add gradient Merge optimizer to meta (#25763)

* add gradient Merge optimizer to meta, test=develop

python/paddle/fleet/base/distributed_strategy.py

@@ -521,6 +521,23 @@ class DistributedStrategy(object):
 
     @property
     def gradient_merge(self):
+        """
+        Gradient Merge, also called as Gradient Accumulation,
+        is a strategy for large batch training. With this strategy,
+        model parameter will not be updated until user-defined steps.
+        For each step, the forward network and the backward network
+        will run to calculate the gradient of model parameters.
+        For every k step, the optimization network will run,
+        applying a specific optimization method (such as SGD, Adam)
+        to model parameters.
+
+        Examples:
+        .. code-block:: python
+            import paddle.fleet as fleet
+            strategy = fleet.DistributedStrategy()
+            strategy.gradient_merge = True
+            strategy.gradient_merge_configs = {"k_steps": 4, "avg": True}
+        """
         return self.strategy.gradient_merge
 
     @gradient_merge.setter
@@ -532,6 +549,18 @@ class DistributedStrategy(object):
 
     @property
     def gradient_merge_configs(self):
+        """
+        the key-value configs of distribute_strategy
+        Keys: 
+            k_steps (int): the update period of the parameters
+            avg (bool): whether to average the gradients of each mini-batch,
+                the default value is `True`
+        Example:
+            import paddle.fleet as fleet
+            strategy = fleet.DistributedStrategy()
+            strategy.gradient_merge = True
+            strategy.gradient_merge_configs = {"k_steps": 4, "avg": True}
+        """
         return get_msg_dict(self.strategy.gradient_merge_configs)
 
     @gradient_merge_configs.setter

python/paddle/fleet/base/meta_optimizer_factory.py

@@ -13,11 +13,16 @@
 # limitations under the License.
 
 from ..meta_optimizers import RecomputeOptimizer
+from ..meta_optimizers import GradientMergeOptimizer
 from ..meta_optimizers import GraphExecutionOptimizer
 
 __all__ = ["MetaOptimizerFactory"]
 
-meta_optimizer_names = ["RecomputeOptimizer", "GraphExecutionOptimizer"]
+meta_optimizer_names = [
+    "RecomputeOptimizer",
+    "GradientMergeOptimizer",
+    "GraphExecutionOptimizer",
+]
 
 
 class MetaOptimizerFactory(object):

python/paddle/fleet/meta_optimizers/__init__.py

@@ -12,6 +12,10 @@
 # See the License for the specific language governing permissions and
 
 from .recompute_optimizer import RecomputeOptimizer
+from .gradient_merge_optimizer import GradientMergeOptimizer
 from .graph_execution_optimizer import GraphExecutionOptimizer
 
-__all__ = ['RecomputeOptimizer']
+__all__ = [
+    'RecomputeOptimizer',
+    'GradientMergeOptimizer',
+]

python/paddle/fleet/meta_optimizers/gradient_merge_optimizer.py

+#   Copyright (c) 2019 PaddlePaddle 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
+
+from paddle.fluid.optimizer import GradientMergeOptimizer as GM
+from .meta_optimizer_base import MetaOptimizerBase
+
+__all__ = ["GradientMergeOptimizer"]
+
+
+class GradientMergeOptimizer(MetaOptimizerBase):
+    def __init__(self, optimizer):
+        super(GradientMergeOptimizer, self).__init__(optimizer)
+        self.inner_opt = optimizer
+        self.wrapped_opt = GM(optimizer)
+        self.meta_optimizers_white_list = []
+
+    def _set_basic_info(self, loss, role_maker, user_defined_optimizer,
+                        user_defined_strategy):
+        super(GradientMergeOptimizer, self)._set_basic_info(
+            loss, role_maker, user_defined_optimizer, user_defined_strategy)
+        self.wrapped_opt._set_k_steps(
+            self.user_defined_strategy.gradient_merge_configs["k_steps"])
+        self.wrapped_opt._set_avg(
+            self.user_defined_strategy.gradient_merge_configs["avg"])
+
+    def _can_apply(self):
+        can_apply = (self.user_defined_strategy.gradient_merge == True) and \
+                  self.user_defined_strategy.gradient_merge_configs["k_steps"] > 1
+        return can_apply
+
+    def _disable_strategy(self, dist_strategy):
+        dist_strategy.gradient_merge = False
+        dist_strategy.gradient_merge_configs = {"k_steps": 1, "avg": True}
+
+    def minimize_impl(self,
+                      loss,
+                      startup_program=None,
+                      parameter_list=None,
+                      no_grad_set=None):
+        optimize_ops, params_grads = \
+            self.wrapped_opt.minimize(loss, startup_program,
+                                      parameter_list, no_grad_set)
+        return optimize_ops, params_grads

python/paddle/fluid/optimizer.py

@@ -5017,6 +5017,12 @@ class GradientMergeOptimizer(object):
         self.type = "gradient_merge"
         self.avg = avg
 
