Coverage for /pythoncovmergedfiles/medio/medio/usr/local/lib/python3.8/site-packages/keras/src/optimizers/adamax.py: 27%

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

7# http://www.apache.org/licenses/LICENSE-2.0

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"""Adamax optimizer implementation."""

17import tensorflow.compat.v2 as tf

19from keras.src.optimizers import optimizer

20from keras.src.saving.object_registration import register_keras_serializable

22# isort: off

23from tensorflow.python.util.tf_export import keras_export

26@register_keras_serializable()

27@keras_export(

28 "keras.optimizers.experimental.Adamax", "keras.optimizers.Adamax", v1=[]

29)

30class Adamax(optimizer.Optimizer):

31 """Optimizer that implements the Adamax algorithm.

33 Adamax, a variant of Adam based on the infinity norm, is a first-order

34 gradient-based optimization method. Due to its capability of adjusting the

35 learning rate based on data characteristics, it is suited to learn

36 time-variant process, e.g., speech data with dynamically changed noise

37 conditions. Default parameters follow those provided in the paper (see

38 references below).

40 Initialization:

42 ```python

43 m = 0 # Initialize initial 1st moment vector

44 u = 0 # Initialize the exponentially weighted infinity norm

45 t = 0 # Initialize timestep

46 ```

48 The update rule for parameter `w` with gradient `g` is described at the end

49 of section 7.1 of the paper (see the referenece section):

51 ```python

52 t += 1

53 m = beta1 * m + (1 - beta) * g

54 u = max(beta2 * u, abs(g))

55 current_lr = learning_rate / (1 - beta1 ** t)

56 w = w - current_lr * m / (u + epsilon)

57 ```

59 Args:

60 learning_rate: A `tf.Tensor`, floating point value, a schedule that is a

61 `tf.keras.optimizers.schedules.LearningRateSchedule`, or a callable

62 that takes no arguments and returns the actual value to use. The

63 learning rate. Defaults to `0.001`.

64 beta_1: A float value or a constant float tensor. The exponential decay

65 rate for the 1st moment estimates.

66 beta_2: A float value or a constant float tensor. The exponential decay

67 rate for the exponentially weighted infinity norm.

68 epsilon: A small constant for numerical stability.

69 {{base_optimizer_keyword_args}}

71 Reference:

72 - [Kingma et al., 2014](http://arxiv.org/abs/1412.6980)

73 """

75 def __init__(

76 self,

77 learning_rate=0.001,

78 beta_1=0.9,

79 beta_2=0.999,

80 epsilon=1e-7,

81 weight_decay=None,

82 clipnorm=None,

83 clipvalue=None,

84 global_clipnorm=None,

85 use_ema=False,

86 ema_momentum=0.99,

87 ema_overwrite_frequency=None,

88 jit_compile=True,

89 name="Adamax",

90 **kwargs

91 ):

92 super().__init__(

93 name=name,

94 weight_decay=weight_decay,

95 clipnorm=clipnorm,

96 clipvalue=clipvalue,

97 global_clipnorm=global_clipnorm,

98 use_ema=use_ema,

99 ema_momentum=ema_momentum,

100 ema_overwrite_frequency=ema_overwrite_frequency,

101 jit_compile=jit_compile,

102 **kwargs

103 )

104 self._learning_rate = self._build_learning_rate(learning_rate)

105 self.beta_1 = beta_1

106 self.beta_2 = beta_2

107 self.epsilon = epsilon

108

109 def build(self, var_list):

110 """Initialize optimizer variables.

111

112 Adamax optimizer has 2 types of variables: momentums (denoted as m),

113 exponentially weighted infinity norm (denoted as u).

114

115 Args:

116 var_list: list of model variables to build Adamax variables on.

117 """

118 super().build(var_list)

119 if hasattr(self, "_built") and self._built:

120 return

121 self._built = True

122 self._m = []

123 self._u = []

124 for var in var_list:

125 self._m.append(

126 self.add_variable_from_reference(

127 model_variable=var, variable_name="m"

128 )

129 )

130 self._u.append(

131 self.add_variable_from_reference(

132 model_variable=var, variable_name="u"

133 )

134 )

135

136 def update_step(self, gradient, variable):

137 """Update step given gradient and the associated model variable."""

138 lr = tf.cast(self.learning_rate, variable.dtype)

139 local_step = tf.cast(self.iterations + 1, variable.dtype)

140 beta_1_power = tf.pow(tf.cast(self.beta_1, variable.dtype), local_step)

141

142 var_key = self._var_key(variable)

143 m = self._m[self._index_dict[var_key]]

144 u = self._u[self._index_dict[var_key]]

145

146 if isinstance(gradient, tf.IndexedSlices):

147 # Sparse gradients.

148 indices = gradient.indices

149 m.assign_add(-m * (1 - self.beta_1))

150 m.scatter_add(

151 tf.IndexedSlices(gradient.values * (1 - self.beta_1), indices)

152 )

153 u.assign(u * self.beta_2)

154 u_slice = tf.gather(u, indices)

155 u_slice_incremental = (

156 tf.maximum(u_slice, tf.abs(gradient.values)) - u_slice

157 )

158 u.scatter_add(tf.IndexedSlices(u_slice_incremental, indices))

159 variable.assign_sub(

160 (lr * m) / ((1 - beta_1_power) * (u + self.epsilon))

161 )

162 else:

163 # Dense gradients.

164 m.assign_add((gradient - m) * (1 - self.beta_1))

165 u.assign(tf.maximum(self.beta_2 * u, tf.abs(gradient)))

166 variable.assign_sub(

167 (lr * m) / ((1 - beta_1_power) * (u + self.epsilon))

168 )

169

170 def get_config(self):

171 config = super().get_config()

172

173 config.update(

174 {

175 "learning_rate": self._serialize_hyperparameter(

176 self._learning_rate

177 ),

178 "beta_1": self.beta_1,

179 "beta_2": self.beta_2,

180 "epsilon": self.epsilon,

181 }

182 )

183 return config

184

185

186Adamax.__doc__ = Adamax.__doc__.replace(

187 "{{base_optimizer_keyword_args}}", optimizer.base_optimizer_keyword_args

188)

189