fleet_base.py 35.0 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
#   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.

from __future__ import print_function
16
import copy
17
import warnings
18
import paddle
19
from paddle.fluid.framework import dygraph_only
20
from paddle.fluid import compiler
21
from .role_maker import UserDefinedRoleMaker, PaddleCloudRoleMaker, RoleMakerBase
22
from .strategy_compiler import StrategyCompiler
23
from .distributed_strategy import DistributedStrategy
24 25 26
from .meta_optimizer_factory import MetaOptimizerFactory
from .runtime_factory import RuntimeFactory
from .util_factory import UtilFactory
27
from paddle.fluid.wrapped_decorator import wrap_decorator
28
from paddle.fluid.dygraph import parallel_helper
29

30

31 32 33 34 35 36 37 38 39 40 41 42
def _inited_runtime_handler_(func):
    def __impl__(*args, **kwargs):
        cls = args[0]

        if cls._runtime_handle is None:
            raise ValueError("Fleet can not find suitable runtime handler")

        return func(*args, **kwargs)

    return __impl__


43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58
def _is_non_distributed_check_(func):
    def __impl__(*args, **kwargs):
        cls = args[0]

        if cls._role_maker is not None and cls._role_maker._is_non_distributed(
        ) is True:
            warnings.warn(
                "%s() function doesn't work when use non_distributed fleet." %
                (func.__name__))
            return

        return func(*args, **kwargs)

    return __impl__


59
inited_runtime_handler = wrap_decorator(_inited_runtime_handler_)
60
is_non_distributed_check = wrap_decorator(_is_non_distributed_check_)
61 62


63 64 65
class Fleet(object):
    """
    Unified API for distributed training of PaddlePaddle
66
    Please reference the https://github.com/PaddlePaddle/FleetX for details
67 68 69 70 71


    Returns:
        Fleet: A Fleet instance

72
    Example for collective training:
73 74
        .. code-block:: python

75
            import paddle.distributed.fleet as fleet
76 77 78

            fleet.init(is_collective=True)

79 80 81
            strategy = fleet.DistributedStrategy()
            optimizer = paddle.optimizer.SGD(learning_rate=0.001)
            optimizer = fleet.distributed_optimizer(optimizer, strategy=strategy)
82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97

            # do distributed training


    Example for parameter server training:

        .. code-block:: python

            import paddle.distributed.fleet as fleet

            fleet.init()

            strategy = fleet.DistributedStrategy()
            optimizer = paddle.optimizer.SGD(learning_rate=0.001)
            optimizer = fleet.distributed_optimizer(optimizer, strategy=strategy)

98 99
            if fleet.is_first_worker():
                print("this is first worker")
100

101 102
            print("current node index: {}".format(fleet.worker_index()))
            print("total number of worker num: {}".format(fleet.worker_num()))
103

104 105 106
            if fleet.is_worker():
                print("this is worker")
            print("worker endpoints: {}".format(fleet.worker_endpoints(to_string=True)))
107

108 109
            print("server num: {}".format(fleet.server_num()))
            print("server endpoints: {}".format(fleet.server_endpoints(to_string=True)))
110

111 112 113
            if fleet.is_server():
                print("this is server")
            fleet.stop_worker()
114 115


116 117 118
    """

    def __init__(self):
119
        self._role_maker = None
120
        self.strategy_compiler = None
121
        self._is_collective = False
122 123
        self._runtime_handle = None
        self._util = None
124

125 126 127 128
    def init(self, role_maker=None, is_collective=False):
        """
        Initialize role_maker in Fleet.

129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163
        This function is responsible for the distributed architecture
        what you want to run your code behind.

