fleet_base.py 34.9 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
from .meta_optimizer_factory import MetaOptimizerFactory
from .runtime_factory import RuntimeFactory
26
from paddle.fluid.wrapped_decorator import wrap_decorator
27
from paddle.fluid.dygraph import parallel_helper
28

29

30 31 32 33 34 35 36 37 38 39 40 41
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__


42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57
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__


58
inited_runtime_handler = wrap_decorator(_inited_runtime_handler_)
59
is_non_distributed_check = wrap_decorator(_is_non_distributed_check_)
60 61


62 63 64
class Fleet(object):
    """
    Unified API for distributed training of PaddlePaddle
65
    Please reference the https://github.com/PaddlePaddle/FleetX for details
66 67 68 69 70


    Returns:
        Fleet: A Fleet instance

71
    Example for collective training:
72 73
        .. code-block:: python

74
            import paddle.distributed.fleet as fleet
75 76 77

            fleet.init(is_collective=True)

78 79 80
            strategy = fleet.DistributedStrategy()
            optimizer = paddle.optimizer.SGD(learning_rate=0.001)
            optimizer = fleet.distributed_optimizer(optimizer, strategy=strategy)
81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96

            # 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)

97 98
            if fleet.is_first_worker():
                print("this is first worker")
99

100 101
            print("current node index: {}".format(fleet.worker_index()))
            print("total number of worker num: {}".format(fleet.worker_num()))
102

103 104 105
            if fleet.is_worker():
                print("this is worker")
            print("worker endpoints: {}".format(fleet.worker_endpoints(to_string=True)))
106

107 108
            print("server num: {}".format(fleet.server_num()))
            print("server endpoints: {}".format(fleet.server_endpoints(to_string=True)))
109

110 111 112
            if fleet.is_server():
                print("this is server")
            fleet.stop_worker()
113 114


115 116 117
    """

    def __init__(self):
118
        self._role_maker = None
119
        self.strategy_compiler = None
120
        self._is_collective = False
121
        self._runtime_handle = None
D
Dong Daxiang 已提交
122 123
        self._util = None
        self._context = {}
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._role_maker._generate_role()
184

185 186 187
        import paddle.distributed.fleet as fleet
        fleet.util._set_role_maker(self._role_maker)

188
        self.strategy_compiler = StrategyCompiler()
189 190 191 192 193 194
        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()
195 196 197 198 199 200 201 202

    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.
203

204 205 206 207 208 209 210 211
        Examples:

            .. code-block:: python

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

212
        """
213
        return self._role_maker._is_first_worker()
214 215 216 217 218 219 220

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

        Returns:
            int: node id
221 222 223 224 225 226 227 228

        Examples:

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

229
        """
230
        return self._role_maker._worker_index()
231 232 233 234 235 236 237

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

        Returns:
            int: worker numbers
D
Dong Daxiang 已提交
238
        
239 240 241 242 243 244 245
        Examples:
            .. code-block:: python

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

246
        """
247
        return self._role_maker._worker_num()
248 249 250 251 252 253 254 255

    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.
256 257 258 259 260 261 262 263

        Examples:
            .. code-block:: python

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

264
        """
265
        return self._role_maker._is_worker()
266 267 268

    def worker_endpoints(self, to_string=False):
        """
269
        Get current worker endpoints, such as ["127.0.0.1:1001", "127.0.0.1:1002"].
270 271 272

        Returns:
            list/string: server endpoints
273 274 275 276 277 278 279 280

        Examples:
            .. code-block:: python

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

281 282
        """
        if to_string:
283
            return ",".join(self._role_maker._get_trainer_endpoints())
284
        else:
285
            return self._role_maker._get_trainer_endpoints()
286 287 288 289 290 291 292

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

        Returns:
            int: server number
293 294 295 296 297 298

        Examples:
            .. code-block:: python
            import paddle.distributed.fleet as fleet
            fleet.init()
            fleet.server_num()
299
        """
300
        return len(self._role_maker._get_pserver_endpoints())
301 302 303 304 305 306 307

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

        Returns:
            int: node id
308 309 310 311 312 313 314 315

        Examples:
            .. code-block:: python

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

316
        """
317
        return self._role_maker._server_index()
318 319 320 321 322 323 324

    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
325 326 327 328 329 330 331 332

        Examples:
            .. code-block:: python

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

333
        """
334

335
        if to_string:
336
            return ",".join(self._role_maker._get_pserver_endpoints())
337
        else:
338
            return self._role_maker._get_pserver_endpoints()
339 340 341 342 343 344 345 346

