Coverage for /pythoncovmergedfiles/medio/medio/usr/local/lib/python3.8/site-packages/keras/src/optimizers/ftrl.py: 23%
57 statements
« prev ^ index » next coverage.py v7.4.0, created at 2024-01-03 07:57 +0000
1# Copyright 2022 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"""FTRL 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.Ftrl", "keras.optimizers.Ftrl", v1=[]
29)
30class Ftrl(optimizer.Optimizer):
31 r"""Optimizer that implements the FTRL algorithm.
33 "Follow The Regularized Leader" (FTRL) is an optimization algorithm
34 developed at Google for click-through rate prediction in the early 2010s. It
35 is most suitable for shallow models with large and sparse feature spaces.
36 The algorithm is described by
37 [McMahan et al., 2013](https://research.google.com/pubs/archive/41159.pdf).
38 The Keras version has support for both online L2 regularization
39 (the L2 regularization described in the paper
40 above) and shrinkage-type L2 regularization
41 (which is the addition of an L2 penalty to the loss function).
43 Initialization:
45 ```python
46 n = 0
47 sigma = 0
48 z = 0
49 ```
51 Update rule for one variable `w`:
53 ```python
54 prev_n = n
55 n = n + g ** 2
56 sigma = (n ** -lr_power - prev_n ** -lr_power) / lr
57 z = z + g - sigma * w
58 if abs(z) < lambda_1:
59 w = 0
60 else:
61 w = (sgn(z) * lambda_1 - z) / ((beta + sqrt(n)) / alpha + lambda_2)
62 ```
64 Notation:
66 - `lr` is the learning rate
67 - `g` is the gradient for the variable
68 - `lambda_1` is the L1 regularization strength
69 - `lambda_2` is the L2 regularization strength
70 - `lr_power` is the power to scale n.
72 Check the documentation for the `l2_shrinkage_regularization_strength`
73 parameter for more details when shrinkage is enabled, in which case gradient
74 is replaced with a gradient with shrinkage.
76 Args:
77 learning_rate: A `Tensor`, floating point value, a schedule that is a
78 `tf.keras.optimizers.schedules.LearningRateSchedule`, or a callable that
79 takes no arguments and returns the actual value to use. The learning
80 rate. Defaults to `0.001`.
81 learning_rate_power: A float value, must be less or equal to zero.
82 Controls how the learning rate decreases during training. Use zero for a
83 fixed learning rate.
84 initial_accumulator_value: The starting value for accumulators. Only zero
85 or positive values are allowed.
86 l1_regularization_strength: A float value, must be greater than or equal
87 to zero. Defaults to `0.0`.
88 l2_regularization_strength: A float value, must be greater than or equal
89 to zero. Defaults to `0.0`.
90 l2_shrinkage_regularization_strength: A float value, must be greater than
91 or equal to zero. This differs from L2 above in that the L2 above is a
92 stabilization penalty, whereas this L2 shrinkage is a magnitude penalty.
93 When input is sparse shrinkage will only happen on the active weights.
94 beta: A float value, representing the beta value from the paper. Defaults
95 to 0.0.
96 {{base_optimizer_keyword_args}}
97 """
99 def __init__(
100 self,
101 learning_rate=0.001,
102 learning_rate_power=-0.5,
103 initial_accumulator_value=0.1,
104 l1_regularization_strength=0.0,
105 l2_regularization_strength=0.0,
106 l2_shrinkage_regularization_strength=0.0,
107 beta=0.0,
108 weight_decay=None,
109 clipnorm=None,
110 clipvalue=None,
111 global_clipnorm=None,
112 use_ema=False,
113 ema_momentum=0.99,
114 ema_overwrite_frequency=None,
115 jit_compile=True,
116 name="Ftrl",
117 **kwargs,
118 ):
119 super().__init__(
120 name=name,
121 weight_decay=weight_decay,
122 clipnorm=clipnorm,
123 clipvalue=clipvalue,
124 global_clipnorm=global_clipnorm,
125 use_ema=use_ema,
126 ema_momentum=ema_momentum,
127 ema_overwrite_frequency=ema_overwrite_frequency,
128 jit_compile=jit_compile,
129 **kwargs,
130 )
132 if initial_accumulator_value < 0.0:
133 raise ValueError(
134 "`initial_accumulator_value` needs to be positive or zero. "
135 "Received: initial_accumulator_value="
136 f"{initial_accumulator_value}."
