optimizer.py 110.8 KB
Newer Older
1
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
D
dzhwinter 已提交
2
#
D
dzhwinter 已提交
3 4 5
# 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
D
dzhwinter 已提交
6
#
D
dzhwinter 已提交
7
#     http://www.apache.org/licenses/LICENSE-2.0
D
dzhwinter 已提交
8
#
D
dzhwinter 已提交
9 10 11 12 13
# 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.
14 15

from __future__ import print_function
16

17
import numpy as np
18
from collections import defaultdict
19

Q
Qiao Longfei 已提交
20
from paddle.fluid.distribute_lookup_table import find_distributed_lookup_table
21
from paddle.fluid.framework import Program, Variable, name_scope, default_main_program, default_startup_program
22

23 24
from . import framework
from . import layers
25
from . import unique_name
H
hutuxian 已提交
26
from .backward import append_backward, _some_in_set_, _append_grad_suffix_
27
from .clip import append_gradient_clip_ops, error_clip_callback
28 29 30
from .framework import program_guard
from .initializer import Constant
from .layer_helper import LayerHelper
S
sneaxiy 已提交
31
from .layers import ops
32
from .regularizer import append_regularization_ops
33 34 35 36 37
from .dygraph import base as imperative_base
from .dygraph.learning_rate_scheduler import LearningRateDecay
from paddle.fluid import core
from paddle.fluid.layers import tensor
from functools import reduce
38
from .wrapped_decorator import signature_safe_contextmanager
39

40
__all__ = [
Q
qiaolongfei 已提交
41
    'SGD', 'Momentum', 'Adagrad', 'Adam', 'Adamax', 'DecayedAdagrad', 'Ftrl',
42
    'SGDOptimizer', 'MomentumOptimizer', 'AdagradOptimizer', 'AdamOptimizer',
W
weixing02 已提交
43
    'AdamaxOptimizer', 'DecayedAdagradOptimizer', 'RMSPropOptimizer',
44
    'FtrlOptimizer', 'Adadelta', 'ModelAverage', 'LarsMomentum',
45
    'LarsMomentumOptimizer', 'DGCMomentumOptimizer', 'LambOptimizer',
H
hutuxian 已提交
46
    'ExponentialMovingAverage', 'PipelineOptimizer'
47
]
Q
Qiao Longfei 已提交
48 49 50 51 52 53


class Optimizer(object):
    """Optimizer Base class.

    Define the common interface of an optimizer.
54 55
    User should not use this class directly,
    but need to use one of it's implementation.
Q
Qiao Longfei 已提交
56 57
    """

58
    @imperative_base.no_grad
X
Xin Pan 已提交
59
    def __init__(self, learning_rate, regularization=None, name=None):
L
lujun 已提交
60
        if framework.in_dygraph_mode():
M
minqiyang 已提交
61 62 63 64 65
            if not isinstance(learning_rate, float) and \
                    not isinstance(learning_rate, LearningRateDecay):
                raise TypeError(
                    "learning rate should be float or LearningRateDecay, got %s here"
                    % type(learning_rate))
66 67 68 69
            if name is not None:
                self._name = unique_name.generate(name)
            else:
                self._name = unique_name.generate(self.__class__.__name__)
M
minqiyang 已提交
70 71 72 73 74 75
        else:
            if not isinstance(learning_rate, float) and \
                    not isinstance(learning_rate, framework.Variable):
                raise TypeError(
                    "learning rate should be float or Variable, got %s here" %
                    type(learning_rate))
76
            self._name = name
M
minqiyang 已提交
77

D
dzhwinter 已提交
78
        self.regularization = regularization
79
        self._learning_rate = learning_rate
D
dzhwinter 已提交
80 81
        # the learning rate type should be inferenced from loss
        self._dtype = None
82
        # each program should have a independent learning rate
83
        # program -> Variable(learning_rate)
Q
qiaolongfei 已提交
84
        self._learning_rate_map = dict()
85
        if isinstance(self._learning_rate, framework.Variable):
86 87
            self._learning_rate_map[framework.default_main_program(
            )] = self._learning_rate
88 89 90 91 92
        # Dictionary of accumulators. Some optimizer subclasses need to
        # allocate and manage extra variables associated with the parameters
        # to train. These variables are called accumulators.
        # {accum_name : { paramter_name : accumulator_for_parameter, ...}, ...}
        self._accumulators = defaultdict(lambda: dict())
Q
Qiao Longfei 已提交
93
        self.helper = None
94 95
        self._opti_name_list = []

96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179
    def load(self, stat_dict):
        """
        load optimizer with learning rate decay in dygraph mode
        :return: None

        Args:
            stat_dict: the dict load by load_persistable method

        Examples:

        .. code-block:: python

            from __future__ import print_function
            import numpy as np
            import paddle
            import paddle.fluid as fluid
            from paddle.fluid.optimizer import SGDOptimizer
            from paddle.fluid.dygraph.nn import FC
            from paddle.fluid.dygraph.base import to_variable

            class MLP(fluid.Layer):
                def __init__(self, name_scope):
                    super(MLP, self).__init__(name_scope)

                    self._fc1 = FC(self.full_name(), 10)
                    self._fc2 = FC(self.full_name(), 10)

                def forward(self, inputs):
                    y = self._fc1(inputs)
                    y = self._fc2(y)
                    return y

            with fluid.dygraph.guard():
                mlp = MLP('mlp')
                optimizer2 = SGDOptimizer(
                    learning_rate=fluid.layers.natural_exp_decay(
                    learning_rate=0.1,
                    decay_steps=10000,
                    decay_rate=0.5,
                    staircase=True))

                train_reader = paddle.batch(
                        paddle.dataset.mnist.train(), batch_size=128, drop_last=True)

                for batch_id, data in enumerate(train_reader()):
                    dy_x_data = np.array(
                            [x[0].reshape(1, 28, 28) for x in data]).astype('float32')

                    y_data = np.array([x[1] for x in data]).astype('int64').reshape(
                            128, 1)

                    img = to_variable(dy_x_data)
                    label = to_variable(y_data)
                    label._stop_gradient = True
                    cost = mlp(img)
                    avg_loss = fluid.layers.reduce_mean(cost)
                    avg_loss.backward()
                    optimizer.minimize(avg_loss)
                    mlp.clear_gradients()
                    fluid.dygraph.save_persistables(
                            mlp.state_dict(), [optimizer, optimizer2], "save_dir_2")
                    if batch_id == 2:
                            break

            with fluid.dygraph.guard():
                mlp_load = MLP('mlp')
                optimizer_load2 = SGDOptimizer(
                        learning_rate=fluid.layers.natural_exp_decay(
                        learning_rate=0.1,
                        decay_steps=10000,
                        decay_rate=0.5,
                        staircase=True))
                parameters, optimizers = fluid.dygraph.load_persistables(
                    "save_dir_2")
                mlp_load.load_dict(parameters)
                optimizer_load2.load(optimizers)
            self.assertTrue(optimizer2._learning_rate.__dict__ == optimizer_load2._learning_rate.__dict__)

        """
        if framework.in_dygraph_mode():
            self._learning_rate = stat_dict[self._name]
        else:
            raise TypeError("load can only be used under DyGraph mode")

180 181
    def get_opti_var_name_list(self):
        return self._opti_name_list
Q
Qiao Longfei 已提交
182

Q
Qiao Longfei 已提交
183
    def _create_global_learning_rate(self):
184 185 186
        if imperative_base.enabled():
            # create learning rate Variable
            if isinstance(self._learning_rate, float):
M
minqiyang 已提交
187 188 189 190 191 192 193 194 195 196 197 198
                lr = self._global_learning_rate()

                if isinstance(lr, framework.Variable):
                    return
                else:
                    self._learning_rate_map[framework.default_main_program(
                    )] = layers.create_global_var(
                        name=unique_name.generate("learning_rate"),
                        shape=[1],
                        value=float(self._learning_rate),
                        dtype='float32' if self._dtype is None else self._dtype,
                        persistable=True)
199
            # get learning rate Variable from LearningRateDecay
M
minqiyang 已提交
200
            elif isinstance(self._learning_rate, LearningRateDecay):
201 202 203
                self._learning_rate_map[framework.default_main_program(
                )] = self._learning_rate()
            else:
Q
qiaolongfei 已提交
204
                raise TypeError(
205 206
                    "optimizer's learning rate must be float or LearningRateDecay"
                )
207
        else:
208 209 210 211
            lr = self._global_learning_rate()

            if isinstance(lr, framework.Variable):
                return
M
minqiyang 已提交
212 213 214 215 216 217
            else:
                if not isinstance(self._learning_rate, float):
                    raise TypeError(
                        "learning rate variable is create outside optimizer,"
                        "can not create new learning rate variable for new program"
                    )
Q
Qiao Longfei 已提交
218

219 220 221 222 223 224 225 226
            # create learning rate in the current main program
            self._learning_rate_map[framework.default_main_program(
            )] = layers.create_global_var(
                name=unique_name.generate("learning_rate"),
                shape=[1],
                value=float(self._learning_rate),
                dtype='float32' if self._dtype is None else self._dtype,
                persistable=True)
227

Y
yuyang18 已提交
228
    def _global_learning_rate(self, program=None):
Q
Qiao Longfei 已提交
229 230 231 232
        """
        get global decayed learning rate
        :return:
        """
233 234
        if program is None:
            program = framework.default_main_program()
Q
qiaolongfei 已提交
235
        return self._learning_rate_map.get(program, None)
Q
Qiao Longfei 已提交
236

Q
Qiao Longfei 已提交
237 238 239 240 241
    def _append_optimize_op(self, block, param_and_grad):
        """ append optimize operator to block and return all the added optimize_op
        """
        raise NotImplementedError()