    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.
347 348 349 350 351 352 353 354

        Examples:

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

355
        """
356
        return self._role_maker._is_server(
357
        ) or self._role_maker._is_heter_worker()
358 359 360

    def barrier_worker(self):
        """
361 362 363 364
        barrier all workers

        Returns:
            None
365
        """
366
        self._role_maker._barrier("worker")
367

368
    @is_non_distributed_check
369
    @inited_runtime_handler
370 371
    def init_worker(self):
        """
372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389
        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()

390 391 392
        """
        self._runtime_handle._init_worker()

393
    @is_non_distributed_check
394
    @inited_runtime_handler
395
    def init_server(self, *args, **kwargs):
396
        """
397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415
        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()

416
        """
417
        self._runtime_handle._init_server(*args, **kwargs)
418

419
    @is_non_distributed_check
420
    @inited_runtime_handler
421 422
    def run_server(self):
        """
423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440
        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()

441 442 443
        """
        self._runtime_handle._run_server()

444
    @is_non_distributed_check
445
    @inited_runtime_handler
446 447
    def stop_worker(self):
        """
448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464
        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()

465 466 467
        """
        self._runtime_handle._stop_worker()

468 469 470 471 472 473 474
    def save_inference_model(self,
                             executor,
                             dirname,
                             feeded_var_names,
                             target_vars,
                             main_program=None,
                             export_for_deployment=True):
475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494
        """
        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()

        """

495 496 497 498 499
        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):
500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539
        """

        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())

        """

540 541
        self._runtime_handle._save_persistables(executor, dirname, main_program)

542
    def distributed_optimizer(self, optimizer, strategy=None):
543
        """
544 545 546 547 548 549 550 551 552
        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.

553
        Returns:
554
            Fleet: instance of fleet.
555 556

        Examples:
557

558
            .. code-block:: python
559 560 561 562 563 564 565 566

                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)

567 568
        """
        self.user_defined_optimizer = optimizer
569 570 571
        if paddle.fluid.framework.in_dygraph_mode():
            return self

572 573
        if strategy == None:
            strategy = DistributedStrategy()
D
Dong Daxiang 已提交
574 575 576

        self._user_defined_strategy = copy.deepcopy(strategy)
        self._context = {}
577 578
        return self

579 580 581
    @dygraph_only
    def distributed_model(self, model):
        """
582 583 584 585 586 587 588
        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.
589 590

        Examples:
591

592 593
            .. code-block:: python

594 595 596 597 598 599 600 601 602
                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)
603

604 605
                    def forward(self, x):
                        return self._linear2(self._linear1(x))
606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637

                # 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()

638

639 640 641 642 643 644 645 646 647
        """
        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.
648
        (Only work in dygraph mode)
649 650 651 652 653 654 655

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

        Examples:
            .. code-block:: python

656 657 658 659 660 661
                import numpy as np
                import paddle
                from paddle.distributed import fleet

                paddle.disable_static()
                fleet.init(is_collective=True)
662

663 664
                value = np.arange(26).reshape(2, 13).astype("float32")
                a = paddle.fluid.dygraph.to_variable(value)
665

666 667
                layer = paddle.nn.Linear(13, 5)
                adam = paddle.optimizer.Adam(learning_rate=0.01, parameters=layer.parameters())
668

669 670 671
                adam = fleet.distributed_optimizer(adam)
                dp_layer = fleet.distributed_model(layer)
                state_dict = adam.state_dict()
672 673 674 675 676 677 678 679
        """
        # imitate target optimizer retrieval
        return self.user_defined_optimizer.state_dict()

    @dygraph_only
    def set_state_dict(self, state_dict):
        """
        Load optimizer state dict.
680
        (Only work in dygraph mode)
681 682 683 684

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

685 686
        Returns:
            None
687 688 689 690

        Examples:
            .. code-block:: python

691 692 693
                import numpy as np
                import paddle
                from paddle.distributed import fleet
694

695 696 697 698 699
                paddle.disable_static()
                fleet.init(is_collective=True)

                value = np.arange(26).reshape(2, 13).astype("float32")
                a = paddle.fluid.dygraph.to_variable(value)
700

701 702
                layer = paddle.nn.Linear(13, 5)
                adam = paddle.optimizer.Adam(learning_rate=0.01, parameters=layer.parameters())
703