137 )
138 if learning_rate_power > 0.0:
139 raise ValueError(
140 "`learning_rate_power` needs to be negative or zero. Received: "
141 f"learning_rate_power={learning_rate_power}."
142 )
143 if l1_regularization_strength < 0.0:
144 raise ValueError(
145 "`l1_regularization_strength` needs to be positive or zero. "
146 "Received: l1_regularization_strength="
147 f"{l1_regularization_strength}."
148 )
149 if l2_regularization_strength < 0.0:
150 raise ValueError(
151 "`l2_regularization_strength` needs to be positive or zero. "
152 "Received: l2_regularization_strength="
153 f"{l2_regularization_strength}."
154 )
155 if l2_shrinkage_regularization_strength < 0.0:
156 raise ValueError(
157 "`l2_shrinkage_regularization_strength` needs to be positive "
158 "or zero. Received: l2_shrinkage_regularization_strength"
159 f"={l2_shrinkage_regularization_strength}."
160 )
162 self._learning_rate = self._build_learning_rate(learning_rate)
163 self.learning_rate_power = learning_rate_power
164 self.initial_accumulator_value = initial_accumulator_value
165 self.l1_regularization_strength = l1_regularization_strength
166 self.l2_regularization_strength = l2_regularization_strength
167 self.l2_shrinkage_regularization_strength = (
168 l2_shrinkage_regularization_strength
169 )
170 self.beta = beta
172 def build(self, var_list):
173 """Initialize optimizer variables.
175 Args:
176 var_list: list of model variables to build Ftrl variables on.
177 """
178 super().build(var_list)
179 if hasattr(self, "_built") and self._built:
180 return
181 self._accumulators = []
182 self._linears = []
183 for var in var_list:
184 self._accumulators.append(
185 self.add_variable_from_reference(
186 model_variable=var,
187 variable_name="accumulator",
188 initial_value=tf.cast(
189 tf.fill(
190 dims=var.shape, value=self.initial_accumulator_value
191 ),
192 dtype=var.dtype,
193 ),
194 )
195 )
196 self._linears.append(
197 self.add_variable_from_reference(
198 model_variable=var, variable_name="linear"
199 )
200 )
201 self._built = True
203 def update_step(self, gradient, variable):
204 """Update step given gradient and the associated model variable."""
206 lr = tf.cast(self.learning_rate, variable.dtype)
207 var_key = self._var_key(variable)
208 accum = self._accumulators[self._index_dict[var_key]]
209 linear = self._linears[self._index_dict[var_key]]
211 lr_power = self.learning_rate_power
212 l2_reg = self.l2_regularization_strength
213 l2_reg = l2_reg + self.beta / (2.0 * lr)
215 # Ftrl optimizer has the same implementation for sparse and dense
216 # gradients update.
217 grad_to_use = (
218 gradient + 2 * self.l2_shrinkage_regularization_strength * variable
219 )
220 new_accum = accum + tf.pow(gradient, 2)
221 linear.assign_add(
222 grad_to_use
223 - (tf.pow(new_accum, -lr_power) - tf.pow(accum, -lr_power))
224 / lr
225 * variable
226 )
227 quadratic = tf.pow(new_accum, (-lr_power)) / lr + 2 * l2_reg
228 linear_clipped = tf.clip_by_value(
229 linear,
230 -self.l1_regularization_strength,
231 self.l1_regularization_strength,
232 )
233 variable.assign((linear_clipped - linear) / quadratic)
234 accum.assign(new_accum)
236 def get_config(self):
237 config = super().get_config()
239 config.update(
240 {
241 "learning_rate": self._serialize_hyperparameter(
242 self._learning_rate
243 ),
244 "learning_rate_power": self.learning_rate_power,
245 "initial_accumulator_value": self.initial_accumulator_value,
246 "l1_regularization_strength": self.l1_regularization_strength,
247 "l2_regularization_strength": self.l2_regularization_strength,
248 "l2_shrinkage_regularization_strength": self.l2_shrinkage_regularization_strength, # noqa: E501
249 "beta": self.beta,
250 }
251 )
252 return config
255Ftrl.__doc__ = Ftrl.__doc__.replace(
256 "{{base_optimizer_keyword_args}}", optimizer.base_optimizer_keyword_args
257)