parallel.py 21.2 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except jin 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 os
15
import six
Y
Yan Xu 已提交
16
import numpy as np
17
from collections import OrderedDict
18
from .. import core
19
from . import layers
C
chengduo 已提交
20
from . import parallel_helper
21
from .. import framework
J
Jiabin Yang 已提交
22
from . import to_variable, no_grad
23

24
__all__ = ["prepare_context", "ParallelEnv", "DataParallel"]
25 26 27 28

ParallelStrategy = core.ParallelStrategy


C
chengduo 已提交
29
def prepare_context(strategy=None):
30 31 32
    '''
    :api_attr: imperative
    '''
C
chengduo 已提交
33 34 35 36 37 38 39 40
    if strategy is None:
        strategy = ParallelStrategy()
        strategy.nranks = Env().nranks
        strategy.local_rank = Env().local_rank
        strategy.trainer_endpoints = Env().trainer_endpoints
        strategy.current_endpoint = Env().current_endpoint
    if strategy.nranks < 2:
        return
41
    assert framework.in_dygraph_mode() is True, \
42
        "dygraph.prepare_context should be used with dygrahp mode."
43
    place = framework._current_expected_place()
C
chengduo 已提交
44
    assert place is not None, \
45
        "dygraph.prepare_context should be used in fluid.dygraph.guard(place) guard."
46
    if isinstance(place, core.CUDAPlace):
C
chengduo 已提交
47 48
        parallel_helper._set_parallel_ctx(
            core.NCCLParallelContext(strategy, place))
49 50 51
    else:
        # TODO(Yancey1989): add Gloo Parallel Context to support CPU parallel computation
        assert ("Only support CUDAPlace for now.")
C
chengduo 已提交
52 53
    parallel_helper._init_parallel_ctx()
    return strategy
54 55


56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113
class ParallelEnv(object):
    """
    **Notes**:
        **The old class name was Env and will be deprecated. Please use new class name ParallelEnv.**

    This class is used to obtain the environment variables required for 
    the parallel execution of dynamic graph model.

    The dynamic graph parallel mode needs to be started using paddle.distributed.launch.
    By default, the related environment variable is automatically configured by this module.

    This class is generally used in with `fluid.dygraph.DataParallel` to configure dynamic graph models
    to run in parallel.

    Examples:
      .. code-block:: python

        # This example needs to run with paddle.distributed.launch, The usage is:
        #   python -m paddle.distributed.launch --selected_gpus=0,1 example.py
        # And the content of `example.py` is the code of following example.

        import numpy as np
        import paddle.fluid as fluid
        import paddle.fluid.dygraph as dygraph
        from paddle.fluid.optimizer import AdamOptimizer
        from paddle.fluid.dygraph.nn import Linear
        from paddle.fluid.dygraph.base import to_variable

        place = fluid.CUDAPlace(fluid.dygraph.ParallelEnv().dev_id)
        with fluid.dygraph.guard(place=place):

            # prepare the data parallel context
            strategy=dygraph.prepare_context()

            linear = Linear(1, 10, act="softmax")
            adam = fluid.optimizer.AdamOptimizer()

            # make the module become the data parallelism module
            linear = dygraph.DataParallel(linear, strategy)

            x_data = np.random.random(size=[10, 1]).astype(np.float32)
            data = to_variable(x_data)

            hidden = linear(data)
            avg_loss = fluid.layers.mean(hidden)

            # scale the loss according to the number of trainers.
            avg_loss = linear.scale_loss(avg_loss)

            avg_loss.backward()

            # collect the gradients of trainers.
            linear.apply_collective_grads()

            adam.minimize(avg_loss)
            linear.clear_gradients()
    """

114 115 116 117 118 119 120 121 122 123
    def __init__(self):
        self._nranks = int(os.getenv("PADDLE_TRAINERS_NUM", "1"))
        self._local_rank = int(os.getenv("PADDLE_TRAINER_ID", "0"))
        self._dev_id = int(os.getenv("FLAGS_selected_gpus", "0"))
        self._trainer_endpoints = os.getenv("PADDLE_TRAINER_ENDPOINTS",
                                            "").split(",")
        self._current_endpoint = os.getenv("PADDLE_CURRENT_ENDPOINT", "")

    @property
    def nranks(self):
124 125 126 127 128 129 130 131 132 133 134 135 136 137 138
        """
        The number of trainers, generally refers to the number of GPU cards used in training.

