【paddle.fleet】add lamb to fleet meta optimizer (#26025)

add lamb to fleet meta optimizer

paddle/fluid/framework/distributed_strategy.proto

@@ -55,9 +55,8 @@ message LarsConfig {
 }
 
 message LambConfig {
-  optional float beta1 = 1 [ default = 0.001 ];
-  optional float beta2 = 2 [ default = 0.999 ];
-  optional float epsilon = 3 [ default = 0.000001 ];
+  optional float lamb_weight_decay = 1 [ default = 0.01 ];
+  repeated string exclude_from_weight_decay = 2;
 }
 
 message BuildStrategy {

python/paddle/fleet/base/distributed_strategy.py

@@ -627,6 +627,15 @@ class DistributedStrategy(object):
         else:
             print("WARNING: lamb should have value of bool type")
 
+    @property
+    def lamb_configs(self):
+        return get_msg_dict(self.strategy.lamb_configs)
+
+    @lamb_configs.setter
+    def lamb_configs(self, configs):
+        check_configs_key(self.strategy.lamb_configs, configs, "lamb_configs")
+        assign_configs_value(self.strategy.lamb_configs, configs)
+
     @property
     def elastic(self):
         return self.strategy.elastic

python/paddle/fleet/meta_optimizers/__init__.py

@@ -21,6 +21,7 @@ from .localsgd_optimizer import LocalSGDOptimizer
 from .lars_optimizer import LarsOptimizer
 from .async_graph_execution_optimizer import AsyncGraphExecutionOptimizer
 from .dgc_optimizer import DGCOptimizer
+from .lamb_optimizer import LambOptimizer
 
 __all__ = [
     'AMPOptimizer',
@@ -33,4 +34,5 @@ __all__ = [
     'LarsOptimizer',
     'AsyncGraphExecutionOptimizer',
     'DGCOptimizer',
+    'LambOptimizer',
 ]

python/paddle/fleet/meta_optimizers/lamb_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
+
+from paddle.fluid.optimizer import AdamOptimizer
+from paddle.fluid.optimizer import LambOptimizer as LAMB
+from .meta_optimizer_base import MetaOptimizerBase
+import logging
+
+__all__ = ["LambOptimizer"]
+
+
+class LambOptimizer(MetaOptimizerBase):
+    def __init__(self, optimizer):
+        super(LambOptimizer, self).__init__(optimizer)
+        self.inner_opt = optimizer
+        self.lamb_opt = None
+        # we do not allow meta optimizer to be inner optimizer currently
+        self.meta_optimizers_white_list = []
+
+    def _set_basic_info(self, loss, role_maker, user_defined_optimizer,
+                        user_defined_strategy):
+        super(LambOptimizer, self)._set_basic_info(
+            loss, role_maker, user_defined_optimizer, user_defined_strategy)
+
+        opt = self.inner_opt
+        if not isinstance(opt, AdamOptimizer):
+            return
+
+        configs = self.user_defined_strategy.lamb_configs
+        if len(configs['exclude_from_weight_decay']) == 0:
+            _exclude_from_weight_decay_fn = None
+        else:
+
+            def exclude_fn(param):
+                exclude_list = configs['exclude_from_weight_decay']
+                for name in exclude_list:
+                    if param.name.endswith(name):
+                        return True
+                return False
+
+            _exclude_from_weight_decay_fn = exclude_fn
+
+        self.lamb_opt = LAMB(
+            learning_rate=opt._learning_rate,
+            lamb_weight_decay=configs['lamb_weight_decay'],
+            beta1=opt._beta1,
+            beta2=opt._beta2,
+            epsilon=opt._epsilon,
+            parameter_list=opt._parameter_list,
+            regularization=opt.regularization,
+            grad_clip=opt._grad_clip,
+            exclude_from_weight_decay_fn=_exclude_from_weight_decay_fn,
+            name=opt._name)
+
+    def _can_apply(self):
+        if self.user_defined_strategy.lamb:
+            if not isinstance(self.inner_opt, AdamOptimizer):
+                logging.warn(
+                    "lamb need the inner optimizer to be AdamOptimizer optimizer but got {}.".
+                    format(self.inner_opt.type))
+                return False
+            return True
+        return False
+
+    def _disable_strategy(self, dist_strategy):
+        dist_strategy.lamb = False
+        dist_strategy.lamb_configs = {
+            'lamb_weight_decay': 0.01,
+            'exclude_from_weight_decay': [],
+        }
+
+    def backward(self,
+                 loss,
+                 startup_program=None,
+                 parameter_list=None,
+                 no_grad_set=None,
+                 callbacks=None):
+        return self.lamb_opt.backward(loss, startup_program, parameter_list,
+                                      no_grad_set, callbacks)
+
+    def minimize_impl(self,
+                      loss,
+                      startup_program=None,
+                      parameter_list=None,
+                      no_grad_set=None):
+        optimize_ops, params_grads = \
+            self.lamb_opt.minimize(loss, startup_program,
+                                      parameter_list, no_grad_set)
+        return optimize_ops, params_grads

