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"""Momentum for TensorFlow.""" 17from tensorflow.python.framework import ops 18from tensorflow.python.ops import math_ops 19from tensorflow.python.training import optimizer 20from tensorflow.python.training import training_ops 21from tensorflow.python.util.tf_export import tf_export 22 23 24@tf_export(v1=["train.MomentumOptimizer"]) 25class MomentumOptimizer(optimizer.Optimizer): 26 """Optimizer that implements the Momentum algorithm. 27 28 Computes (if `use_nesterov = False`): 29 30 ``` 31 accumulation = momentum * accumulation + gradient 32 variable -= learning_rate * accumulation 33 ``` 34 35 Note that in the dense version of this algorithm, `accumulation` is updated 36 and applied regardless of a gradient's value, whereas the sparse version (when 37 the gradient is an `IndexedSlices`, typically because of `tf.gather` or an 38 embedding) only updates variable slices and corresponding `accumulation` terms 39 when that part of the variable was used in the forward pass. 40 41 @compatibility(TF2) 42 tf.compat.v1.train.MomentumOptimizer is compatible with eager mode and 43 `tf.function`. 44 When eager execution is enabled, `learning_rate`,`momentum`, can each be a 45 callable that takes no arguments and returns the actual value to use. This 46 can be useful for changing these values across different invocations of 47 optimizer functions. 48 49 To switch to native TF2 style, please directly use 50 [`tf.keras.optimizers.SGD`] 51 (https://www.tensorflow.org/api_docs/python/tf/keras/optimizers/SGD) 52 with the `momentum` argument. 53 54 #### Structural mapping to native TF2 55 56 Before: 57 58 ```python 59 optimizer = tf.compat.v1.train.MomentumOptimizer( 60 learning_rate=learning_rate, 61 momentum=momentum, 62 use_nesterov=use_nesterov) 63 ``` 64 65 After: 66 67 ```python 68 optimizer = tf.keras.optimizers.SGD( 69 learning_rate=learning_rate, 70 momentum=momentum, 71 nesterov=use_nesterov) 72 ``` 73 74 #### How to map arguments 75 | TF1 Arg Name | TF2 Arg Name | Note | 76 | ------------------ | ------------- | ------------------------------- | 77 | `learning_rate` | `learning_rate`| Be careful of setting | 78 : : : learning_rate tensor value computed from the global step. : 79 : : : In TF1 this was usually meant to imply a dynamic learning rate and : 80 : : : would recompute in each step. In TF2 (eager + function) it will : 81 : : : treat it as a scalar value that only gets computed once instead of : 82 : : : a symbolic placeholder to be computed each time. : 83 | `momentum` | `momentum` | - | 84 | `use_locking` | - | Not applicable in TF2. | 85 | `use_nesterov` | `nesterov` | - | 86 87 #### Before & after usage example 88 Before: 89 90 ```python 91 x = tf.Variable([1,2,3], dtype=tf.float32) 92 grad = tf.constant([0.1, 0.2, 0.3]) 93 optimizer = tf.compat.v1.train.MomentumOptimizer( 94 learning_rate=0.001, 95 momentum=0.9, 96 use_nesterov=False) 97 optimizer.apply_gradients(zip([grad], [x])) 98 ``` 99 100 After: 101 102 ```python 103 x = tf.Variable([1,2,3], dtype=tf.float32) 104 grad = tf.constant([0.1, 0.2, 0.3]) 105 optimizer = tf.keras.optimizers.SGD( 106 learning_rate=0.001, 107 momentum=0.9, 108 nesterov=False) 109 optimizer.apply_gradients(zip([grad], [x])) 110 ``` 111 112 @end_compatibility 113 114 """ 115 116 def __init__(self, learning_rate, momentum, 117 use_locking=False, name="Momentum", use_nesterov=False): 118 """Construct a new Momentum optimizer. 119 120 Args: 121 learning_rate: A `Tensor` or a floating point value. The learning rate. 122 momentum: A `Tensor` or a floating point value. The momentum. 123 use_locking: If `True` use locks for update operations. 124 name: Optional name prefix for the operations created when applying 125 gradients. Defaults to "Momentum". 126 use_nesterov: If `True` use Nesterov Momentum. 127 See (Sutskever et al., 2013). 128 This implementation always computes gradients at the value of the 129 variable(s) passed to the optimizer. Using Nesterov Momentum makes the 130 variable(s) track the values called `theta_t + mu*v_t` in the paper. 131 This implementation is an approximation of the original formula, valid 132 for high values of momentum. It will compute the "adjusted gradient" 133 in NAG by assuming that the new gradient will be estimated by the 134 current average gradient plus the product of momentum and the change 135 in the average gradient. 136 137 References: 138 On the importance of initialization and momentum in deep learning: 139 [Sutskever et al., 2013] 140 (http://proceedings.mlr.press/v28/sutskever13.html) 141 ([pdf](http://proceedings.mlr.press/v28/sutskever13.pdf)) 142 143 144 """ 145 super(MomentumOptimizer, self).__init__(use_locking, name) 146 self._learning_rate = learning_rate 147 self._momentum = momentum 148 self._use_nesterov = use_nesterov 149 150 def _create_slots(self, var_list): 151 for v in var_list: 152 self._zeros_slot(v, "momentum", self._name) 153 154 def _prepare(self): 155 learning_rate = self._learning_rate 156 if callable(learning_rate): 157 learning_rate = learning_rate() 158 self._learning_rate_tensor = ops.convert_to_tensor(learning_rate, 159 name="learning_rate") 160 momentum = self._momentum 161 if callable(momentum): 162 momentum = momentum() 163 self._momentum_tensor = ops.convert_to_tensor(momentum, name="momentum") 164 165 def _apply_dense(self, grad, var): 166 mom = self.get_slot(var, "momentum") 167 return training_ops.apply_momentum( 168 var, mom, 169 math_ops.cast(self._learning_rate_tensor, var.dtype.base_dtype), 170 grad, 171 math_ops.cast(self._momentum_tensor, var.dtype.base_dtype), 172 use_locking=self._use_locking, 173 use_nesterov=self._use_nesterov).op 174 175 def _resource_apply_dense(self, grad, var): 176 mom = self.get_slot(var, "momentum") 177 return training_ops.resource_apply_momentum( 178 var.handle, mom.handle, 179 math_ops.cast(self._learning_rate_tensor, grad.dtype.base_dtype), 180 grad, 181 math_ops.cast(self._momentum_tensor, grad.dtype.base_dtype), 182 use_locking=self._use_locking, 183 use_nesterov=self._use_nesterov) 184 185 def _apply_sparse(self, grad, var): 186 mom = self.get_slot(var, "momentum") 187 return training_ops.sparse_apply_momentum( 188 var, mom, 189 math_ops.cast(self._learning_rate_tensor, var.dtype.base_dtype), 190 grad.values, grad.indices, 191 math_ops.cast(self._momentum_tensor, var.dtype.base_dtype), 192 use_locking=self._use_locking, 193 use_nesterov=self._use_nesterov).op 194 195 def _resource_apply_sparse(self, grad, var, indices): 196 mom = self.get_slot(var, "momentum") 197 return training_ops.resource_sparse_apply_momentum( 198 var.handle, mom.handle, 199 math_ops.cast(self._learning_rate_tensor, grad.dtype), 200 grad, indices, 201 math_ops.cast(self._momentum_tensor, grad.dtype), 202 use_locking=self._use_locking, 203 use_nesterov=self._use_nesterov) 204