        Args:
            role_maker (RoleMakerBase, optional): A ``RoleMakerBase`` containing the configuration
                of environment variables related to distributed training.If you did not initialize 
                the rolemaker by yourself, it will be automatically initialized to PaddleRoleMaker.
                The default value is None.
            is_collective (Boolean, optional): A ``Boolean`` variable determines whether the program 
                runs on the CPU or GPU. False means set distributed training using CPU, and True means
                GPU.The default value is False.The default value is False.
        Returns:
            None

        Examples1:

            .. code-block:: python

                import paddle.distributed.fleet as fleet
                fleet.init()

        Examples2:

            .. code-block:: python

                import paddle.distributed.fleet as fleet
                fleet.init(is_collective=True)

        Examples3:

            .. code-block:: python

                import paddle.distributed.fleet as fleet
                role = fleet.PaddleCloudRoleMaker
                fleet.init(role)
164

165
        """
166 167

        if role_maker is None:
168 169 170 171 172 173
            if isinstance(is_collective, bool):
                self._is_collective = is_collective
                self._role_maker = PaddleCloudRoleMaker(
                    is_collective=self._is_collective)
            else:
                raise ValueError(
174 175
                    "`is_collective` should be instance of `bool`, but got {}".
                    format(type(is_collective)))
176
        else:
177 178 179 180 181 182
            if isinstance(role_maker, RoleMakerBase):
                self._role_maker = role_maker
            else:
                raise ValueError(
                    "`role_maker` should be subclass of `RoleMakerBase`, but got {}".
                    format(type(role_maker)))
183
        self.strategy_compiler = StrategyCompiler()
184 185 186 187 188 189
        if paddle.fluid.framework.in_dygraph_mode():
            if parallel_helper._is_parallel_ctx_initialized():
                warnings.warn(
                    "The dygraph parallel environment has been initialized.")
            else:
                paddle.distributed.init_parallel_env()
190
        return None
191 192 193 194 195 196 197 198

    def is_first_worker(self):
        """
        Check whether the node is the first instance of worker.

        Returns:
            bool: True if this is the first node of worker,
                  False if not.
199

200 201 202 203 204 205 206 207
        Examples:

            .. code-block:: python

                import paddle.distributed.fleet as fleet
                fleet.init()
                fleet.is_first_worker()

208 209 210 211 212 213 214 215 216
        """
        return self._role_maker.is_first_worker()

    def worker_index(self):
        """
        Get current worker index.

        Returns:
            int: node id
217 218 219 220 221 222 223 224

        Examples:

            .. code-block:: python
                import paddle.distributed.fleet as fleet
                fleet.init()
                fleet.worker_index()

225 226 227 228 229 230 231 232 233
        """
        return self._role_maker.worker_index()

    def worker_num(self):
        """
        Get current total worker number.

        Returns:
            int: worker numbers
234

235 236 237 238 239 240 241
        Examples:
            .. code-block:: python

                import paddle.distributed.fleet as fleet
                fleet.init()
                fleet.worker_num()

242 243 244 245 246 247 248 249 250 251
        """
        return self._role_maker.worker_num()

    def is_worker(self):
        """
        Check whether the node is an instance of worker.

        Returns:
            bool: True if this is a node of worker,
                  False if not.
252 253 254 255 256 257 258 259

        Examples:
            .. code-block:: python

                import paddle.distributed.fleet as fleet
                fleet.init()
                fleet.is_worker()

260 261 262 263 264
        """
        return self._role_maker.is_worker()

    def worker_endpoints(self, to_string=False):
        """
265
        Get current worker endpoints, such as ["127.0.0.1:1001", "127.0.0.1:1002"].
266 267 268

        Returns:
            list/string: server endpoints
269 270 271 272 273 274 275 276

        Examples:
            .. code-block:: python

                import paddle.distributed.fleet as fleet
                fleet.init()
                fleet.worker_endpoints()

277 278 279 280 281 282 283 284 285 286 287 288 289 290 291
        """
        '''
        if to_string:
            return ",".join(self._role_maker.get_trainer_endpoints())
        else:
            return self._role_maker.get_trainer_endpoints()
        '''
        return ["127.0.0.1:1001", "127.0.0.1:1002"]

    def server_num(self):
        """
        Get current total worker number.

        Returns:
            int: server number
292 293 294 295 296 297

        Examples:
            .. code-block:: python
            import paddle.distributed.fleet as fleet
            fleet.init()
            fleet.server_num()
298 299 300 301 302 303 304 305 306
        """
        return len(self._role_maker.get_pserver_endpoints())

    def server_index(self):
        """
        Get current server index.