+    def _set_k_steps(self, k_steps):
+        self.k_steps = k_steps
+
+    def _set_avg(self, avg):
+        self.avg = avg
+
     def minimize(self,
                  loss,
                  startup_program=None,

python/paddle/fluid/tests/unittests/CMakeLists.txt

@@ -32,6 +32,7 @@ list(APPEND MIXED_DIST_TEST_OPS test_fleet_checkpoint)
 list(APPEND MIXED_DIST_TEST_OPS test_collective_optimizer)
 list(APPEND MIXED_DIST_TEST_OPS test_fleet_base)
 list(APPEND MIXED_DIST_TEST_OPS test_fleet_meta_optimizer)
+list(APPEND MIXED_DIST_TEST_OPS test_fleet_gradient_merge_meta_optimizer)
 list(APPEND MIXED_DIST_TEST_OPS test_fleet_private_function)
 foreach(TEST_OP ${MIXED_DIST_TEST_OPS})
   list(REMOVE_ITEM TEST_OPS ${TEST_OP})
@@ -364,6 +365,7 @@ if(WITH_DISTRIBUTE)
     if(NOT APPLE)
     	   py_test_modules(test_fleet_base MODULES test_fleet_base ENVS ${dist_ENVS})
     	   py_test_modules(test_fleet_meta_optimizer MODULES test_fleet_meta_optimizer ENVS ${dist_ENVS})
+           py_test_modules(test_fleet_gradient_merge_meta_optimizer MODULES test_fleet_gradient_merge_meta_optimizer ENVS ${dist_ENVS})
     	   py_test_modules(test_fleet_private_function MODULES test_fleet_private_function ENVS ${dist_ENVS})
     endif(NOT APPLE)
     if(WITH_DGC)

python/paddle/fluid/tests/unittests/test_fleet_gradient_merge_meta_optimizer.py

+# Copyright (c) 2020 PaddlePaddle 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.
+
+import unittest
+import paddle
+import os
+import paddle.fleet as fleet
+import paddle.fluid.incubate.fleet.base.role_maker as role_maker
+
+
+class TestFleetGradientMergeMetaOptimizer(unittest.TestCase):
+    def setUp(self):
+        os.environ["POD_IP"] = "127.0.0.1"
+        os.environ["PADDLE_TRAINER_ENDPOINTS"] = "127.0.0.1:36001"
+        os.environ["PADDLE_TRAINERS_NUM"] = "2"
+        os.environ["PADDLE_PSERVERS_IP_PORT_LIST"] = \
+                       "127.0.0.1:36001,127.0.0.2:36001"
+
+    def test_gradient_merge_optimizer(self):
+        role = role_maker.PaddleCloudRoleMaker(is_collective=True)
+        fleet.init(role)
+        input_x = paddle.fluid.layers.data(
+            name="x", shape=[32], dtype='float32')
+        input_y = paddle.fluid.layers.data(name="y", shape=[1], dtype='int64')
+
+        fc_1 = paddle.fluid.layers.fc(input=input_x, size=64, act='tanh')
+        fc_2 = paddle.fluid.layers.fc(input=fc_1, size=64, act='tanh')
+        prediction = paddle.fluid.layers.fc(input=[fc_2], size=2, act='softmax')
+        cost = paddle.fluid.layers.cross_entropy(
+            input=prediction, label=input_y)
+        avg_cost = paddle.fluid.layers.mean(x=cost)
+
+        strategy = paddle.fleet.DistributedStrategy()
+        strategy.gradient_merge = True
+        strategy.gradient_merge_configs = {"k_steps": 2, "avg": True}
+        optimizer = paddle.optimizer.SGD(learning_rate=0.01)
+        optimizer = fleet.distributed_optimizer(optimizer, strategy=strategy)
+        optimizer.minimize(avg_cost)
+
+
+if __name__ == "__main__":
+    unittest.main()