        Its value is equal to the value of the environment variable PADDLE_TRAINERS_NUM. The default value is 1.

        Examples:
          .. code-block:: python

            # execute this command in terminal: export PADDLE_TRAINERS_NUM=4
            import paddle.fluid as fluid
            
            env = fluid.dygraph.ParallelEnv()
            print("The nranks is %d" % env.nranks)
            # The nranks is 4
        """
139 140 141 142
        return self._nranks

    @property
    def local_rank(self):
143 144 145 146 147 148 149 150 151 152 153 154 155 156 157
        """
        The current trainer number.

        Its value is equal to the value of the environment variable PADDLE_TRAINER_ID. The default value is 0.

        Examples:
          .. code-block:: python

            # execute this command in terminal: export PADDLE_TRAINER_ID=0
            import paddle.fluid as fluid
            
            env = fluid.dygraph.ParallelEnv()
            print("The local rank is %d" % env.local_rank)
            # The local rank is 0
        """
158 159 160 161
        return self._local_rank

    @property
    def dev_id(self):
162 163 164 165 166 167 168 169 170 171 172 173 174 175 176
        """
        The ID of selected GPU card for parallel training.

        Its value is equal to the value of the environment variable FLAGS_selected_gpus. The default value is 0.

        Examples:
          .. code-block:: python

            # execute this command in terminal: export FLAGS_selected_gpus=1
            import paddle.fluid as fluid
            
            env = fluid.dygraph.ParallelEnv()
            print("The device id are %d" % env.dev_id)
            # The device id are 1
        """
177 178 179 180
        return self._dev_id

    @property
    def current_endpoint(self):
181 182 183 184 185 186 187 188 189 190 191 192 193 194 195
        """
        The endpoint of current trainer, it is in the form of (node IP + port).

        Its value is equal to the value of the environment variable PADDLE_CURRENT_ENDPOINT. The default value is "".

        Examples:
          .. code-block:: python
            
            # execute this command in terminal: export PADDLE_CURRENT_ENDPOINT=127.0.0.1:6170
            import paddle.fluid as fluid
            
            env = fluid.dygraph.ParallelEnv()
            print("The current endpoint are %s" % env.current_endpoint)
            # The current endpoint are 127.0.0.1:6170
        """
196
        return self._current_endpoint
197 198 199

    @property
    def trainer_endpoints(self):
200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215
        """
        The endpoints of all trainer nodes in the task, 
        which are used to broadcast the NCCL ID when NCCL2 is initialized.

        Its value is equal to the value of the environment variable PADDLE_TRAINER_ENDPOINTS. The default value is "".

        Examples:
          .. code-block:: python

            # execute this command in terminal: export PADDLE_TRAINER_ENDPOINTS=127.0.0.1:6170,127.0.0.1:6171
            import paddle.fluid as fluid
            
            env = fluid.dygraph.ParallelEnv()
            print("The trainer endpoints are %s" % env.trainer_endpoints)
            # The trainer endpoints are ['127.0.0.1:6170', '127.0.0.1:6171']
        """
216 217 218
        return self._trainer_endpoints


219 220 221 222 223 224
# NOTE: [ Compatible ] Originally this class name is `Env`. The semantics of the old class names
# are inaccurate and may confuse users, so replace it with `ParallelEnv`, but to be compatible
# with the old examples, here still need to keep this name.
Env = ParallelEnv


225
class DataParallel(layers.Layer):
C
chengduo 已提交
226
    """
227
    Run the dygraph module with data parallelism.
C
chengduo 已提交
228

229
    Currently, DataParallel class only supports to run the dynamic graph
C
chengduo 已提交
230
    with multi-process. The usage is:
231
    `python -m paddle.distributed.launch --selected_gpus=0,1 dynamic_graph_test.py`.
C
chengduo 已提交
232 233
    And the content of `dynamic_graph_test.py` is the code of examples.

234 235 236 237 238 239 240 241
    Args:
        layers(Layer): The module that should be executed by data parallel.
        strategy(ParallelStrategy): The strategy of data parallelism, contains 
            environment configuration related to parallel execution.

    Returns:
        Layer: The data paralleled module.