        Returns:
            int: node id
307 308 309 310 311 312 313 314

        Examples:
            .. code-block:: python

                import paddle.distributed.fleet as fleet
                fleet.init()
                fleet.server_index()

315 316 317 318 319 320 321 322 323
        """
        return self._role_maker.server_index()

    def server_endpoints(self, to_string=False):
        """
        Get current server endpoints, such as ["127.0.0.1:1001", "127.0.0.1:1002"].

        Returns:
            list/string: server endpoints
324 325 326 327 328 329 330 331

        Examples:
            .. code-block:: python

                import paddle.distributed.fleet as fleet
                fleet.init()
                fleet.server_endpoints()

332
        """
333

334 335 336 337 338 339 340 341 342 343 344 345
        if to_string:
            return ",".join(self._role_maker.get_pserver_endpoints())
        else:
            return self._role_maker.get_pserver_endpoints()

    def is_server(self):
        """
        Check whether the node is an instance of server.

        Returns:
            bool: True if this is a node of server,
                  False if not.
346 347 348 349 350 351 352 353

        Examples:

            .. code-block:: python
                import paddle.distributed.fleet as fleet
                fleet.init()
                fleet.is_server()

354
        """
355 356
        return self._role_maker.is_server(
        ) or self._role_maker._is_heter_worker()
357 358 359 360 361 362

    @property
    def util(self):
        """
        Utility functions that can be used under certain runtime
        return util
363 364 365 366 367 368 369 370 371 372 373 374 375

        Returns:
            UtilBase: instance of UtilBase, can use distributed ops/tools easily.

        Examples:

            .. code-block:: python
                import paddle.distributed.fleet as fleet
                fleet.init()
                util = fleet.util
                files = ["1.log", "2.log", "3.log", "4.log"]
                files = util.get_file_shard()

376 377 378 379 380 381 382
        """
        return self._util

    @util.setter
    def util(self, util):
        """
        Set Utility functions for userd-defined runtime
383 384 385

        Returns:
            None
386 387 388 389 390
        """
        self._util = util

    def barrier_worker(self):
        """
391 392 393 394
        barrier all workers

        Returns:
            None
395 396 397
        """
        self._role_maker.barrier_worker()

398
    @is_non_distributed_check
399
    @inited_runtime_handler
400 401
    def init_worker(self):
        """
402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419
        initialize `Communicator` for parameter server training.


        Returns:
            None

        Examples:

            .. code-block:: python

                import paddle.distributed.fleet as fleet
                fleet.init()

                # build net
                # fleet.distributed_optimizer(...)

                fleet.init_worker()

420 421 422
        """
        self._runtime_handle._init_worker()

423
    @is_non_distributed_check
424
    @inited_runtime_handler
425
    def init_server(self, *args, **kwargs):
426
        """
427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445
        init_server executor to initialize startup program,
        if the `args` is not empty, it will run load_persistables for increment training.


        Returns:
            None

        Examples:

            .. code-block:: python

                import paddle.distributed.fleet as fleet
                fleet.init()

                # build net
                # fleet.distributed_optimizer(...)

                fleet.init_server()

446
        """
447
        self._runtime_handle._init_server(*args, **kwargs)
448

449
    @is_non_distributed_check
450
    @inited_runtime_handler
451 452
    def run_server(self):
        """
453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470
        run server will run pserver main program with executor.

        Returns:
            None

        Examples:

            .. code-block:: python

                import paddle.distributed.fleet as fleet
                fleet.init()

                # build net
                # fleet.distributed_optimizer(...)

                if fleet.is_server():
                    fleet.init_server()

471 472 473
        """
        self._runtime_handle._run_server()

474
    @is_non_distributed_check
475
    @inited_runtime_handler
476 477
    def stop_worker(self):
        """
478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494
        stop `Communicator` and give training complete notice to parameter server.

        Returns:
            None

        Examples:

            .. code-block:: python

                import paddle.distributed.fleet as fleet
                fleet.init()

                # build net
                # fleet.distributed_optimizer(...)

                fleet.init_server()

495 496 497
        """
        self._runtime_handle._stop_worker()

498 499 500 501 502 503 504
    def save_inference_model(self,
                             executor,
                             dirname,
                             feeded_var_names,
                             target_vars,
                             main_program=None,
                             export_for_deployment=True):
505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524
        """
        save inference model for inference.