242 243 244 245
    def _create_param_lr(self, param_and_grad):
        # create learning rate variable for every parameter
        param = param_and_grad[0]
        param_lr = param.optimize_attr['learning_rate']
W
Wu Yi 已提交
246 247
        if type(param_lr) == Variable:
            return param_lr
Q
qiaolongfei 已提交
248
        else:
W
Wu Yi 已提交
249
            if param_lr == 1.0:
Y
yuyang18 已提交
250
                return self._global_learning_rate()
W
Wu Yi 已提交
251
            else:
X
Xin Pan 已提交
252 253 254
                with default_main_program()._lr_schedule_guard(
                        is_with_opt=True), framework.name_scope(
                            'scale_with_param_lr'):
255
                    return self._global_learning_rate() * param_lr
256 257 258 259 260 261 262

    def _create_accumulators(self, block, parameters):
        """Create all accumulators needed by the parameters

        Args:
            block: the block in which the loss variable is present
            parameters: list of parameter variables for the optimizer
Q
Qiao Longfei 已提交
263
        """
264 265
        pass

266
    def _finish_update(self, block, parameters_and_grads):
267 268 269 270 271 272 273 274
        """Finish any custom updates needed
           before completing an optimization step

        Args:
            block: the block in which the loss variable is present
            parameters: list of parameter variables for the optimizer

        Returns:
Q
qiaolongfei 已提交
275
            None
276 277 278
        """
        pass

279 280 281 282 283 284
    def _add_accumulator(self,
                         name,
                         param,
                         dtype=None,
                         fill_value=0.0,
                         shape=None):
285 286 287 288 289 290 291 292 293
        """Utility function to add an accumulator for a parameter

        Args:
            block: the block in which the loss variable is present
            name: name of the accumulator
            param: parameter variable for which accumulator is to be added
            dtype: data type of the accumulator variable
            fill_value: value to initialize the accumulator variable
        """
W
whs 已提交
294 295
        if self._name is not None:
            name = self._name + "_" + name
296 297
        if (name in self._accumulators and
                param.name in self._accumulators[name]):
L
lujun 已提交
298
            if framework.in_dygraph_mode():
X
polish  
Xin Pan 已提交
299
                return self._accumulators[name][param.name]
300
            raise Exception("Accumulator {} already exists for parameter {}".
301
                            format(name, param.name))
302 303
        if shape == None:
            shape = param.shape
Q
Qiao Longfei 已提交
304
        assert isinstance(self.helper, LayerHelper)
305 306 307 308 309

        var_name = param.name + "_" + name
        var_name = unique_name.generate(var_name)
        self._opti_name_list.append(var_name)

Q
Qiao Longfei 已提交
310
        var = self.helper.create_global_variable(
311
            name=var_name,
Q
Qiao Longfei 已提交
312
            persistable=True,
F
fengjiayi 已提交
313
            dtype=dtype or param.dtype,
Q
Qiao Longfei 已提交
314
            type=param.type,
315
            shape=shape)
Q
Qiao Longfei 已提交
316
        self.helper.set_variable_initializer(
317
            var, initializer=Constant(value=float(fill_value)))
Q
Qiao Longfei 已提交
318
        self._accumulators[name][param.name] = var
319
        return var
320 321 322 323 324 325 326 327 328 329 330

    def _get_accumulator(self, name, param):
        """Utility function to fetch an accumulator for a parameter

        Args:
            name: name of the accumulator
            param: parameter variable for which accumulator is to be fetched

        Returns:
            accumulator variable for the parameter
        """
W
whs 已提交
331 332
        if self._name is not None:
            name = self._name + "_" + name
333 334 335 336 337 338
        if (name not in self._accumulators or
                param.name not in self._accumulators[name]):
            raise Exception("Accumulator {} does not exist for parameter {}".
                            format(name, param.name))
        return self._accumulators[name][param.name]

339
    def _create_optimization_pass(self, parameters_and_grads):
Q
Qiao Longfei 已提交
340 341 342
        """Add optimization operators to update gradients to variables.

        Args:
Q
qiaolongfei 已提交
343
          parameters_and_grads(list(tuple(Variable, Variable))):
344
            a list of (variable, gradient) pair to update.
Q
Qiao Longfei 已提交
345 346

        Returns:
347
          return_op_list: a list of operators that will complete one step of
348 349 350
            optimization. This will include parameter update ops, global step
            update ops and any other custom ops required by subclasses to manage
            their internal state.
Q
Qiao Longfei 已提交
351
        """
352 353 354 355 356
        # This is a default implementation of create_optimization_pass that
        # can be shared by most optimizers. This implementation assumes that
        # the subclass will implement the _append_optimize_op method and the
        #  _initialize_tensors method. The subclass can extend the
        # _create_accumulators method if it needs to create accumulators
357
        # for parameters and extend _finish_update method to add custom ops.
358