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

@@ -40,6 +40,7 @@ list(APPEND MIXED_DIST_TEST_OPS test_fleet_amp_meta_optimizer)
 list(APPEND MIXED_DIST_TEST_OPS test_fleet_gradient_merge_meta_optimizer)
 list(APPEND MIXED_DIST_TEST_OPS test_fleet_localsgd_meta_optimizer)
 list(APPEND MIXED_DIST_TEST_OPS test_fleet_lars_meta_optimizer)
+list(APPEND MIXED_DIST_TEST_OPS test_fleet_lamb_meta_optimizer)
 list(APPEND MIXED_DIST_TEST_OPS test_fleet_dgc_meta_optimizer)
 list(APPEND MIXED_DIST_TEST_OPS test_fleet_private_function)
 list(APPEND MIXED_DIST_TEST_OPS test_fleet_graph_executor)
@@ -386,6 +387,7 @@ if(WITH_DISTRIBUTE)
         if(NOT WIN32)
             py_test_modules(test_fleet_localsgd_meta_optimizer MODULES test_fleet_localsgd_meta_optimizer ENVS ${dist_ENVS})
             py_test_modules(test_fleet_lars_meta_optimizer MODULES test_fleet_lars_meta_optimizer ENVS ${dist_ENVS})
+            py_test_modules(test_fleet_lamb_meta_optimizer MODULES test_fleet_lamb_meta_optimizer ENVS ${dist_ENVS})
         endif(NOT WIN32)
     endif(NOT APPLE)
     if(WITH_DGC)

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

@@ -14,6 +14,7 @@
 
 import unittest
 import paddle
+from paddle import fluid
 import os
 import paddle.fleet as fleet
 import paddle.fluid.incubate.fleet.base.role_maker as role_maker
@@ -25,31 +26,40 @@ class TestFleetDGCOptimizer(unittest.TestCase):
         os.environ[
             "PADDLE_TRAINER_ENDPOINTS"] = "127.0.0.1:36001,127.0.0.1:36002"
 
-    def net(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')
+    def net(self, main_prog, startup_prog):
+        with fluid.program_guard(main_prog, startup_prog):
+            with fluid.unique_name.guard():
+                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=256, 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)
+                fc_1 = paddle.fluid.layers.fc(input=input_x,
+                                              size=64,
+                                              act='tanh')
+                fc_2 = paddle.fluid.layers.fc(input=fc_1, size=256, 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.dgc = True
-        strategy.dgc_configs = {
-            "rampup_begin_step": 128,
-            "rampup_step": 100,
-            "sparsity": [0.996, 0.999]
-        }
+                strategy = paddle.fleet.DistributedStrategy()
+                strategy.dgc = True
+                strategy.dgc_configs = {
+                    "rampup_begin_step": 128,
+                    "rampup_step": 100,
+                    "sparsity": [0.996, 0.999]
+                }
         return avg_cost, strategy
 
     def test_dgc_optimizer(self):
-        avg_cost, strategy = self.net()
+        startup_prog = fluid.Program()
+        train_prog = fluid.Program()
+        avg_cost, strategy = self.net(train_prog, startup_prog)
         optimizer = paddle.optimizer.Momentum(learning_rate=0.01, momentum=0.9)
         optimizer = fleet.distributed_optimizer(optimizer, strategy=strategy)
         optimizer.minimize(avg_cost)
@@ -59,7 +69,9 @@ class TestFleetDGCOptimizer(unittest.TestCase):
         self.assertIn('dgc_momentum', ops)
 
     def test_dgc_not_apply_with_adam(self):
-        avg_cost, strategy = self.net()
+        startup_prog = fluid.Program()
+        train_prog = fluid.Program()
+        avg_cost, strategy = self.net(train_prog, startup_prog)
         optimizer = paddle.optimizer.Adam(learning_rate=0.01)
         optimizer = fleet.distributed_optimizer(optimizer, strategy=strategy)
         optimizer.minimize(avg_cost)
@@ -72,7 +84,9 @@ class TestFleetDGCOptimizer(unittest.TestCase):
         os.environ["PADDLE_TRAINER_ID"] = "0"
         os.environ["PADDLE_TRAINER_ENDPOINTS"] = "127.0.0.1:36001"
 
-        avg_cost, strategy = self.net()
+        startup_prog = fluid.Program()
+        train_prog = fluid.Program()
+        avg_cost, strategy = self.net(train_prog, startup_prog)
         optimizer = paddle.optimizer.Momentum(learning_rate=0.01, momentum=0.9)
         optimizer = fleet.distributed_optimizer(optimizer, strategy=strategy)
         optimizer.minimize(avg_cost)

