Internal change

PiperOrigin-RevId: 418724903

Author: saberkun

official/modeling/tf_utils.py

@@ -201,3 +201,74 @@ def safe_mean(losses):
   total = tf.reduce_sum(losses)
   num_elements = tf.cast(tf.size(losses), dtype=losses.dtype)
   return tf.math.divide_no_nan(total, num_elements)
+
+
+def get_replica_id():
+  """Gets replica id depending on the environment."""
+  context = tf.distribute.get_replica_context()
+  if context is not None:
+    return context.replica_id_in_sync_group
+  else:
+    raise RuntimeError("Unknown replica context. The `get_replica_id` method "
+                       "relies on TF 2.x tf.distribute API.")
+
+
+def cross_replica_concat(value, axis, name="cross_replica_concat"):
+  """Concatenates the given `value` across (GPU/TPU) cores, along `axis`.
+
+  In general, each core ("replica") will pass a
+  replica-specific value as `value` (corresponding to some element of a
+  data-parallel computation taking place across replicas).
+
+  The resulting concatenated `Tensor` will have the same shape as `value` for
+  all dimensions except `axis`, where it will be larger by a factor of the
+  number of replicas. It will also have the same `dtype` as `value`.
+
+  The position of a given replica's `value` within the resulting concatenation
+  is determined by that replica's replica ID. For
+  example:
+
+  With `value` for replica 0 given as
+
+      0 0 0
+      0 0 0
+
+  and `value` for replica 1 given as
+
+      1 1 1
+      1 1 1
+
+  the resulting concatenation along axis 0 will be
+
+      0 0 0
+      0 0 0
+      1 1 1
+      1 1 1
+
+  and this result will be identical across all replicas.
+
+  Note that this API only works in TF2 with `tf.distribute`.
+
+  Args:
+    value: The `Tensor` to concatenate across replicas. Each replica will have a
+      different value for this `Tensor`, and these replica-specific values will
+      be concatenated.
+    axis: The axis along which to perform the concatenation as a Python integer
+      (not a `Tensor`). E.g., `axis=0` to concatenate along the batch dimension.
+    name: A name for the operation (used to create a name scope).
+
+  Returns:
+    The result of concatenating `value` along `axis` across replicas.
+
+  Raises:
+    RuntimeError: when the batch (0-th) dimension is None.
+  """
+  with tf.name_scope(name):
+    context = tf.distribute.get_replica_context()
+    # Typically this could be hit only if the tensor is derived from a
+    # dataset with finite epochs and drop_remainder=False, where the last
+    # batch could of different batch size and then the dim-0 is of dynamic
+    # shape.
+    if value.shape.as_list()[0] is None:
+      raise RuntimeError(f"{value} has unknown batch.")
+    return context.all_gather(value, axis=axis)

official/modeling/tf_utils_test.py

@@ -0,0 +1,89 @@
+# Copyright 2021 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Tests for tf_utils."""
+from absl.testing import parameterized
+import numpy as np
+import tensorflow as tf
+
+from tensorflow.python.distribute import combinations
+from tensorflow.python.distribute import strategy_combinations
+from official.modeling import tf_utils
+
+
+def all_strategy_combinations():
+  return combinations.combine(
+      strategy=[
+          strategy_combinations.cloud_tpu_strategy,
+          strategy_combinations.mirrored_strategy_with_two_gpus,
+      ],
+      mode='eager',
+  )
+
+
+class TFUtilsTest(tf.test.TestCase, parameterized.TestCase):
+
+  @combinations.generate(all_strategy_combinations())
+  def test_cross_replica_concat(self, strategy):
+    num_cores = strategy.num_replicas_in_sync
+
+    shape = (2, 3, 4)
+
+    def concat(axis):
+
+      @tf.function
+      def function():
+        replica_value = tf.fill(shape, tf_utils.get_replica_id())
+        return tf_utils.cross_replica_concat(replica_value, axis=axis)
+
+      return function
+
+    def expected(axis):
+      values = [np.full(shape, i) for i in range(num_cores)]
+      return np.concatenate(values, axis=axis)
+
+    per_replica_results = strategy.run(concat(axis=0))
+    replica_0_result = per_replica_results.values[0].numpy()
+    for value in per_replica_results.values[1:]:
+      self.assertAllClose(value.numpy(), replica_0_result)
+    self.assertAllClose(replica_0_result, expected(axis=0))
+
+    replica_0_result = strategy.run(concat(axis=1)).values[0].numpy()
+    self.assertAllClose(replica_0_result, expected(axis=1))
+
+    replica_0_result = strategy.run(concat(axis=2)).values[0].numpy()
+    self.assertAllClose(replica_0_result, expected(axis=2))
+
+  @combinations.generate(all_strategy_combinations())
+  def test_cross_replica_concat_gradient(self, strategy):
+    num_cores = strategy.num_replicas_in_sync
+
+    shape = (10, 5)
+
+    @tf.function
+    def function():
+      replica_value = tf.random.normal(shape)
+      with tf.GradientTape() as tape:
+        tape.watch(replica_value)
+        concat_value = tf_utils.cross_replica_concat(replica_value, axis=0)
+        output = tf.reduce_sum(concat_value)
+      return tape.gradient(output, replica_value)
+
+    per_replica_gradients = strategy.run(function)
+    for gradient in per_replica_gradients.values:
+      self.assertAllClose(gradient, num_cores * tf.ones(shape))
+
+
+if __name__ == '__main__':
+  tf.test.main()