704 705 706 707 708 709
                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)
710 711 712 713 714 715 716 717
        """
        # 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. 
718
        (Only work in dygraph mode)
719

720 721 722
        Args:
            value (float|Tensor): the value of learning rate

723 724
        Returns: 
            None 
725 726 727 728

        Examples:
            .. code-block:: python

729 730 731
                import numpy as np
                import paddle
                from paddle.distributed import fleet
732

733 734
                paddle.disable_static()
                fleet.init(is_collective=True)
735

736 737
                value = np.arange(26).reshape(2, 13).astype("float32")
                a = paddle.fluid.dygraph.to_variable(value)
738

739 740
                layer = paddle.nn.Linear(13, 5)
                adam = paddle.optimizer.Adam(learning_rate=0.01, parameters=layer.parameters())
741

742 743 744 745 746 747 748 749 750 751 752 753 754 755
                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
756 757 758 759 760 761 762 763
        """
        # imitate target optimizer retrieval
        return self.user_defined_optimizer.set_lr(value)

    @dygraph_only
    def get_lr(self):
        """
        Get current step learning rate.
764
        (Only work in dygraph mode)
765 766 767 768 769 770 771

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

        Examples:
            .. code-block:: python

772 773 774 775 776 777
                import numpy as np
                import paddle
                from paddle.distributed import fleet

                paddle.disable_static()
                fleet.init(is_collective=True)
778

779 780
                value = np.arange(26).reshape(2, 13).astype("float32")
                a = paddle.fluid.dygraph.to_variable(value)
781

782 783
                layer = paddle.nn.Linear(13, 5)
                adam = paddle.optimizer.Adam(learning_rate=0.01, parameters=layer.parameters())
784

785 786
                adam = fleet.distributed_optimizer(adam)
                dp_layer = fleet.distributed_model(layer)
787

788 789
                lr = adam.get_lr()
                print(lr) # 0.01
790 791 792 793 794 795 796 797
        """
        # imitate target optimizer retrieval
        return self.user_defined_optimizer.get_lr()

    @dygraph_only
    def step(self):
        """
        Execute the optimizer once.
798
        (Only work in dygraph mode)
799

800 801
        Returns:
            None
802 803 804 805

        Examples:
            .. code-block:: python

806 807 808
                import paddle
                import paddle.nn as nn
                from paddle.distributed import fleet
809

810 811 812 813 814
                class LinearNet(nn.Layer):
                    def __init__(self):
                        super(LinearNet, self).__init__()
                        self._linear1 = nn.Linear(10, 10)
                        self._linear2 = nn.Linear(10, 1)
815

816 817
                    def forward(self, x):
                        return self._linear2(self._linear1(x))
818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857

                # 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):
        """
858 859
        Clear the gradients of all optimized parameters for model.
        (Only work in dygraph mode)
860

861 862
        Returns: 
            None
863 864 865 866

        Examples:
            .. code-block:: python

867 868 869
                import paddle
                import paddle.nn as nn
                from paddle.distributed import fleet
870

871 872 873 874 875
                class LinearNet(nn.Layer):
                    def __init__(self):
                        super(LinearNet, self).__init__()
                        self._linear1 = nn.Linear(10, 10)
                        self._linear2 = nn.Linear(10, 1)
876

877 878
                    def forward(self, x):
                        return self._linear2(self._linear1(x))
879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914

                # 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()

D
Dong Daxiang 已提交
915 916 917 918 919 920 921 922 923
    def _final_strategy(self):
        if "valid_strategy" not in self._context:
            print(
                "WARNING: You may need to call minimize function before this function is called"
            )
            return {}
        else:
            return self._context["valid_strategy"]

924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946
    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.
947 948
            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
949 950 951
            ``fetch_list`` before run, see details in ``Executor``.

        Examples:
952
            .. code-block:: python
953

954 955
                import paddle
                import paddle.distributed.fleet as fleet
956

957 958 959 960 961 962 963 964 965 966 967 968
                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)
969

970
                # for more examples, please reference https://github.com/PaddlePaddle/FleetX
971 972

        """
D
Dong Daxiang 已提交
973 974 975
        context = {}
        context["user_defined_strategy"] = copy.deepcopy(
            self._user_defined_strategy)
976 977 978
        if paddle.fluid.framework.in_dygraph_mode():
            # imitate target optimizer retrieval
            target_opt = self.user_defined_optimizer
D
Dong Daxiang 已提交
979
            self._context = context
980 981
            return target_opt.minimize(loss)