359 360 361 362 363 364 365 366 367
        # Allways called under program_guard use global block as loss block
        global_block = framework.default_main_program().global_block()
        start = len(global_block.ops)
        self.helper = LayerHelper(self.__class__.__name__)
        self._create_accumulators(global_block,
                                  [p[0] for p in parameters_and_grads])
        self._create_global_learning_rate()

        optimize_ops = []
M
minqiyang 已提交
368
        if framework.in_dygraph_mode():
369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387
            for param_and_grad in parameters_and_grads:
                if param_and_grad[1] is None:
                    continue
                with param_and_grad[0].block.program._optimized_guard(
                        param_and_grad):
                    if param_and_grad[0].trainable is True:
                        optimize_op = self._append_optimize_op(global_block,
                                                               param_and_grad)
                        optimize_ops.append(optimize_op)
        else:
            for param_and_grad in parameters_and_grads:
                if param_and_grad[1] is None:
                    continue
                with param_and_grad[0].block.program._optimized_guard(
                        param_and_grad), name_scope("optimizer"):
                    if param_and_grad[0].trainable is True:
                        optimize_op = self._append_optimize_op(global_block,
                                                               param_and_grad)
                        optimize_ops.append(optimize_op)
388 389 390 391 392 393 394 395 396

        # Get custom finish ops for subclasses
        # FIXME: Need to fix this once we figure out how to handle dependencies
        self._finish_update(global_block, parameters_and_grads)

        end = len(global_block.ops)
        return global_block._slice_ops(start, end)

    def _process_distribute_lookuptable(self, param_grads):
Q
Qiao Longfei 已提交
397 398 399 400 401 402 403 404 405
        """
        Because distribute lookup table only support SGD optimizer for now, not support
        other optimizer and regularization, so we should find the table parameter out,
        and avoid to add regularization and other op for it, and add sgd optimize op
        for it independently.
        :param param_grads(list((Var, Var))): list of (param, grad) pair.
        :param loss: the loss variable.
        :param startup_program: the startup program
        """
406 407
        program = framework.default_main_program()
        global_block = framework.default_main_program().global_block()
Q
Qiao Longfei 已提交
408 409 410 411 412 413 414 415 416 417 418 419 420 421 422
        table_name = find_distributed_lookup_table(program)
        table_param = None
        table_grad = None
        new_param_grads = []
        for p, g in param_grads:
            if p.name == table_name:
                if table_param is not None:
                    raise RuntimeError(
                        "multi dist table var found, only support one now!")
                table_param = p
                table_grad = g
            else:
                new_param_grads.append((p, g))
        sgd_op = None
        if table_param is not None:
423 424 425 426 427 428 429 430 431 432 433 434 435
            param_and_grad = [table_param, table_grad]
            with table_param.block.program._optimized_guard(param_and_grad), \
                    framework.name_scope("optimizer"):
                self._create_global_learning_rate()
                # create the optimize op
                sgd_op = global_block.append_op(
                    type='sgd',
                    inputs={
                        "Param": table_param,
                        "Grad": table_grad,
                        "LearningRate": self._create_param_lr(param_and_grad)
                    },
                    outputs={"ParamOut": param_and_grad[0]})
Q
Qiao Longfei 已提交
436 437
        return new_param_grads, (table_param, table_grad), sgd_op

438 439 440
    def _append_dgc_ops(self, param_and_grad):
        pass

441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458
    def backward(self,
                 loss,
                 startup_program=None,
                 parameter_list=None,
                 no_grad_set=None,
                 callbacks=None):
        """
        First part of `minimize`, do auto-diff to append backward ops for
        the current program.

        Args:
            loss (Variable): loss variable to run optimizations.
            startup_program (Program): startup_program for initializing parameters
                in `parameter_list`.
            parameter_list (list): list of Variables to update.
            no_grad_set (set|None): set of Variables should be ignored.
            callbacks (list|None): list of callables to run when appending backward
                operator for one parameter.
M
minqiyang 已提交
459

460 461
        Return:
            list: list of (param, grad) pair, grad is the output of backward.
M
minqiyang 已提交
462

463 464 465
        Examples:
            See examples in `apply_gradients`.
        """
C
chengduo 已提交
466
        self._dtype = loss.dtype
L
lujun 已提交
467
        if framework.in_dygraph_mode():
C
chengduo 已提交
468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484
            if parameter_list is not None:
                parameters = parameter_list
            else:
                parameters = framework._dygraph_tracer().all_parameters()

            params_grads = []
            for param in parameters:
                if not param.trainable:
                    continue
                if param._ivar._grad_ivar() is not None:
                    # create gradient variable
                    grad_var = Variable(
                        block=loss.block,
                        name=param._ivar._grad_name(),
                        stop_gradient=True,
                        ivar=param._ivar._grad_ivar())
                    params_grads.append((param, grad_var))
485
        else:
C
chengduo 已提交
486 487 488 489 490 491 492 493 494 495 496 497
            if callbacks is None:
                callbacks = [error_clip_callback]
            else:
                assert (isinstance(callbacks, list))
            program = loss.block.program
            with program_guard(program, startup_program):
                params_grads = append_backward(loss, parameter_list,
                                               no_grad_set, callbacks)
                # Note: since we can't use all_reduce_op now,
                #  dgc_op should be the last op of one grad.
                self._append_dgc_ops(params_grads)
        return params_grads
498 499 500 501 502 503 504 505

    def apply_gradients(self, params_grads):
        """
        Second part of `minimize`, appending optimization operators for
        given `params_grads` pairs.

