Home | History | Annotate | Download | only in training 
      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."""
     17 from __future__ import absolute_import
     18 from __future__ import division
     19 from __future__ import print_function
     20 
     21 from tensorflow.python.framework import ops
     22 from tensorflow.python.ops import array_ops
     23 from tensorflow.python.ops import gen_array_ops
     24 from tensorflow.python.ops import init_ops
     25 from tensorflow.python.ops import math_ops
     26 from tensorflow.python.training import optimizer
     27 from tensorflow.python.training import training_ops
     28 from tensorflow.python.util.tf_export import tf_export
     29 
     30 
     31 @tf_export("train.AdagradOptimizer")
     32 class AdagradOptimizer(optimizer.Optimizer):
     33   """Optimizer that implements the Adagrad algorithm.
     34 
     35   See this [paper](http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf)
     36   or this
     37   [intro](http://cs.stanford.edu/~ppasupat/a9online/uploads/proximal_notes.pdf).
     38   """
     39 
     40   def __init__(self, learning_rate, initial_accumulator_value=0.1,
     41                use_locking=False, name="Adagrad"):
     42     """Construct a new Adagrad optimizer.
     43 
     44     Args:
     45       learning_rate: A `Tensor` or a floating point value.  The learning rate.
     46       initial_accumulator_value: A floating point value.
     47         Starting value for the accumulators, must be positive.
     48       use_locking: If `True` use locks for update operations.
     49       name: Optional name prefix for the operations created when applying
     50         gradients.  Defaults to "Adagrad".
     51 
     52     Raises:
     53       ValueError: If the `initial_accumulator_value` is invalid.
     54     """
     55     if initial_accumulator_value <= 0.0:
     56       raise ValueError("initial_accumulator_value must be positive: %s" %
     57                        initial_accumulator_value)
     58     super(AdagradOptimizer, self).__init__(use_locking, name)
     59     self._learning_rate = learning_rate
     60     self._initial_accumulator_value = initial_accumulator_value
     61     # Created in Initialize.
     62     self._learning_rate_tensor = None
     63 
     64   def _create_slots(self, var_list):
     65     for v in var_list:
     66       with ops.colocate_with(v):
     67         dtype = v.dtype.base_dtype
     68         if v.get_shape().is_fully_defined():
     69           init = init_ops.constant_initializer(self._initial_accumulator_value,
     70                                                dtype=dtype)
     71         else:
     72           # Use a Tensor instead of initializer if variable does not have static
     73           # shape.
     74           init_constant = gen_array_ops.fill(array_ops.shape(v),
     75                                              self._initial_accumulator_value)
     76           init = math_ops.cast(init_constant, dtype)
     77       self._get_or_make_slot_with_initializer(v, init, v.get_shape(), dtype,
     78                                               "accumulator", self._name)
     79 
     80   def _prepare(self):
     81     self._learning_rate_tensor = ops.convert_to_tensor(self._learning_rate,
     82                                                        name="learning_rate")
     83 
     84   def _apply_dense(self, grad, var):
     85     acc = self.get_slot(var, "accumulator")
     86     return training_ops.apply_adagrad(
     87         var,
     88         acc,
     89         math_ops.cast(self._learning_rate_tensor, var.dtype.base_dtype),
     90         grad,
     91         use_locking=self._use_locking)
     92 
     93   def _resource_apply_dense(self, grad, var):
     94     acc = self.get_slot(var, "accumulator")
     95     return training_ops.resource_apply_adagrad(
     96         var.handle,
     97         acc.handle,
     98         math_ops.cast(self._learning_rate_tensor, grad.dtype.base_dtype),
     99         grad,
    100         use_locking=self._use_locking)
    101 
    102   def _apply_sparse(self, grad, var):
    103     acc = self.get_slot(var, "accumulator")
    104     return training_ops.sparse_apply_adagrad(
    105         var,
    106         acc,
    107         math_ops.cast(self._learning_rate_tensor, var.dtype.base_dtype),
    108         grad.values,
    109         grad.indices,
    110         use_locking=self._use_locking)
    111 
    112   def _resource_apply_sparse(self, grad, var, indices):
    113     acc = self.get_slot(var, "accumulator")
    114     return training_ops.resource_sparse_apply_adagrad(
    115         var.handle,
    116         acc.handle,
    117         math_ops.cast(self._learning_rate_tensor, grad.dtype),
    118         grad,
    119         indices,
    120         use_locking=self._use_locking)
    121