From 2114ed7447c8239531aae073336140c8ca861303 Mon Sep 17 00:00:00 2001 From: Yuefeng Zhou Date: Tue, 4 Dec 2018 21:26:16 -0800 Subject: [PATCH] Expose tf.distribute.MirroredStrategy. RELNOTES: tf.distribute.MirroredStrategy is now part of core TensorFlow. It is an updated version of tf.contrib.distribute.MirroredStrategy. It can be used for distributing training to multiple GPUs in Keras, Estimator, or directly. It is an implementation of the tf.distribute.Strategy API, which is also now in core TensorFlow. This API has been updated based on the public design review https://github.com/tensorflow/community/blob/master/rfcs/20181016-replicator.md. Other distribution strategies will be moved to core TF in future releases. PiperOrigin-RevId: 224097441 --- tensorflow/python/BUILD | 1 + tensorflow/python/__init__.py | 1 + tensorflow/python/distribute/BUILD | 12 ++ tensorflow/python/distribute/__init__.py | 25 ++++ .../python/distribute/mirrored_strategy.py | 2 + ...orflow.distribute.-mirrored-strategy.pbtxt | 138 ++++++++++++++++++ .../api/golden/v1/tensorflow.distribute.pbtxt | 4 + ...orflow.distribute.-mirrored-strategy.pbtxt | 138 ++++++++++++++++++ .../api/golden/v2/tensorflow.distribute.pbtxt | 4 + 9 files changed, 325 insertions(+) create mode 100644 tensorflow/python/distribute/__init__.py create mode 100644 tensorflow/tools/api/golden/v1/tensorflow.distribute.-mirrored-strategy.pbtxt create mode 100644 tensorflow/tools/api/golden/v2/tensorflow.distribute.-mirrored-strategy.pbtxt diff --git a/tensorflow/python/BUILD b/tensorflow/python/BUILD index 1384a2ea738..dccbc401432 100644 --- a/tensorflow/python/BUILD +++ b/tensorflow/python/BUILD @@ -145,6 +145,7 @@ py_library( "//tensorflow/lite/python:lite", "//tensorflow/python/compat", "//tensorflow/python/data", + "//tensorflow/python/distribute", "//tensorflow/python/distribute:estimator_training", "//tensorflow/python/eager:def_function", "//tensorflow/python/feature_column:feature_column_py", diff --git a/tensorflow/python/__init__.py b/tensorflow/python/__init__.py index 547043030b1..62a6e9f6d47 100644 --- a/tensorflow/python/__init__.py +++ b/tensorflow/python/__init__.py @@ -78,6 +78,7 @@ from tensorflow.python.ops import initializers_ns as initializers # Bring in subpackages. from tensorflow.python import data +from tensorflow.python import distribute from tensorflow.python import keras from tensorflow.python.feature_column import feature_column_lib as feature_column from tensorflow.python.layers import layers diff --git a/tensorflow/python/distribute/BUILD b/tensorflow/python/distribute/BUILD index 2d9a1764db7..887c61cb8fd 100644 --- a/tensorflow/python/distribute/BUILD +++ b/tensorflow/python/distribute/BUILD @@ -103,6 +103,18 @@ cuda_py_test( ], ) +py_library( + name = "distribute", + srcs = [ + "__init__.py", + ], + srcs_version = "PY2AND3", + deps = [ + ":distribute_lib", + ":mirrored_strategy", + ], +) + py_library( name = "distribute_lib", srcs = [ diff --git a/tensorflow/python/distribute/__init__.py b/tensorflow/python/distribute/__init__.py new file mode 100644 index 00000000000..4ff912ae10d --- /dev/null +++ b/tensorflow/python/distribute/__init__.py @@ -0,0 +1,25 @@ +# Copyright 2017 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. +# ============================================================================== +"""Distribution Strategy library.""