        Args:
            params_grads (list): list of (param, grad) pair to do optimization.
M
minqiyang 已提交
506

507 508
        Returns:
            list: A list of operators appended to the current program.
M
minqiyang 已提交
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
        Examples:
            .. code-block:: python

                loss = network()
                optimizer = fluid.optimizer.SGD(learning_rate=0.1)
                params_grads = optimizer.backward(loss)
                # you may append operations for params_grads here
                # ...
                optimizer.apply_gradients(params_grads)
        """
        params_grads = sorted(params_grads, key=lambda x: x[0].name)

        params_grads, table_param_and_grad, table_optimize_op = \
            self._process_distribute_lookuptable(params_grads)

        params_grads = append_gradient_clip_ops(params_grads)

        # Add regularization if any
        params_grads = append_regularization_ops(params_grads,
                                                 self.regularization)

        optimize_ops = self._create_optimization_pass(params_grads)
        if table_optimize_op is not None:
            optimize_ops.append(table_optimize_op)
            params_grads.append(table_param_and_grad)

        return optimize_ops

C
chengduo 已提交
538 539 540 541 542 543 544 545 546 547 548 549 550 551
    def apply_optimize(self, loss, startup_program, params_grads):
        """
        Second part of `minimize`, appending optimization operators for
        given `params_grads` pairs.

        Args:
            loss (Variable): loss variable to run optimizations.
            startup_program (Program): startup_program for initializing parameters
                in `parameter_list`.
            params_grads (list): list of (param, grad) pair to do optimization.

        Returns:
            list: A list of operators appended to the current program.
        """
L
lujun 已提交
552
        if framework.in_dygraph_mode():
C
chengduo 已提交
553 554
            with program_guard(framework.default_main_program(),
                               framework.default_startup_program()):
555 556
                params_grads = append_regularization_ops(params_grads,
                                                         self.regularization)
C
chengduo 已提交
557 558 559 560 561 562 563
                optimize_ops = self._create_optimization_pass(params_grads)
        else:
            program = loss.block.program
            with program_guard(program, startup_program):
                optimize_ops = self.apply_gradients(params_grads)
        return optimize_ops

564
    @imperative_base.no_grad
Q
Qiao Longfei 已提交
565 566
    def minimize(self,
                 loss,
567
                 startup_program=None,
Q
Qiao Longfei 已提交
568
                 parameter_list=None,
569 570
                 no_grad_set=None,
                 grad_clip=None):
571 572 573 574 575
        """
        Add operations to minimize `loss` by updating `parameter_list`.

        This method combines interface `backward()` and
        `apply_gradients()` into one.
M
minqiyang 已提交
576

577 578 579 580 581 582
        Args:
            loss (Variable): loss variable to run optimizations.
            startup_program (Program): startup_program for initializing parameters
                in `parameter_list`.
            parameter_list (list): list of Variables to update.
            no_grad_set (set|None): set of Variables should be ignored.
583
            grad_clip (GradClipBase|None) : Gradient clip strategy
Q
Qiao Longfei 已提交
584

585 586 587
        Returns:
            tuple: (optimize_ops, params_grads) which are, list of operators appended;
            and list of (param, grad) Variables pair for optimization.
Q
Qiao Longfei 已提交
588
        """
C
chengduo 已提交
589 590 591 592 593
        params_grads = self.backward(
            loss,
            startup_program=startup_program,
            parameter_list=parameter_list,
            no_grad_set=no_grad_set)
594 595 596 597 598

        if grad_clip is not None and framework.in_dygraph_mode():
            # TODO(hongyu): FIX later, this is only for dygraph, should be work for static mode
            params_grads = grad_clip(params_grads)

C
chengduo 已提交
599 600
        optimize_ops = self.apply_optimize(
            loss, startup_program=startup_program, params_grads=params_grads)
M
minqiyang 已提交
601

602 603 604
        if framework.in_dygraph_mode():
            framework._dygraph_tracer()._clear_ops()

Q
Qiao Longfei 已提交
605
        return optimize_ops, params_grads
Q
Qiao Longfei 已提交
606 607 608


class SGDOptimizer(Optimizer):
Q
qiaolongfei 已提交
609 610 611 612 613 614 615 616 617 618
    """
    Optimizer of the stochastic gradient descent algorithm.

    .. math::

        param\_out = param - learning\_rate * grad

    Args:
        learning_rate (float|Variable): the learning rate used to update parameters. \
        Can be a float value or a Variable with one float value as data element.
X
Xin Pan 已提交
619 620 621
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
Q
qiaolongfei 已提交
622 623 624 625