        Returns:
            None

        Examples:

            .. code-block:: python

                import paddle.distributed.fleet as fleet
                fleet.init()

                # build net
                # fleet.distributed_optimizer(...)

                fleet.init_server()

        """

525 526 527 528 529
        self._runtime_handle._save_inference_model(
            executor, dirname, feeded_var_names, target_vars, main_program,
            export_for_deployment)

    def save_persistables(self, executor, dirname, main_program=None):
530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569
        """

        saves all persistable variables from :code:`main_program` to
        the folder :code:`dirname`. You can refer to

        The :code:`dirname` is used to specify the folder where persistable variables
        are going to be saved. If you would like to save variables in separate
        files, set :code:`filename` None.

        Args:
            executor(Executor): The executor to run for saving persistable variables.
                                You can refer to :ref:`api_guide_executor_en` for
                                more details.

            dirname(str, optional): The saving directory path.
                                When you need to save the parameter to the memory, set it to None.
            main_program(Program, optional): The program whose persistbale variables will
                                             be saved. Default: None.


        Returns:
            None

        Examples:

            .. code-block:: text

                import paddle.distributed.fleet as fleet
                import paddle.fluid as fluid

                fleet.init()

                # build net
                # fleet.distributed_optimizer(...)

                exe = fluid.Executor(fluid.CPUPlace())
                fleet.save_persistables(exe, "dirname", fluid.default_main_program())

        """

570 571
        self._runtime_handle._save_persistables(executor, dirname, main_program)

572
    def distributed_optimizer(self, optimizer, strategy=None):
573
        """
574 575 576 577 578 579 580 581 582
        Optimizer for distributed training.

        For the distributed training, this method would rebuild a new instance of DistributedOptimizer.
        Which has basic Optimizer function and special features for distributed training.

        Args:
            optimizer(Optimizer): The executor to run for init server.
            strategy(DistributedStrategy): Extra properties for distributed optimizer.

583
        Returns:
584
            Fleet: instance of fleet.
585 586

        Examples:
587

588
            .. code-block:: python
589 590 591 592 593 594 595 596

                import paddle.distributed.fleet as fleet
                role = fleet.role_maker.PaddleCloudRoleMaker(is_collective=True)
                fleet.init(role)
                strategy = fleet.DistributedStrategy()
                optimizer = paddle.optimizer.SGD(learning_rate=0.001)
                optimizer = fleet.distributed_optimizer(optimizer, strategy=strategy)

597 598
        """
        self.user_defined_optimizer = optimizer
599 600 601
        if paddle.fluid.framework.in_dygraph_mode():
            return self

602 603
        if strategy == None:
            strategy = DistributedStrategy()
604
        self.user_defined_strategy = strategy
D
Dong Daxiang 已提交
605
        self.valid_strategy = None
606 607
        return self

608 609 610
    @dygraph_only
    def distributed_model(self, model):
        """
611 612 613 614 615 616 617
        Return distributed data parallel model (Only work in dygraph mode)

        Args:
            model (Layer): the user-defind model which inherits Layer.

        Returns:
            distributed data parallel model which inherits Layer.
618 619

        Examples:
620

621 622
            .. code-block:: python

623 624 625 626 627 628 629 630 631
                import paddle
                import paddle.nn as nn
                from paddle.distributed import fleet

                class LinearNet(nn.Layer):
                    def __init__(self):
                        super(LinearNet, self).__init__()
                        self._linear1 = nn.Linear(10, 10)
                        self._linear2 = nn.Linear(10, 1)
632

633 634
                    def forward(self, x):
                        return self._linear2(self._linear1(x))
635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666

                # 1. enable dynamic mode
                paddle.disable_static()

                # 2. initialize fleet environment
                fleet.init(is_collective=True)

                # 3. create layer & optimizer
                layer = LinearNet()
                loss_fn = nn.MSELoss()
                adam = paddle.optimizer.Adam(
                    learning_rate=0.001, parameters=layer.parameters())

                # 4. get data_parallel model using fleet
                adam = fleet.distributed_optimizer(adam)
                dp_layer = fleet.distributed_model(layer)