C
chengduo 已提交
242 243 244 245 246 247 248
    Examples:
        .. code-block:: python

           import numpy as np
           import paddle.fluid as fluid
           import paddle.fluid.dygraph as dygraph
           from paddle.fluid.optimizer import AdamOptimizer
249
           from paddle.fluid.dygraph.nn import Linear
C
chengduo 已提交
250 251
           from paddle.fluid.dygraph.base import to_variable

252
           place = place = fluid.CUDAPlace(fluid.dygraph.ParallelEnv().dev_id)
C
chengduo 已提交
253 254 255
           with fluid.dygraph.guard(place=place):

               # prepare the data parallel context
256
               strategy=dygraph.prepare_context()
C
chengduo 已提交
257

258
               linear = Linear(1, 10, act="softmax")
C
chengduo 已提交
259 260 261
               adam = fluid.optimizer.AdamOptimizer()

               # make the module become the data parallelism module
262
               linear = dygraph.DataParallel(linear, strategy)
C
chengduo 已提交
263 264 265 266

               x_data = np.random.random(size=[10, 1]).astype(np.float32)
               data = to_variable(x_data)

267
               hidden = linear(data)
C
chengduo 已提交
268 269 270
               avg_loss = fluid.layers.mean(hidden)

               # scale the loss according to the number of trainers.
271
               avg_loss = linear.scale_loss(avg_loss)
C
chengduo 已提交
272 273 274 275

               avg_loss.backward()

               # collect the gradients of trainers.
276
               linear.apply_collective_grads()
C
chengduo 已提交
277 278

               adam.minimize(avg_loss)
279
               linear.clear_gradients()
C
chengduo 已提交
280 281
    """

Y
Yan Xu 已提交
282
    def __init__(self, layers, strategy):
283 284
        super(DataParallel,
              self).__init__(layers.full_name() + "_data_parallel")
C
chengduo 已提交
285

286
        self._layers = layers
Y
Yan Xu 已提交
287
        self._strategy = strategy
288 289

    def forward(self, *inputs, **kwargs):
Y
Yan Xu 已提交
290 291 292
        return self._layers(*inputs, **kwargs)

    def scale_loss(self, loss):
C
chengduo 已提交
293 294 295 296 297 298
        """
        Scale the loss. In data parallel mode, the loss should be scale with
        the number of trainers. If not in data parallel mode, return the loss
        directly.

        Args:
299
            loss(Variable): The loss of the current Model.
C
chengduo 已提交
300 301

        Returns:
302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333
            Variable: the scaled loss.

        Examples:
            .. code-block:: python

                import numpy as np
                import paddle.fluid as fluid
                import paddle.fluid.dygraph as dygraph
                from paddle.fluid.optimizer import AdamOptimizer
                from paddle.fluid.dygraph.nn import Linear
                from paddle.fluid.dygraph.base import to_variable

                place = place = fluid.CUDAPlace(fluid.dygraph.ParallelEnv().dev_id)
                with fluid.dygraph.guard(place=place):
                    strategy=dygraph.prepare_context()
                    linear = Linear(1, 10, act="softmax")
                    adam = fluid.optimizer.AdamOptimizer()
                    linear = dygraph.DataParallel(linear, strategy)

                    x_data = np.random.random(size=[10, 1]).astype(np.float32)
                    data = to_variable(x_data)
                    hidden = linear(data)
                    avg_loss = fluid.layers.mean(hidden)

                    # scale the loss according to the number of trainers.
                    avg_loss = linear.scale_loss(avg_loss)

                    avg_loss.backward()
                    linear.apply_collective_grads()

                    adam.minimize(avg_loss)
                    linear.clear_gradients()
C
chengduo 已提交
334 335
        """
        if not self._is_data_parallel_mode():
Y
Yan Xu 已提交
336
            return loss
C
chengduo 已提交
337

Y
Yan Xu 已提交
338 339 340 341 342 343
        loss_scale = to_variable(
            np.array([self._strategy.nranks]).astype("float32"))
        loss_scale.stop_gradient = True
        loss = loss / loss_scale
        return loss