    Examples:
        .. code-block:: python

626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650
            import paddle
            import paddle.fluid as fluid
            import numpy as np

            place = fluid.CPUPlace()
            main = fluid.Program()
            with fluid.program_guard(main):
                x = fluid.layers.data(name='x', shape=[13], dtype='float32')
                y = fluid.layers.data(name='y', shape=[1], dtype='float32')
                y_predict = fluid.layers.fc(input=x, size=1, act=None)
                cost = fluid.layers.square_error_cost(input=y_predict, label=y)
                avg_cost = fluid.layers.mean(cost)

                sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001)
                sgd_optimizer.minimize(avg_cost)

                fetch_list = [avg_cost]
                train_reader = paddle.batch(
                    paddle.dataset.uci_housing.train(), batch_size=1)
                feeder = fluid.DataFeeder(place=place, feed_list=[x, y])
                exe = fluid.Executor(place)
                exe.run(fluid.default_startup_program())
                for data in train_reader():
                    exe.run(main, feed=feeder.feed(data), fetch_list=fetch_list)

Q
Qiao Longfei 已提交
651 652
    """

X
Xin Pan 已提交
653
    def __init__(self, learning_rate, regularization=None, name=None):
Q
Qiao Longfei 已提交
654
        assert learning_rate is not None
Q
Qiao Longfei 已提交
655
        super(SGDOptimizer, self).__init__(
X
Xin Pan 已提交
656 657 658
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
Q
Qiao Longfei 已提交
659 660
        self.type = "sgd"

661 662
    def _append_optimize_op(self, block, param_and_grad):
        assert isinstance(block, framework.Block)
663

Q
Qiao Longfei 已提交
664 665 666 667 668 669
        # create the optimize op
        sgd_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
670
                "LearningRate": self._create_param_lr(param_and_grad)
Q
Qiao Longfei 已提交
671
            },
M
minqiyang 已提交
672 673
            outputs={"ParamOut": param_and_grad[0]},
            stop_gradient=True)
Q
Qiao Longfei 已提交
674 675

        return sgd_op
676 677 678


class MomentumOptimizer(Optimizer):
Q
qiaolongfei 已提交
679 680 681 682 683 684 685 686 687 688 689 690 691 692
    """

    Simple Momentum optimizer with velocity state

    This optimizer has a flag for Nestrov Momentum.

    The update equations are as follows:

    .. math::

        & velocity = mu * velocity + gradient

        & if (use\_nesterov):

693
        &\quad   param = param - (gradient + mu * velocity) * learning\_rate
Q
qiaolongfei 已提交
694 695 696

        & else:

Q
qiaolongfei 已提交
697
        &\quad   param = param - learning\_rate * velocity
Q
qiaolongfei 已提交
698 699 700 701 702 703

    Args:
        learning_rate (float|Variable): the learning rate used to update parameters. \
        Can be a float value or a Variable with one float value as data element.
        momentum (float): momentum factor
        use_nesterov (bool): enables Nesterov momentum
X
Xin Pan 已提交
704 705 706
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
Q
qiaolongfei 已提交
707 708 709 710

    Examples:
        .. code-block:: python

711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735
            import paddle
            import paddle.fluid as fluid
            import numpy as np

            place = fluid.CPUPlace()
            main = fluid.Program()
            with fluid.program_guard(main):
                x = fluid.layers.data(name='x', shape=[13], dtype='float32')
                y = fluid.layers.data(name='y', shape=[1], dtype='float32')
                y_predict = fluid.layers.fc(input=x, size=1, act=None)
                cost = fluid.layers.square_error_cost(input=y_predict, label=y)
                avg_cost = fluid.layers.mean(cost)

                moment_optimizer = fluid.optimizer.MomentumOptimizer(learning_rate=0.001, momentum=0.9)
                moment_optimizer.minimize(avg_cost)

                fetch_list = [avg_cost]
                train_reader = paddle.batch(
                    paddle.dataset.uci_housing.train(), batch_size=1)
                feeder = fluid.DataFeeder(place=place, feed_list=[x, y])
                exe = fluid.Executor(place)
                exe.run(fluid.default_startup_program())
                for data in train_reader():
                    exe.run(main, feed=feeder.feed(data), fetch_list=fetch_list)

736 737 738
    """
    _velocity_acc_str = "velocity"

X
Xin Pan 已提交
739 740 741 742 743 744
    def __init__(self,
                 learning_rate,
                 momentum,
                 use_nesterov=False,
                 regularization=None,
                 name=None):
745 746
        assert learning_rate is not None
        assert momentum is not None
Q
Qiao Longfei 已提交
747
        super(MomentumOptimizer, self).__init__(
X
Xin Pan 已提交
748 749 750
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
751 752
        self.type = "momentum"
        self._momentum = momentum
753
        self._use_nesterov = bool(use_nesterov)
754 755 756 757 758

    def _create_accumulators(self, block, parameters):
        assert isinstance(block, framework.Block)

        for p in parameters:
Q
Qiao Longfei 已提交
759
            self._add_accumulator(self._velocity_acc_str, p)
760 761 762 763 764 765 766 767 768 769 770 771 772

    def _append_optimize_op(self, block, param_and_grad):
        assert isinstance(block, framework.Block)

        velocity_acc = self._get_accumulator(self._velocity_acc_str,
                                             param_and_grad[0])
        # create the momentum optimize op
        momentum_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
                "Velocity": velocity_acc,
773
                "LearningRate": self._create_param_lr(param_and_grad)
774 775 776 777 778
            },
            outputs={
                "ParamOut": param_and_grad[0],
                "VelocityOut": velocity_acc
            },
779
            attrs={"mu": self._momentum,
M
minqiyang 已提交
780 781
                   "use_nesterov": self._use_nesterov},
            stop_gradient=True)
782 783

        return momentum_op
784 785


786 787 788 789 790
class DGCMomentumOptimizer(MomentumOptimizer):
    """

    Original paper is https://arxiv.org/abs/1712.01887

G
gongweibao 已提交
791
    DGC reduces the communication bandwidth by sending only the important gradients (sparse update):\
792 793
        only gradients larger than a threshold are transmitted.