982 983
        # cache original feed forward program
        self.origin_main_program = loss.block.program
984 985
        context["origin_main_program"] = self.origin_main_program
        context["loss"] = loss
986 987
        if startup_program == None:
            self.origin_startup_program = \
988 989
                paddle.static.default_startup_program().clone(for_test=False)
            startup_program = paddle.static.default_startup_program()
990 991 992
        else:
            self.origin_startup_program = \
                startup_program.clone(for_test=False)
993

994 995
        context["origin_startup_program"] = startup_program
        context["role_maker"] = self._role_maker
996 997 998 999 1000

        # compile time
        distributed_optimizer_list = \
            MetaOptimizerFactory()._get_valid_meta_optimizers(
                self.user_defined_optimizer)
D
Dong Daxiang 已提交
1001

D
Dong Daxiang 已提交
1002 1003 1004
        context["user_defined_strategy"] = copy.deepcopy(
            self._user_defined_strategy)
        copy_user_defined_strategy = copy.deepcopy(self._user_defined_strategy)
1005 1006 1007 1008 1009 1010

        # 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)
D
Dong Daxiang 已提交
1011
        if copy_user_defined_strategy._is_strict_auto():
1012 1013
            # turn on all the strategy for each optimizer
            for opt in distributed_optimizer_list:
D
Dong Daxiang 已提交
1014
                opt._enable_strategy(copy_user_defined_strategy, context)
1015

1016 1017
        valid_optimizer_list = []
        valid_graph_optimizer_list = []
D
Dong Daxiang 已提交
1018
        can_not_apply_optimizer_list = []
1019 1020 1021 1022
        # recall meta optimizers for ranking
        for opt in distributed_optimizer_list:
            opt._set_basic_info(loss, self._role_maker,
                                self.user_defined_optimizer,
D
Dong Daxiang 已提交
1023
                                copy_user_defined_strategy)
1024 1025
            if opt._can_apply() and not opt._is_graph_out():
                valid_optimizer_list.append(opt)
D
Dong Daxiang 已提交
1026
            elif opt._can_apply() and opt._is_graph_out():
1027
                valid_graph_optimizer_list.append(opt)
D
Dong Daxiang 已提交
1028 1029
            else:
                can_not_apply_optimizer_list.append(opt)
1030
        # combine recalled meta optimizers to be a valid meta optimizer
D
Dong Daxiang 已提交
1031
        meta_optimizer, graph_optimizer = \
1032 1033
            self.strategy_compiler.generate_optimizer(
                loss, self._role_maker, self.user_defined_optimizer,
D
Dong Daxiang 已提交
1034
                copy_user_defined_strategy, valid_optimizer_list,
1035
                valid_graph_optimizer_list)
D
Dong Daxiang 已提交
1036

D
Dong Daxiang 已提交
1037
        valid_strategy = self.strategy_compiler._get_valid_strategy(
D
Dong Daxiang 已提交
1038 1039 1040
            copy_user_defined_strategy, can_not_apply_optimizer_list)

        context["valid_strategy"] = copy.deepcopy(valid_strategy)
1041

D
Dong Daxiang 已提交
1042
        self._context = context
1043

D
Dong Daxiang 已提交
1044
        self.valid_strategy = valid_strategy
1045
        self.valid_strategy._enable_env()
D
Dong Daxiang 已提交
1046

1047 1048
        optimize_ops = []
        params_grads = []
1049

1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063
        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)

1064 1065 1066 1067 1068 1069
        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)
1070

1071
            default_program = paddle.static.default_main_program()
1072 1073 1074 1075

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

1076 1077 1078 1079 1080 1081
        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)
1082

1083 1084
        context["program_optimize_ops"] = optimize_ops
        context["program_params_grads"] = params_grads
1085

1086
        if graph_optimizer:
D
Dong Daxiang 已提交
1087
            optimize_ops, params_grads = graph_optimizer.minimize(
1088 1089 1090 1091 1092 1093 1094 1095
                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
1096 1097 1098
            context["graph_optimize_ops"] = optimize_ops
            context["graph_optimize_grads"] = params_grads

1099
        if self._runtime_handle is None:
1100
            self._runtime_handle = RuntimeFactory()._create_runtime(context)
1101

1102 1103
        import paddle.distributed.fleet as fleet
        fleet.util._set_strategy(context["valid_strategy"])
1104 1105

        return optimize_ops, params_grads