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +# pylint: disable=unused-import +from tensorflow.python.distribute import distribute_lib +from tensorflow.python.distribute import distribution_strategy_context +from tensorflow.python.distribute import mirrored_strategy +# pylint: enable=unused-import diff --git a/tensorflow/python/distribute/mirrored_strategy.py b/tensorflow/python/distribute/mirrored_strategy.py index 0775920a7d6..7589720724c 100644 --- a/tensorflow/python/distribute/mirrored_strategy.py +++ b/tensorflow/python/distribute/mirrored_strategy.py @@ -43,6 +43,7 @@ from tensorflow.python.ops import control_flow_ops from tensorflow.python.ops import variable_scope from tensorflow.python.training import coordinator from tensorflow.python.util import nest +from tensorflow.python.util.tf_export import tf_export # TODO(josh11b): Replace asserts in this file with if ...: raise ... @@ -422,6 +423,7 @@ def all_local_devices(num_gpus=None): ("/device:CPU:0",)) +@tf_export("distribute.MirroredStrategy") class MirroredStrategy(distribute_lib.DistributionStrategy): """Mirrors vars to distribute across multiple devices and machines. diff --git a/tensorflow/tools/api/golden/v1/tensorflow.distribute.-mirrored-strategy.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.distribute.-mirrored-strategy.pbtxt new file mode 100644 index 00000000000..a613e2d3d1d --- /dev/null +++ b/tensorflow/tools/api/golden/v1/tensorflow.distribute.-mirrored-strategy.pbtxt @@ -0,0 +1,138 @@ +path: "tensorflow.distribute.MirroredStrategy" +tf_class { + is_instance: "" + is_instance: "" + is_instance: "" + member { + name: "between_graph" + mtype: "" + } + member { + name: "extended" + mtype: "" + } + member { + name: "num_replicas_in_sync" + mtype: "" + } + member { + name: "parameter_devices" + mtype: "" + } + member { + name: "require_static_shapes" + mtype: "" + } + member { + name: "should_checkpoint" + mtype: "" + } + member { + name: "should_init" + mtype: "" + } + member { + name: "should_save_summary" + mtype: "" + } + member { + name: "worker_devices" + mtype: "" + } + member_method { + name: "__init__" + argspec: "args=[\'self\', \'devices\', \'cross_device_ops\'], varargs=None, keywords=None, defaults=[\'None\', \'None\'], " + } + member_method { + name: "batch_reduce" + argspec: "args=[\'self\', \'aggregation\', \'value_destination_pairs\'], varargs=None, keywords=None, defaults=None" + } + member_method { + name: "broadcast" + argspec: "args=[\'self\', \'tensor\', \'destinations\'], varargs=None, keywords=None, defaults=[\'None\'], " + } + member_method { + name: "call_for_each_replica" + argspec: "args=[\'self\', \'fn\'], varargs=args, keywords=kwargs, defaults=None" + } + member_method { + name: "colocate_vars_with" + argspec: "args=[\'self\', \'colocate_with_variable\'], varargs=None, keywords=None, defaults=None" + } + member_method { + name: "configure" + argspec: "args=[\'self\', \'session_config\', \'cluster_spec\', \'task_type\', \'task_id\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\'], " + } + member_method { + name: "distribute_dataset" + argspec: "args=[\'self\', \'dataset_fn\'], varargs=None, keywords=None, defaults=None" + } + member_method { + name: "experimental_finalize" + argspec: "args=[\'self\'], varargs=None, keywords=None, defaults=None" + } + member_method { + name: "experimental_initialize" + argspec: "args=[\'self\'], varargs=None, keywords=None, defaults=None" + } + member_method { + name: "finalize" + argspec: "args=[\'self\'], varargs=None, keywords=None, defaults=None" + } + member_method { + name: "group" + argspec: "args=[\'self\', \'value\', \'name\'], varargs=None, keywords=None, defaults=[\'None\'], " + } + member_method { + name: "initialize" + argspec: "args=[\'self\'], varargs=None, keywords=None, defaults=None" + } + member_method { + name: "make_dataset_iterator" + argspec: "args=[\'self\', \'dataset\'], varargs=None, keywords=None, defaults=None" + } + member_method { + name: "make_input_fn_iterator" + argspec: "args=[\'self\', \'input_fn\', \'replication_mode\'], varargs=None, keywords=None, defaults=[\'InputReplicationMode.PER_WORKER\'], " + } + member_method { + name: "non_slot_devices" + argspec: "args=[\'self\', \'var_list\'], varargs=None, keywords=None, defaults=None" + } + member_method { + name: "read_var" + argspec: "args=[\'self\', \'v\'], varargs=None, keywords=None, defaults=None" + } + member_method { + name: "reduce" + argspec: "args=[\'self\', \'reduce_op\', \'value\'], varargs=None, keywords=None, defaults=None" + } + member_method { + name: "run_steps_on_dataset" + argspec: "args=[\'self\', \'fn\', \'iterator\', \'iterations\', \'initial_loop_values\'], varargs=None, keywords=None, defaults=[\'1\', \'None\'], " + } + member_method { + name: "scope" + argspec: "args=[\'self\'], varargs=None, keywords=None, defaults=None" + } + member_method { + name: "unwrap" + argspec: "args=[\'self\', \'value\'], varargs=None, keywords=None, defaults=None" + } + member_method { + name: "update" + argspec: "args=[\'self\', \'var\', \'fn\'], varargs=args, keywords=kwargs, defaults=None" + } + member_method { + name: "update_config_proto" + argspec: "args=[\'self\', \'config_proto\'], varargs=None, keywords=None, defaults=None" + } + member_method { + name: "update_non_slot" + argspec: "args=[\'self\', \'colocate_with\', \'fn\'], varargs=args, keywords=kwargs, defaults=None" + } + member_method { + name: "value_container" + argspec: "args=[\'self\', \'value\'], varargs=None, keywords=None, defaults=None" + } +} diff --git a/tensorflow/tools/api/golden/v1/tensorflow.distribute.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.distribute.pbtxt index 4d833b54ba0..b0dd73ca1d4 100644 --- a/tensorflow/tools/api/golden/v1/tensorflow.distribute.pbtxt +++ b/tensorflow/tools/api/golden/v1/tensorflow.distribute.pbtxt @@ -8,6 +8,10 @@ tf_module { name: "InputReplicationMode" mtype: "" } + member { + name: "MirroredStrategy" + mtype: "" + } member { name: "ReduceOp" mtype: "" diff --git a/tensorflow/tools/api/golden/v2/tensorflow.distribute.-mirrored-strategy.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.distribute.-mirrored-strategy.pbtxt new file mode 100644 index 00000000000..a613e2d3d1d --- /dev/null +++ b/tensorflow/tools/api/golden/v2/tensorflow.distribute.-mirrored-strategy.pbtxt @@ -0,0 +1,138 @@ +path: "tensorflow.distribute.MirroredStrategy" +tf_class { + is_instance: "" + is_instance: "" + is_instance: "" + member { + name: "between_graph" + mtype: "" + } + member { + name: "extended" + mtype: "" + } + member { + name: "num_replicas_in_sync" + mtype: "" + } + member { + name: "parameter_devices" + mtype: "" + } + member { + name: "require_static_shapes" + mtype: "" + } + member { + name: "should_checkpoint" + mtype: "" + } + member { + name: "should_init" + mtype: "" + } + member { + name: "should_save_summary" + mtype: "" + } + member { + name: "worker_devices" + mtype: "" + } + member_method { + name: "__init__" + argspec: "args=[\'self\', \'devices\', \'cross_device_ops\'], varargs=None, keywords=None, defaults=[\'None\', \'None\'], " + } + member_method { + name: "batch_reduce" + argspec: "args=[\'self\', \'aggregation\', \'value_destination_pairs\'], varargs=None, keywords=None, defaults=None" + } + member_method { + name: "broadcast" + argspec: "args=[\'self\', \'tensor\', \'destinations\'], varargs=None, keywords=None, defaults=[\'None\'], " + } + member_method { + name: "call_for_each_replica" + argspec: "args=[\'self\', \'fn\'], varargs=args, keywords=kwargs, defaults=None" + } + member_method { + name: "colocate_vars_with" + argspec: "args=[\'self\', \'colocate_with_variable\'], varargs=None, keywords=None, defaults=None" + } + member_method { + name: "configure" + argspec: "args=[\'self\', \'session_config\', \'cluster_spec\', \'task_type\', \'task_id\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\'], " + } + member_method { + name: "distribute_dataset" + argspec: "args=[\'self\', \'dataset_fn\'], varargs=None, keywords=None, defaults=None" + } + member_method { + name: "experimental_finalize" + argspec: "args=[\'self\'], varargs=None, keywords=None, defaults=None" + } + member_method { + name: "experimental_initialize" + argspec: "args=[\'self\'], varargs=None, keywords=None, defaults=None" + } + member_method { + name: "finalize" + argspec: "args=[\'self\'], varargs=None, keywords=None, defaults=None" + } + member_method { + name: "group" + argspec: "args=[\'self\', \'value\', \'name\'], varargs=None, keywords=None, defaults=[\'None\'], " + } + member_method { + name: "initialize" + argspec: "args=[\'self\'], varargs=None, keywords=None, defaults=None" + } + member_method { + name: "make_dataset_iterator" + argspec: "args=[\'self\', \'dataset\'], varargs=None, keywords=None, defaults=None" + } + member_method { + name: "make_input_fn_iterator" + argspec: "args=[\'self\', \'input_fn\', \'replication_mode\'], varargs=None, keywords=None, defaults=[\'InputReplicationMode.PER_WORKER\'], " + } + member_method { + name: "non_slot_devices" + argspec: "args=[\'self\', \'var_list\'], varargs=None, keywords=None, defaults=None" + } + member_method { + name: "read_var" + argspec: "args=[\'self\', \'v\'], varargs=None, keywords=None, defaults=None" + } + member_method { + name: "reduce" + argspec: "args=[\'self\', \'reduce_op\', \'value\'], varargs=None, keywords=None, defaults=None" + } + member_method { + name: "run_steps_on_dataset" + argspec: "args=[\'self\', \'fn\', \'iterator\', \'iterations\', \'initial_loop_values\'], varargs=None, keywords=None, defaults=[\'1\', \'None\'], " + } + member_method { + name: "scope" + argspec: "args=[\'self\'], varargs=None, keywords=None, defaults=None" + } + member_method { + name: "unwrap" + argspec: "args=[\'self\', \'value\'], varargs=None, keywords=None, defaults=None" + } + member_method { + name: "update" + argspec: "args=[\'self\', \'var\', \'fn\'], varargs=args, keywords=kwargs, defaults=None" + } + member_method { + name: "update_config_proto" + argspec: "args=[\'self\', \'config_proto\'], varargs=None, keywords=None, defaults=None" + } + member_method { + name: "update_non_slot" + argspec: "args=[\'self\', \'colocate_with\', \'fn\'], varargs=args, keywords=kwargs, defaults=None" + } + member_method { + name: "value_container" + argspec: "args=[\'self\', \'value\'], varargs=None, keywords=None, defaults=None" + } +} diff --git a/tensorflow/tools/api/golden/v2/tensorflow.distribute.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.distribute.pbtxt index 4d833b54ba0..b0dd73ca1d4 100644 --- a/tensorflow/tools/api/golden/v2/tensorflow.distribute.pbtxt +++ b/tensorflow/tools/api/golden/v2/tensorflow.distribute.pbtxt @@ -8,6 +8,10 @@ tf_module { name: "InputReplicationMode" mtype: "" } + member { + name: "MirroredStrategy" + mtype: "" + } member { name: "ReduceOp" mtype: ""