                # 5. run layer
                inputs = paddle.randn([10, 10], 'float32')
                outputs = dp_layer(inputs)
                labels = paddle.randn([10, 1], 'float32')
                loss = loss_fn(outputs, labels)

                print("loss:", loss.numpy())

                loss = dp_layer.scale_loss(loss)
                loss.backward()
                dp_layer.apply_collective_grads()

                adam.step()
                adam.clear_grad()

667

668 669 670 671 672 673 674 675 676
        """
        assert model is not None
        self.model = paddle.DataParallel(model)
        return self.model

    @dygraph_only
    def state_dict(self):
        """
        Get state dict information from optimizer.
677
        (Only work in dygraph mode)
678 679 680 681 682 683 684

        Returns: 
            state_dict(dict) : dict contains all the Tensor used by optimizer

        Examples:
            .. code-block:: python

685 686 687 688 689 690
                import numpy as np
                import paddle
                from paddle.distributed import fleet

                paddle.disable_static()
                fleet.init(is_collective=True)
691

692 693
                value = np.arange(26).reshape(2, 13).astype("float32")
                a = paddle.fluid.dygraph.to_variable(value)
694

695 696
                layer = paddle.nn.Linear(13, 5)
                adam = paddle.optimizer.Adam(learning_rate=0.01, parameters=layer.parameters())
697

698 699 700
                adam = fleet.distributed_optimizer(adam)
                dp_layer = fleet.distributed_model(layer)
                state_dict = adam.state_dict()
701 702 703 704 705 706 707 708
        """
        # imitate target optimizer retrieval
        return self.user_defined_optimizer.state_dict()

    @dygraph_only
    def set_state_dict(self, state_dict):
        """
        Load optimizer state dict.
709
        (Only work in dygraph mode)
710 711 712 713

        Args: 
            state_dict(dict) : Dict contains all the Tensor needed by optimizer

714 715
        Returns:
            None
716 717 718 719

        Examples:
            .. code-block:: python

720 721 722
                import numpy as np
                import paddle
                from paddle.distributed import fleet
723

724 725 726 727 728
                paddle.disable_static()
                fleet.init(is_collective=True)

                value = np.arange(26).reshape(2, 13).astype("float32")
                a = paddle.fluid.dygraph.to_variable(value)
729

730 731
                layer = paddle.nn.Linear(13, 5)
                adam = paddle.optimizer.Adam(learning_rate=0.01, parameters=layer.parameters())
732

733 734 735 736 737 738
                adam = fleet.distributed_optimizer(adam)
                dp_layer = fleet.distributed_model(layer)
                state_dict = adam.state_dict()
                paddle.framework.save(state_dict, "paddle_dy")
                para_state_dict, opti_state_dict = paddle.framework.load( "paddle_dy")
                adam.set_state_dict(opti_state_dict)
739 740 741 742 743 744 745 746
        """
        # imitate target optimizer retrieval
        return self.user_defined_optimizer.set_state_dict(state_dict)

    @dygraph_only
    def set_lr(self, value):
        """
        Set the value of the learning rate manually in the optimizer. 
747
        (Only work in dygraph mode)
748

749 750 751
        Args:
            value (float|Tensor): the value of learning rate

752 753
        Returns: 
            None 
754 755 756 757

        Examples:
            .. code-block:: python

758 759 760
                import numpy as np
                import paddle
                from paddle.distributed import fleet
761

762 763
                paddle.disable_static()
                fleet.init(is_collective=True)
764

765 766
                value = np.arange(26).reshape(2, 13).astype("float32")
                a = paddle.fluid.dygraph.to_variable(value)
767

768 769
                layer = paddle.nn.Linear(13, 5)
                adam = paddle.optimizer.Adam(learning_rate=0.01, parameters=layer.parameters())
770

771 772 773 774 775 776 777 778 779 780 781 782 783 784
                adam = fleet.distributed_optimizer(adam)
                dp_layer = fleet.distributed_model(layer)

                lr_list = [0.2, 0.3, 0.4, 0.5, 0.6]
                for i in range(5):
                    adam.set_lr(lr_list[i])
                    lr = adam.get_lr()
                    print("current lr is {}".format(lr))
                # Print:
                #    current lr is 0.2
                #    current lr is 0.3
                #    current lr is 0.4
                #    current lr is 0.5
                #    current lr is 0.6
785 786 787 788 789 790 791 792
        """
        # imitate target optimizer retrieval
        return self.user_defined_optimizer.set_lr(value)

    @dygraph_only
    def get_lr(self):
        """
        Get current step learning rate.
793
        (Only work in dygraph mode)
794 795 796 797 798 799 800

        Returns:
            float: The learning rate of the current step.