python/paddle/fluid/tests/unittests/test_fleet_lamb_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
+from paddle import fluid
+import os
+import paddle.fleet as fleet
+import paddle.fluid.incubate.fleet.base.role_maker as role_maker
+
+
+class TestFleetLambMetaOptimizer(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 net(self, main_prog, startup_prog):
+        with fluid.program_guard(main_prog, startup_prog):
+            with fluid.unique_name.guard():
+                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=256, 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.lamb = True
+                strategy.lamb_configs = {
+                    'lamb_weight_decay': 0.01,
+                    'exclude_from_weight_decay': [],
+                }
+
+        return avg_cost, strategy
+
+    def test_lamb_optimizer(self):
+        role = role_maker.PaddleCloudRoleMaker(is_collective=True)
+        fleet.init(role)
+        startup_prog = fluid.Program()
+        train_prog = fluid.Program()
+        avg_cost, strategy = self.net(train_prog, startup_prog)
+        optimizer = paddle.optimizer.Adam(learning_rate=0.01)
+        optimizer = fleet.distributed_optimizer(optimizer, strategy=strategy)
+        optimizer.minimize(avg_cost)
+
+        ops = [op.type for op in avg_cost.block.ops]
+        self.assertIn('lamb', ops)
+
+    def test_lamb_not_apply_with_momentum(self):
+        role = role_maker.PaddleCloudRoleMaker(is_collective=True)
+        fleet.init(role)
+        startup_prog = fluid.Program()
+        train_prog = fluid.Program()
+        avg_cost, strategy = self.net(train_prog, startup_prog)
+        optimizer = paddle.optimizer.Momentum(learning_rate=0.1, momentum=0.9)
+        optimizer = fleet.distributed_optimizer(optimizer, strategy=strategy)
+        optimizer.minimize(avg_cost)
+
+        ops = [op.type for op in avg_cost.block.ops]
+        self.assertNotIn('lamb', ops)
+
+    def test_lamb_exclude_fn(self):
+        role = role_maker.PaddleCloudRoleMaker(is_collective=True)
+        fleet.init(role)
+        startup_prog = fluid.Program()
+        train_prog = fluid.Program()
+        avg_cost, strategy = self.net(train_prog, startup_prog)
+        optimizer = paddle.optimizer.Adam(learning_rate=0.01)
+        strategy.lamb_configs = {
+            'lamb_weight_decay': 0.01,
+            'exclude_from_weight_decay': ['.b_0'],
+        }
+        optimizer = fleet.distributed_optimizer(optimizer, strategy=strategy)
+        optimizer.minimize(avg_cost)
+
+        ops_with_bias = [
+            op for op in avg_cost.block.ops
+            if op.type == 'lamb' and op.attr('op_role_var')[0].endswith('.b_0')
+        ]
+        for op in ops_with_bias:
+            self.assertEqual(op.attr('weight_decay'), 0)
+
+
+if __name__ == "__main__":
+    unittest.main()

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

@@ -14,6 +14,7 @@
 
 import unittest
 import paddle
+from paddle import fluid
 import os
 import paddle.fleet as fleet
 import paddle.fluid.incubate.fleet.base.role_maker as role_maker
@@ -27,31 +28,40 @@ class TestFleetLarsMetaOptimizer(unittest.TestCase):
         os.environ["PADDLE_PSERVERS_IP_PORT_LIST"] = \
                        "127.0.0.1:36001,127.0.0.2:36001"
 
-    def net(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')
+    def net(self, main_prog, startup_prog):
+        with fluid.program_guard(main_prog, startup_prog):
+            with fluid.unique_name.guard():
+                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=256, 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)
+                fc_1 = paddle.fluid.layers.fc(input=input_x,
+                                              size=64,
+                                              act='tanh')
+                fc_2 = paddle.fluid.layers.fc(input=fc_1, size=256, 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.lars = True
-        strategy.lars_configs = {
-            "lars_coeff": 0.001,
-            "lars_weight_decay": 0.0005,
-        }
+                strategy = paddle.fleet.DistributedStrategy()
+                strategy.lars = True
+                strategy.lars_configs = {
+                    "lars_coeff": 0.001,
+                    "lars_weight_decay": 0.0005,
+                }
 
         return avg_cost, strategy
 
     def test_lars_optimizer(self):
-        avg_cost, strategy = self.net()
+        role = role_maker.PaddleCloudRoleMaker(is_collective=True)
+        fleet.init(role)
+        startup_prog = fluid.Program()
+        train_prog = fluid.Program()
+        avg_cost, strategy = self.net(train_prog, startup_prog)
         optimizer = paddle.optimizer.Momentum(learning_rate=0.01, momentum=0.9)
         optimizer = fleet.distributed_optimizer(optimizer, strategy=strategy)
         optimizer.minimize(avg_cost)
@@ -60,7 +70,11 @@ class TestFleetLarsMetaOptimizer(unittest.TestCase):
         self.assertIn('lars_momentum', ops)
 
     def test_lars_not_apply_with_adam(self):
-        avg_cost, strategy = self.net()
+        role = role_maker.PaddleCloudRoleMaker(is_collective=True)
+        fleet.init(role)
+        startup_prog = fluid.Program()
+        train_prog = fluid.Program()
+        avg_cost, strategy = self.net(train_prog, startup_prog)
         optimizer = paddle.optimizer.Adam(learning_rate=0.01)
         optimizer = fleet.distributed_optimizer(optimizer, strategy=strategy)
         optimizer.minimize(avg_cost)