xref: /aosp_15_r20/external/tensorflow/tensorflow/python/tpu/tpu_optimizer.py (revision b6fb3261f9314811a0f4371741dbb8839866f948)
1# Copyright 2017 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"""Optimizer that implements cross-shard gradient reduction for TPU."""
17
18
19from tensorflow.python.framework import ops
20from tensorflow.python.ops.losses import losses
21from tensorflow.python.platform import tf_logging as logging
22from tensorflow.python.tpu import tpu_function
23from tensorflow.python.tpu.ops import tpu_ops
24from tensorflow.python.training import optimizer
25from tensorflow.python.util.tf_export import tf_export
26
27
28@tf_export(v1=["tpu.CrossShardOptimizer"])
29class CrossShardOptimizer(optimizer.Optimizer):
30  """An optimizer that averages gradients across TPU shards."""
31
32  def __init__(self,
33               opt,
34               reduction=losses.Reduction.MEAN,
35               name="CrossShardOptimizer",
36               group_assignment=None):
37    """Construct a new cross-shard optimizer.
38
39    Args:
40      opt: An existing `Optimizer` to encapsulate.
41      reduction: The reduction to apply to the shard losses.
42      name: Optional name prefix for the operations created when applying
43        gradients. Defaults to "CrossShardOptimizer".
44      group_assignment: Optional 2d int32 lists with shape
45        [num_groups, num_replicas_per_group] which describles how to apply
46        optimizer to subgroups.
47
48    Raises:
49      ValueError: If reduction is not a valid cross-shard reduction.
50    """
51    accepted_reductions = (losses.Reduction.SUM, losses.Reduction.MEAN)
52    if reduction not in accepted_reductions:
53      raise ValueError(
54          f"Argument `reduction` should be one of {accepted_reductions}. "
55          f"Received: {reduction}")
56    if not isinstance(opt, optimizer.Optimizer):
57      raise TypeError(
58          "CrossShardOptimizer only works with tf.training.Optimizer and not "
59          f"Keras Optimizer. Received: {opt}. "
60          "If you are using TPUStrategy, "
61          "Keras Optimizer will sum gradients across replicas."
62          "If you are using TPUEstimator, you may instead sum your gradients "
63          "with:\n"
64          "`grads = [tf.compat.v1.tpu.cross_replica_sum(g) for g in grads]`\n"
65          "If you want to average your gradients, rescale your loss with: "
66          "`loss /= global_batch_size`")
67
68    super(CrossShardOptimizer, self).__init__(False, name)
69    self._opt = opt
70    self._reduction = reduction
71    self._group_assignment = group_assignment
72
73  def _verify_and_get_subgroup_size(self, group_assignment, num_shards):
74    """Verify group_assignment and get the subgroup size".
75
76    Args:
77      group_assignment: list of group ids for applying the optimizer
78        to subgroups.
79      num_shards: The number of TPU shards.
80
81    Returns:
82      The size of one subgroup in group_assignment.
83
84    Raises:
85      ValueError: If group_assignment is invalid.
86    """
87    if not group_assignment:
88      return None
89    if not (isinstance(group_assignment, list) and
90            all(isinstance(i, list) for i in group_assignment)):
91      raise ValueError(
92          f"Argument `group_assignment` must be a list of lists. "
93          f"Received: {group_assignment}")
94
95    replica_ids = set()
96    for g in group_assignment:
97      for i in g:
98        replica_ids.add(i)
99
100    if set(range(num_shards)) != replica_ids:
101      raise ValueError(
102          f"Argument `group_assignment` must be a permutation of "
103          f"range({num_shards}). Received: {group_assignment}")
104
105    subgroup_size_list = [len(group) for group in group_assignment]
106    if all(subgroup_size_list[0] == size for size in subgroup_size_list):
107      return subgroup_size_list[0]
108    else:
109      raise ValueError("The size of each subgroup in `group_assignment` must "
110                       f"be equal. Received: {group_assignment}")
111
112  def compute_gradients(self, loss, var_list=None, **kwargs):
113    """Compute gradients of "loss" for the variables in "var_list".
114
115    This simply wraps `compute_gradients()` from the real optimizer. The
116    gradients will be aggregated in `apply_gradients()` so that user can
117    modify the gradients like clipping with per replica global norm if needed.