G
gongweibao 已提交
794
    To avoid losing information, DGC accumulates the rest of the gradients locally.
795 796 797

    Eventually, these gradients become large enough to be transmitted.

G
gongweibao 已提交
798
    Thus, DGC sends the large gradients immediately but eventually send all of the gradients over time.
799

G
gongweibao 已提交
800
    To ensure no loss of accuracy, DGC employs momentum correction and local gradient clipping on top of the gradient sparsification to maintain model performance.
801 802 803 804

    DGC also uses momentum factor masking and warmup training to overcome the staleness problem caused by reduced communication.

    This optimizer will do two things:
805

806 807
        1. Compress the gradient by get TopK import value from tensor \
            and use it for allreduce to reduce network bandwidth.
808

809 810 811 812 813 814
        2. Call momentum to optimize on the cost.

    Args:
        learning_rate (float|Variable): the learning rate used to update parameters. \
            Can be a float value or a Variable with one float value as data element.
        momentum (float): Momentum factor.
G
gongweibao 已提交
815
        rampup_begin_step (int): The beginning step from which gradient compression is implemented.
816 817 818 819 820 821 822
        rampup_step (int): How long it use the sparsity periods. Default is 1.
            for example: If the sparsity is [0.75, 0.9375, 0.984375, 0.996, 0.999], and the rampup_step is 5, \
                it will use 0.75 at 0 step, and 0.9375 at 1 step, and so on. And when reach sparsity array ends, \
                it will use 0.999 then and after.
        sparsity (list[float]): Get top important element from gradient tensor, the ratio is (1 - current sparsity).
        use_nesterov (bool): Enables Nesterov momentum. True means use nesterov.
        local_grad_clip_norm (float): Clip norm value if needed.
G
gongweibao 已提交
823
        num_trainers: The number of training nodes.
824
        regularization: A Regularizer, such as fluid.regularizer.L2DecayRegularizer.
G
gongweibao 已提交
825
        name: An optional name prefix.
826 827 828 829 830

    Examples:
        .. code-block:: python

            optimizer = fluid.optimizer.DGCMomentumOptimizer(
G
gongweibao 已提交
831 832 833 834 835
                        learning_rate=0.0001,
                        momentum=0.9,
                        rampup_step=1000,
                        rampup_begin_step=1252,
                        sparsity=[0.999, 0.999])
836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901

    """

    def __init__(self,
                 learning_rate,
                 momentum,
                 rampup_begin_step,
                 rampup_step=1,
                 sparsity=[0.999],
                 use_nesterov=False,
                 local_grad_clip_norm=None,
                 num_trainers=None,
                 regularization=None,
                 name=None):
        self._sparsity = sparsity
        self._rampup_step = rampup_step
        self._rampup_step_var = None

        self._rampup_begin_step = rampup_begin_step
        self._rampup_begin_step_var = None

        self._global_step_var = None
        self._local_grad_clip_norm = None
        self._clip_norm = None

        if local_grad_clip_norm is not None:
            assert isinstance(num_trainers, int)
            assert isinstance(local_grad_clip_norm, float)
            assert num_trainers > 0

            self._local_grad_clip_norm = local_grad_clip_norm
            self._num_trainers = num_trainers
            self._clip_norm = local_grad_clip_norm / (num_trainers *
                                                      num_trainers)

        super(DGCMomentumOptimizer, self).__init__(
            learning_rate, momentum, use_nesterov, regularization, name)

        core.init_dgc()

    def _add_auto_increment_var(self, counter_name, begin, step=1):
        helper = LayerHelper('global_step_counter')
        counter, is_new_var = helper.create_or_get_global_variable(
            name=counter_name, dtype='float32', shape=[1], persistable=True)
        if is_new_var:
            helper.set_variable_initializer(
                counter,
                initializer=Constant(
                    value=float(begin - 1), force_cpu=True))
            helper.main_program.global_block()._prepend_op(
                type='increment',
                inputs={'X': [counter]},
                outputs={'Out': [counter]},
                attrs={'step': float(step)},
                stop_gradient=True)
            counter.stop_gradient = True

        return counter

    def _append_dgc_ops(self, param_and_grads):
        start_program = default_startup_program()
        main_program = default_main_program()
        main_program._enable_dgc = True

        # step counter
        self._global_step_var = self._add_auto_increment_var(
G
gongweibao 已提交
902
            counter_name=core.dgc.kDGCCounterName(), begin=0)
903 904 905 906 907 908

