, seed=seed), rate)
184        kept_idx = math_ops.cast(kept_idx, K.floatx())
185
186        # Get affine transformation params
187        a = ((1 - rate) * (1 + rate * alpha_p**2))**-0.5
188        b = -a * alpha_p * rate
189
190        # Apply mask
191        x = inputs * kept_idx + alpha_p * (1 - kept_idx)
192
193        # Do affine transformation
194        return a * x + b
195
196      return K.in_train_phase(dropped_inputs, inputs, training=training)
197    return inputs
198
199  def get_config(self):
200    config = {'rate': self.rate}
201    base_config = super(AlphaDropout, self).get_config()
202    return dict(list(base_config.items()) + list(config.items()))
203
204  @tf_utils.shape_type_conversion
205  def compute_output_shape(self, input_shape):
206    return input_shape
207