118    The global norm with aggregated gradients can be bad as one replica's huge
119    gradients can hurt the gradients from other replicas.
120
121    When the CrossShardOptimizer is constructed with
122    `reduction == losses.Reduction.MEAN` (default), this function scales the
123    loss by `1.0 / num_shards` before computing the gradients. Assuming the
124    optimizer uses the default implementation of `compute_gradients()`, the
125    gradients of the scaled loss are scaled by `1.0 / num_shards` compared to
126    the gradients of the original loss. This scaling factor is important because
127    `apply_gradients()` sums gradients across shards, rather than averaging
128    them. However, the scaling factor must be taken into account when clipping
129    the norm of the gradients or performing other postprocessing.
130
131    Args:
132      loss: A Tensor containing the value to minimize.
133      var_list: Optional list or tuple of `tf.Variable` to update to minimize
134        `loss`.  Defaults to the list of variables collected in the graph
135        under the key `GraphKey.TRAINABLE_VARIABLES`.
136      **kwargs: Keyword arguments for compute_gradients().
137
138    Returns:
139      A list of (gradient, variable) pairs.
140
141    Raises:
142      ValueError: If not within a tpu_shard_context or group_assignment is
143        invalid.
144    """
145    num_shards = tpu_function.get_tpu_context().number_of_shards
146    if num_shards is None:
147      logging.warning(
148          "CrossShardOptimizer should be used within a tpu_shard_context, but "
149          "got unset number_of_shards. Assuming 1.")
150      num_shards = 1
151
152    subgroup_size = self._verify_and_get_subgroup_size(self._group_assignment,
153                                                       num_shards)
154
155    if num_shards > 1 and self._reduction == losses.Reduction.MEAN:
156      if self._group_assignment:
157        scale = 1.0 / subgroup_size
158      else:
159        scale = 1.0 / num_shards
160      loss *= scale
161
162    return self._opt.compute_gradients(loss, var_list=var_list, **kwargs)
163
164  def apply_gradients(self, grads_and_vars, global_step=None, name=None):
165    """Apply gradients to variables.
166
167    Calls tpu_ops.cross_replica_sum() to sum gradient contributions across
168    replicas, and then applies the real optimizer.
169
170    Args:
171      grads_and_vars: List of (gradient, variable) pairs as returned by
172        compute_gradients().
173      global_step: Optional Variable to increment by one after the
174        variables have been updated.
175      name: Optional name for the returned operation.  Default to the
176        name passed to the Optimizer constructor.
177
178    Returns:
179      An `Operation` that applies the gradients. If `global_step` was not None,
180      that operation also increments `global_step`.
181
182    Raises:
183      ValueError: If the grads_and_vars is malformed.
184    """
185    summed_grads_and_vars = []
186    for (grad, var) in grads_and_vars:
187      if grad is None:
188        summed_grads_and_vars.append((grad, var))
189      else:
190        with ops.colocate_with(grad):
191          summed_grads_and_vars.append((tpu_ops.cross_replica_sum(
192              grad, self._group_assignment), var))
193    return self._opt.apply_gradients(summed_grads_and_vars, global_step, name)
194
195  def get_slot(self, *args, **kwargs):
196    """Return a slot named "name" created for "var" by the Optimizer.
197
198    This simply wraps the get_slot() from the actual optimizer.
199
200    Args:
201      *args: Arguments for get_slot().
202      **kwargs: Keyword arguments for get_slot().
203
204    Returns:
205      The `Variable` for the slot if it was created, `None` otherwise.
206    """
207    return self._opt.get_slot(*args, **kwargs)
208
209  def get_slot_names(self, *args, **kwargs):
210    """Return a list of the names of slots created by the `Optimizer`.
211
212    This simply wraps the get_slot_names() from the actual optimizer.
213
214    Args:
215      *args: Arguments for get_slot().
216      **kwargs: Keyword arguments for get_slot().
217
218    Returns:
219      A list of strings.
220    """
221    return self._opt.get_slot_names(*args, **kwargs)
222
223  def variables(self):
224    """Forwarding the variables from the underlying optimizer."""
225    return self._opt.variables()
226