        Examples:
            .. code-block:: python

801 802 803 804 805 806
                import numpy as np
                import paddle
                from paddle.distributed import fleet

                paddle.disable_static()
                fleet.init(is_collective=True)
807

808 809
                value = np.arange(26).reshape(2, 13).astype("float32")
                a = paddle.fluid.dygraph.to_variable(value)
810

811 812
                layer = paddle.nn.Linear(13, 5)
                adam = paddle.optimizer.Adam(learning_rate=0.01, parameters=layer.parameters())
813

814 815
                adam = fleet.distributed_optimizer(adam)
                dp_layer = fleet.distributed_model(layer)
816

817 818
                lr = adam.get_lr()
                print(lr) # 0.01
819 820 821 822 823 824 825 826
        """
        # imitate target optimizer retrieval
        return self.user_defined_optimizer.get_lr()

    @dygraph_only
    def step(self):
        """
        Execute the optimizer once.
827
        (Only work in dygraph mode)
828

829 830
        Returns:
            None
831 832 833 834

        Examples:
            .. code-block:: python

835 836 837
                import paddle
                import paddle.nn as nn
                from paddle.distributed import fleet
838

839 840 841 842 843
                class LinearNet(nn.Layer):
                    def __init__(self):
                        super(LinearNet, self).__init__()
                        self._linear1 = nn.Linear(10, 10)
                        self._linear2 = nn.Linear(10, 1)
844

845 846
                    def forward(self, x):
                        return self._linear2(self._linear1(x))
847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886

                # 1. enable dynamic mode
                paddle.disable_static()

                # 2. initialize fleet environment
                fleet.init(is_collective=True)

                # 3. create layer & optimizer
                layer = LinearNet()
                loss_fn = nn.MSELoss()
                adam = paddle.optimizer.Adam(
                    learning_rate=0.001, parameters=layer.parameters())

                # 4. get data_parallel model using fleet
                adam = fleet.distributed_optimizer(adam)
                dp_layer = fleet.distributed_model(layer)

                # 5. run layer
                inputs = paddle.randn([10, 10], 'float32')
                outputs = dp_layer(inputs)
                labels = paddle.randn([10, 1], 'float32')
                loss = loss_fn(outputs, labels)

                print("loss:", loss.numpy())

                loss = dp_layer.scale_loss(loss)
                loss.backward()
                dp_layer.apply_collective_grads()

                adam.step()
                adam.clear_grad()


        """
        # imitate target optimizer retrieval
        return self.user_defined_optimizer.step()

    @dygraph_only
    def clear_grad(self):
        """
887 888
        Clear the gradients of all optimized parameters for model.
        (Only work in dygraph mode)
889

890 891
        Returns: 
            None
892 893 894 895

        Examples:
            .. code-block:: python

896 897 898
                import paddle
                import paddle.nn as nn
                from paddle.distributed import fleet
899

900 901 902 903 904
                class LinearNet(nn.Layer):
                    def __init__(self):
                        super(LinearNet, self).__init__()
                        self._linear1 = nn.Linear(10, 10)
                        self._linear2 = nn.Linear(10, 1)
905

906 907
                    def forward(self, x):
                        return self._linear2(self._linear1(x))
908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943

                # 1. enable dynamic mode
                paddle.disable_static()

                # 2. initialize fleet environment
                fleet.init(is_collective=True)

                # 3. create layer & optimizer
                layer = LinearNet()
                loss_fn = nn.MSELoss()
                adam = paddle.optimizer.Adam(
                    learning_rate=0.001, parameters=layer.parameters())

                # 4. get data_parallel model using fleet
                adam = fleet.distributed_optimizer(adam)
                dp_layer = fleet.distributed_model(layer)

                # 5. run layer
                inputs = paddle.randn([10, 10], 'float32')
                outputs = dp_layer(inputs)
                labels = paddle.randn([10, 1], 'float32')
                loss = loss_fn(outputs, labels)

                print("loss:", loss.numpy())

                loss = dp_layer.scale_loss(loss)
                loss.backward()
                dp_layer.apply_collective_grads()

                adam.step()
                adam.clear_grad()

        """
        # imitate target optimizer retrieval
        return self.user_defined_optimizer.clear_grad()

944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966
    def minimize(self,
                 loss,
                 startup_program=None,
                 parameter_list=None,
                 no_grad_set=None):
        """
        Add distributed operations to minimize ``loss`` by updating ``parameter_list``.