344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359
    def _coalesce_tensors(self, var_groups):
        from ..layers import nn
        coalesced_grads_and_grad_vars = []
        for group_id, grad_vars in var_groups.items():
            flattened_vars = []
            g_var_shapes = []
            for g_var in grad_vars:
                g_var_shapes.append(g_var.shape)
                flattened_vars.append(
                    nn.reshape(
                        x=g_var, shape=[np.prod(g_var.shape)], inplace=True))
            coalesced_grad = nn.concat(flattened_vars)
            coalesced_grads_and_grad_vars.append(
                [coalesced_grad, grad_vars, g_var_shapes])
        return coalesced_grads_and_grad_vars

360 361 362 363 364 365 366 367 368
    def _reshape_inplace(self, x, shape):
        x_shape = self._helper.create_variable_for_type_inference(dtype=x.dtype)
        self._helper.append_op(
            type="reshape2",
            inputs={'X': x},
            attrs={'shape': shape},
            outputs={'Out': x,
                     'XShape': x_shape})

369 370 371 372
    def _split_tensors(self, coalesced_grads_and_grad_vars):
        from ..layers import nn
        for coalesced_grad, origin_grad_vars, grad_shapes in coalesced_grads_and_grad_vars:
            grad_var_len = [np.prod(g_shape) for g_shape in grad_shapes]
373 374 375 376 377 378 379
            self._helper.main_program.current_block().append_op(
                type='split',
                inputs={'X': coalesced_grad},
                outputs={'Out': origin_grad_vars},
                attrs={'sections': grad_var_len,
                       'axis': 0})
            for g_var, g_shape in zip(origin_grad_vars, grad_shapes):
380 381
                self._reshape_inplace(x=g_var, shape=g_shape)
                assert g_var.shape == g_shape
382

383
    @no_grad()
Y
Yan Xu 已提交
384
    def apply_collective_grads(self):
C
chengduo 已提交
385 386
        """
        AllReduce the Parameters' gradient.
387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416

        Examples:
            .. code-block:: python

                import numpy as np
                import paddle.fluid as fluid
                import paddle.fluid.dygraph as dygraph
                from paddle.fluid.optimizer import AdamOptimizer
                from paddle.fluid.dygraph.nn import Linear
                from paddle.fluid.dygraph.base import to_variable

                place = place = fluid.CUDAPlace(fluid.dygraph.ParallelEnv().dev_id)
                with fluid.dygraph.guard(place=place):
                    strategy=dygraph.prepare_context()
                    linear = Linear(1, 10, act="softmax")
                    adam = fluid.optimizer.AdamOptimizer()
                    linear = dygraph.DataParallel(linear, strategy)

                    x_data = np.random.random(size=[10, 1]).astype(np.float32)
                    data = to_variable(x_data)
                    hidden = linear(data)
                    avg_loss = fluid.layers.mean(hidden)
                    avg_loss = linear.scale_loss(avg_loss)
                    avg_loss.backward()

                    # collect the gradients of trainers.
                    linear.apply_collective_grads()

                    adam.minimize(avg_loss)
                    linear.clear_gradients()
C
chengduo 已提交
417 418
        """
        if not self._is_data_parallel_mode():
Y
Yan Xu 已提交
419 420
            return

421 422
        grad_var_set = set()
        grad_vars = []
423
        sparse_grad_vars = []
Y
Yan Xu 已提交
424
        for param in self._layers.parameters():
C
chengduo 已提交
425
            # NOTE(zcd): The grad_ivar maybe no generated.
426
            if param.trainable and (param._grad_ivar() is not None):
427
                g_var = param._grad_ivar()
428 429 430
                if g_var._is_sparse():
                    sparse_grad_vars.append(g_var)
                    continue
431 432 433 434
                grad_vars.append(g_var)
                assert g_var not in grad_var_set
                grad_var_set.add(g_var)

435 436 437 438 439
        if sparse_grad_vars:
            sparse_grad_vars.sort(key=lambda x: x.name)
            for grad_var in sparse_grad_vars:
                grad_var._allreduce(self._strategy)

440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460
        # FIXME(zcd): the type of the var should be LoDTensor, i.e
        # the gradients should be dense, otherwise, the following
        # logic should be updated.
        # 128 MB as a group
        mega_bytes = 128 * 1024 * 1024
        group_idx = 0
        memory_counter = 0
        grad_var_groups = OrderedDict()
        dtype = grad_vars[0].dtype
        for g_var in grad_vars:
            # Note: the dtype of the same group should be the same.
            bytes = np.prod(g_var.shape) * core.size_of_dtype(g_var.dtype)
            if memory_counter < mega_bytes and dtype == g_var.dtype:
                memory_counter += bytes
            else:
                memory_counter = bytes
                group_idx += 1
            grad_var_groups.setdefault(group_idx, []).append(g_var)

        coalesced_grads_and_vars = self._coalesce_tensors(grad_var_groups)

