• 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
16"""Adagrad for TensorFlow."""
17from __future__ import absolute_import
18from __future__ import division
19from __future__ import print_function
20
21from tensorflow.python.framework import ops
22from tensorflow.python.ops import array_ops
23from tensorflow.python.ops import gen_array_ops
24from tensorflow.python.ops import init_ops
25from tensorflow.python.ops import math_ops
26from tensorflow.python.training import optimizer
27from tensorflow.python.training import training_ops
28from tensorflow.python.util.tf_export import tf_export
29
30
31@tf_export(v1=["train.AdagradOptimizer"])
32class AdagradOptimizer(optimizer.Optimizer):
33  """Optimizer that implements the Adagrad algorithm.
34
35  References:
36    Adaptive Subgradient Methods for Online Learning and Stochastic Optimization
37      :[Duchi et al., 2011](http://jmlr.org/papers/v12/duchi11a.html)
38      ([pdf](http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf))
39  """
40
41  def __init__(self, learning_rate, initial_accumulator_value=0.1,
42               use_locking=False, name="Adagrad"):
43    """Construct a new Adagrad optimizer.
44
45    Args:
46      learning_rate: A `Tensor` or a floating point value.  The learning rate.
47      initial_accumulator_value: A floating point value.
48        Starting value for the accumulators, must be positive.
49      use_locking: If `True` use locks for update operations.
50      name: Optional name prefix for the operations created when applying
51        gradients.  Defaults to "Adagrad".
52
53    Raises:
54      ValueError: If the `initial_accumulator_value` is invalid.
55
56    @compatibility(eager)
57    When eager execution is enabled, `learning_rate` can be a callable that
58    takes no arguments and returns the actual value to use. This can be useful
59    for changing these values across different invocations of optimizer
60    functions.
61    @end_compatibility
62    """
63    if initial_accumulator_value <= 0.0:
64      raise ValueError("initial_accumulator_value must be positive: %s" %
65                       initial_accumulator_value)
66    super(AdagradOptimizer, self).__init__(use_locking, name)
67    self._learning_rate = learning_rate
68    self._initial_accumulator_value = initial_accumulator_value
69    # Created in Initialize.
70    self._learning_rate_tensor = None
71
72  def _create_slots(self, var_list):
73    for v in var_list:
74      dtype = v.dtype.base_dtype
75      if v.get_shape().is_fully_defined():
76        init = init_ops.constant_initializer(self._initial_accumulator_value,
77                                             dtype=dtype)
78      else:
79        init = self._init_constant_op(v, dtype)
80      self._get_or_make_slot_with_initializer(v, init, v.get_shape(), dtype,
81                                              "accumulator", self._name)
82
83  def _init_constant_op(self, v, dtype):
84    def init():
85      # Use a Tensor instead of initializer if variable does not have
86      # static shape.
87      init_constant = gen_array_ops.fill(array_ops.shape(v),
88                                         self._initial_accumulator_value)
89      return math_ops.cast(init_constant, dtype)
90    return init
91
92  def _prepare(self):
93    learning_rate = self._call_if_callable(self._learning_rate)
94    self._learning_rate_tensor = ops.convert_to_tensor(
95        learning_rate, name="learning_rate")
96
97  def _apply_dense(self, grad, var):
98    acc = self.get_slot(var, "accumulator")
99    return training_ops.apply_adagrad(
100        var,
101        acc,
102        math_ops.cast(self._learning_rate_tensor, var.dtype.base_dtype),
103        grad,
104        use_locking=self._use_locking)
105
106  def _resource_apply_dense(self, grad, var):
107    acc = self.get_slot(var, "accumulator")
108    return training_ops.resource_apply_adagrad(
109        var.handle,
110        acc.handle,
111        math_ops.cast(self._learning_rate_tensor, grad.dtype.base_dtype),
112        grad,
113        use_locking=self._use_locking)
114
115  def _apply_sparse(self, grad, var):
116    acc = self.get_slot(var, "accumulator")
117    return training_ops.sparse_apply_adagrad(
118        var,
119        acc,
120        math_ops.cast(self._learning_rate_tensor, var.dtype.base_dtype),
121        grad.values,
122        grad.indices,
123        use_locking=self._use_locking)
124
125  def _resource_apply_sparse(self, grad, var, indices):
126    acc = self.get_slot(var, "accumulator")
127    return training_ops.resource_sparse_apply_adagrad(
128        var.handle,
129        acc.handle,
130        math_ops.cast(self._learning_rate_tensor, grad.dtype),
131        grad,
132        indices,
133        use_locking=self._use_locking)
134