        Args:
            loss (Variable): A ``Variable`` containing the value to minimize.
            startup_program (Program, optional): :ref:`api_fluid_Program` for
                initializing parameters in ``parameter_list``. The default value
                is None, at this time :ref:`api_fluid_default_startup_program` will be used.
            parameter_list (Iterable, optional): Iterable of ``Variable`` or ``Variable.name`` to update
                to minimize ``loss``. The default value is None, at this time all parameters
                will be updated.
            no_grad_set (set, optional): Set of ``Variable``  or ``Variable.name`` that don't need
                to be updated. The default value is None.

        Returns:
            tuple: tuple (optimize_ops, params_grads), A list of operators appended
            by minimize and a list of (param, grad) variable pairs, param is
            ``Parameter``, grad is the gradient value corresponding to the parameter.
967 968
            The returned tuple can be passed to ``fetch_list`` in ``Executor.run()`` to
            indicate program pruning. If so, the program will be pruned by ``feed`` and
969 970 971
            ``fetch_list`` before run, see details in ``Executor``.

        Examples:
972
            .. code-block:: python
973

974 975
                import paddle
                import paddle.distributed.fleet as fleet
976

977 978 979 980 981 982 983 984 985 986 987 988
                fc_1 = paddle.fluid.layers.fc(input=input_x, size=hid_dim, act='tanh')
                fc_2 = paddle.fluid.layers.fc(input=fc_1, size=hid_dim, act='tanh')
                prediction = paddle.fluid.layers.fc(input=[fc_2], size=label_dim, act='softmax')
                cost = paddle.fluid.layers.cross_entropy(input=prediction, label=input_y)
                avg_cost = paddle.fluid.layers.mean(x=cost)

                role = fleet.role_maker.PaddleCloudRoleMaker(is_collective=True)
                fleet.init(role)
                strategy = fleet.DistributedStrategy()
                optimizer = paddle.optimizer.SGD(learning_rate=0.001)
                optimizer = fleet.distributed_optimizer(optimizer, strategy=strategy)
                optimizer.minimize(avg_cost)
989

990
                # for more examples, please reference https://github.com/PaddlePaddle/FleetX
991 992

        """
993 994 995 996 997
        if paddle.fluid.framework.in_dygraph_mode():
            # imitate target optimizer retrieval
            target_opt = self.user_defined_optimizer
            return target_opt.minimize(loss)

998
        context = {}
999 1000
        # cache original feed forward program
        self.origin_main_program = loss.block.program
1001 1002
        context["origin_main_program"] = self.origin_main_program
        context["loss"] = loss
1003 1004
        if startup_program == None:
            self.origin_startup_program = \
1005 1006
                paddle.static.default_startup_program().clone(for_test=False)
            startup_program = paddle.static.default_startup_program()
1007 1008 1009
        else:
            self.origin_startup_program = \
                startup_program.clone(for_test=False)
1010

1011 1012
        context["origin_startup_program"] = startup_program
        context["role_maker"] = self._role_maker
1013 1014 1015 1016 1017

        # compile time
        distributed_optimizer_list = \
            MetaOptimizerFactory()._get_valid_meta_optimizers(
                self.user_defined_optimizer)
D
Dong Daxiang 已提交
1018

1019 1020 1021 1022 1023 1024 1025 1026 1027 1028
        context["user_defined_strategy"] = copy.copy(self.user_defined_strategy)

        # trigger the auto-parallel in very strict condition
        # strategy = DistributedStrategy()
        # strategy.auto = True
        # optimizer = paddle.optimizer.SGD(learning_rate=0.1)
        # optimizer = fleet.distributed_optimizer(optimizer, strategy)
        if self.user_defined_strategy._is_strict_auto():
            # turn on all the strategy for each optimizer
            for opt in distributed_optimizer_list:
1029
                opt._enable_strategy(self.user_defined_strategy, context)
1030