461 462
        for coalesced_grad, _, _ in coalesced_grads_and_vars:
            coalesced_grad._allreduce(self._strategy)
463 464

        self._split_tensors(coalesced_grads_and_vars)
C
chengduo 已提交
465 466 467

    def _is_data_parallel_mode(self):
        return self._strategy.nranks > 1
468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489

    def state_dict(self,
                   destination=None,
                   include_sublayers=True,
                   structured_name_prefix=""):
        '''
        Get all parameters of self._layers and its sub-layers. And set all the parameters into a dict

        Parameters:
            destination(dict, optional) : If provide, all the parameters will set to this dict . Default: None
            include_sublayers(bool, optional) : If true, also include the parameters from sublayers. Default: True
            structured_name_prefix(str, optional): If not empty str, all the key in state dict will start 
                                                 with structured_name_prefix

        Retruns:
            dict: a dict contains all the parameters of self._layers

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
                with fluid.dygraph.guard():
490
                    strategy=fluid.dygraph.prepare_context()
491
                    emb = fluid.dygraph.Embedding([10, 10])
492
                    emb = fluid.dygraph.DataParallel(emb, strategy)
493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523

                    state_dict = emb.state_dict()
                    fluid.save_dygraph( state_dict, "paddle_dy")

        '''

        return self._layers.state_dict(
            destination=destination,
            include_sublayers=include_sublayers,
            structured_name_prefix=structured_name_prefix)

    def set_dict(self,
                 stat_dict,
                 include_sublayers=True,
                 use_structured_name=True):
        '''
        Set parameters of self._layers from stat_dict. All the parameters of self._layers will be reset by the tensor in the stat_dict

        Parameters:
            state_dict(dict) : Dict contains all the parameters
            include_sublayers(bool, optional) : If true, also include the parameters from sublayers. Default: True
            use_structured_name(bool, optional) : If true, use structured name as key, otherwise, use parameter name as key. 
                                                  Default: True
        Returns:
            None

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
                with fluid.dygraph.guard():
524
                    strategy=fluid.dygraph.prepare_context()
525
                    emb = fluid.dygraph.Embedding([10, 10])
526
                    emb = fluid.dygraph.DataParallel(emb, strategy)
527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563

                    state_dict = emb.state_dict()
                    fluid.save_dygraph( state_dict, "paddle_dy")
                    
                    para_state_dict, _ = fluid.load_dygraph( "paddle_dy")

                    emb.set_dict( para_state_dict )

        '''

        self._layers.set_dict(
            stat_dict,
            include_sublayers=include_sublayers,
            use_structured_name=use_structured_name)

    def load_dict(self,
                  stat_dict,
                  include_sublayers=True,
                  use_structured_name=True):
        '''
        Set parameters of self._layers from stat_dict. All the parameters of self._layers will be reset by the tensor in the stat_dict

        This api will be Deprecated. Please use set_dict

        Parameters:
            state_dict(dict) : Dict contains all the parameters
            include_sublayers(bool, optional) : If true, also include the parameters from sublayers. Default: True
            use_structured_name(bool, optional) : If true, use structured name as key, otherwise, use parameter name as key.
                                                  Default: True
        Returns:
            None

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
                with fluid.dygraph.guard():
564
                    strategy=fluid.dygraph.prepare_context()
565
                    emb = fluid.dygraph.Embedding([10, 10])
566
                    emb = fluid.dygraph.DataParallel(emb, strategy)
567 568 569 570 571 572 573 574 575 576 577 578 579 580

                    state_dict = emb.state_dict()
                    fluid.save_dygraph( state_dict, "paddle_dy")
                    
                    para_state_dict, _ = fluid.load_dygraph( "paddle_dy")

                    emb.load_dict( para_state_dict )

        '''

        self._layers.load_dict(
            stat_dict,
            include_sublayers=include_sublayers,
            use_structured_name=use_structured_name)