1031 1032
        valid_optimizer_list = []
        valid_graph_optimizer_list = []
D
Dong Daxiang 已提交
1033
        can_not_apply_optimizer_list = []
1034 1035 1036 1037 1038 1039 1040
        # recall meta optimizers for ranking
        for opt in distributed_optimizer_list:
            opt._set_basic_info(loss, self._role_maker,
                                self.user_defined_optimizer,
                                self.user_defined_strategy)
            if opt._can_apply() and not opt._is_graph_out():
                valid_optimizer_list.append(opt)
D
Dong Daxiang 已提交
1041
            elif opt._can_apply() and opt._is_graph_out():
1042
                valid_graph_optimizer_list.append(opt)
D
Dong Daxiang 已提交
1043 1044
            else:
                can_not_apply_optimizer_list.append(opt)
1045
        # combine recalled meta optimizers to be a valid meta optimizer
D
Dong Daxiang 已提交
1046
        meta_optimizer, graph_optimizer = \
1047 1048 1049 1050
            self.strategy_compiler.generate_optimizer(
                loss, self._role_maker, self.user_defined_optimizer,
                self.user_defined_strategy, valid_optimizer_list,
                valid_graph_optimizer_list)
D
Dong Daxiang 已提交
1051

D
Dong Daxiang 已提交
1052 1053
        valid_strategy = self.strategy_compiler._get_valid_strategy(
            self.user_defined_strategy, can_not_apply_optimizer_list)
1054 1055 1056

        context["valid_strategy"] = valid_strategy

D
Dong Daxiang 已提交
1057
        self.valid_strategy = valid_strategy
1058
        self.valid_strategy._enable_env()
D
Dong Daxiang 已提交
1059

1060 1061
        optimize_ops = []
        params_grads = []
1062

1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076
        if self._role_maker._is_non_distributed() and not self._is_collective:
            if self._runtime_handle is None:
                self._runtime_handle = RuntimeFactory()._create_runtime(context)

            compiled_program = compiler.CompiledProgram(
                self.origin_main_program).with_data_parallel(
                    loss_name=loss.name, share_vars_from=None)
            loss.block.program._graph = compiled_program
            return self.user_defined_optimizer.minimize(
                loss,
                startup_program=startup_program,
                parameter_list=parameter_list,
                no_grad_set=no_grad_set)

1077 1078 1079 1080 1081 1082
        if meta_optimizer:
            optimize_ops, params_grads = meta_optimizer.minimize(
                loss,
                startup_program=startup_program,
                parameter_list=parameter_list,
                no_grad_set=no_grad_set)
1083

1084
            default_program = paddle.static.default_main_program()
1085 1086 1087 1088

            if id(default_program) != id(loss.block.program):
                paddle.fluid.framework.switch_main_program(loss.block.program)

1089 1090 1091 1092 1093 1094
        else:
            optimize_ops, params_grads = self.user_defined_optimizer.minimize(
                loss,
                startup_program=startup_program,
                parameter_list=parameter_list,
                no_grad_set=no_grad_set)
1095

1096 1097
        context["program_optimize_ops"] = optimize_ops
        context["program_params_grads"] = params_grads
1098

1099
        if graph_optimizer:
D
Dong Daxiang 已提交
1100
            optimize_ops, params_grads = graph_optimizer.minimize(
1101 1102 1103 1104 1105 1106 1107 1108
                loss,
                startup_program=startup_program,
                parameter_list=parameter_list,
                no_grad_set=no_grad_set)
            # since we do not encourage users to use graph operations
            # if a graph optimizer takes effect, mostly
            # optimizers_ops and params_grads are None
            # i.e. users can not modify current computation graph anymore
1109 1110 1111
            context["graph_optimize_ops"] = optimize_ops
            context["graph_optimize_grads"] = params_grads

1112
        if self._runtime_handle is None:
1113
            self._runtime_handle = RuntimeFactory()._create_runtime(context)
1114 1115

        if self._util is None:
1116
            self._util = UtilFactory()._create_util(context)
1117 1118

        return optimize_ops, params_grads