        # rampup begin step var for all_reduce_op_handle
        self._rampup_begin_step_var = tensor.create_global_var(
            shape=[1],
            dtype=core.VarDesc.VarType.FP32,
            persistable=True,
G
gongweibao 已提交
909
            name=core.dgc.kDGCRampUpBeginStepName(),
910 911 912 913
            value=self._rampup_begin_step * 1.0,
            force_cpu=True)

        for param_var, grad_var in param_and_grads:
G
gongweibao 已提交
914
            var_numel = abs(reduce(lambda x, y: x * y, param_var.shape))
915 916 917 918 919 920 921 922 923 924
            if var_numel < 16384 or \
                param_var.type == core.VarDesc.VarType.SELECTED_ROWS  or \
                grad_var.type == core.VarDesc.VarType.SELECTED_ROWS  or  \
                    param_var.dtype != core.VarDesc.VarType.FP32 :
                continue

            u_var = tensor.create_global_var(
                shape=param_var.shape,
                dtype=param_var.dtype,
                persistable=True,
G
gongweibao 已提交
925
                name=param_var.name + core.dgc.kDGCUName(),
926 927 928 929 930
                value=0.0)
            v_var = tensor.create_global_var(
                shape=param_var.shape,
                dtype=param_var.dtype,
                persistable=True,
G
gongweibao 已提交
931
                name=param_var.name + core.dgc.kDGCVName(),
932 933 934 935 936 937
                value=0.0)

            k_var = tensor.create_global_var(
                shape=[1],
                dtype=param_var.dtype,
                persistable=True,
G
gongweibao 已提交
938
                name=param_var.name + core.dgc.kDGCKName(),
939 940 941 942 943 944 945
                value=0.0,
                force_cpu=True)

            encoded_var = tensor.create_global_var(
                shape=[1],
                dtype=param_var.dtype,
                persistable=True,
G
gongweibao 已提交
946
                name=param_var.name + core.dgc.kDGCEncodedName(),
947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987
                value=0.0,
                force_cpu=False)

            # del back oprolevarname
            op_maker = core.op_proto_and_checker_maker
            backward = core.op_proto_and_checker_maker.OpRole.Backward
            for op in main_program.global_block().ops:
                if not self._is_the_backward_op(op):
                    continue

                var_attr = op.all_attrs()[op_maker.kOpRoleVarAttrName()]
                if param_var.name not in var_attr:
                    continue

                var_attr.remove(param_var.name)
                var_attr.remove(grad_var.name)
                if len(var_attr) > 1:
                    op._set_attr(op_maker.kOpRoleVarAttrName(), var_attr)
                else:
                    op._remove_attr(op_maker.kOpRoleVarAttrName())

            clip_var = grad_var
            if self._local_grad_clip_norm is not None:
                clip_var = self._append_clip_norm(grad_var, self._clip_norm)
            self._dgc_op(param_var, clip_var, grad_var, u_var, v_var, k_var,
                         encoded_var)

    def _is_the_backward_op(self, op):
        op_maker = core.op_proto_and_checker_maker
        backward = core.op_proto_and_checker_maker.OpRole.Backward
        if op_maker.kOpRoleVarAttrName() in op.attr_names and \
                int(op.all_attrs()[op_maker.kOpRoleAttrName()]) == int(backward):
            return True
        return False

    def _clip_by_norm(self, x, max_norm, name=None):
        args = {'x': x, 'max_norm': max_norm, 'name': name}

        helper = LayerHelper("dgc_clip_by_norm_op", **args)

        if name is None:
988 989
            name = unique_name.generate_with_ignorable_key(".".join(
                [helper.name, 'tmp']))
990 991 992 993 994

        out = helper.create_variable(
            type=x.type, name=name, dtype=x.dtype, persistable=False)

        helper.append_op(
G
gongweibao 已提交
995
            type="dgc_clip_by_norm",
996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007
            inputs={"X": x,
                    "current_step": self._global_step_var},
            attrs={
                "max_norm": max_norm,
                "rampup_begin_step": float(self._rampup_begin_step)
            },
            outputs={"Out": out})
        return out

    def _append_clip_norm(self, grad_var, clip_norm):
        with grad_var.block.program._backward_role_guard():
            return self._clip_by_norm(
G
gongweibao 已提交
1008
                x=grad_var, max_norm=clip_norm, name=grad_var.name)
1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043

    def _dgc_op(self, param_var, clip_var, grad_var, u_var, v_var, k_var,
                encoded_var):
        block = framework.default_main_program().global_block()
        op_maker = core.op_proto_and_checker_maker
        dgc_op = block.append_op(
            type="dgc",
            inputs={
                "U": u_var,
                "V": v_var,
                "Grad": clip_var,
                "current_step": self._global_step_var
            },
            outputs={
                "U_out": u_var,
                "V_out": v_var,
                "EncodeGrad": encoded_var,
                "k": k_var,
                "Grad_out": grad_var
            },
            attrs={
                "m": self._momentum,
                "sparsity": self._sparsity,
                "use_nesterov": self._use_nesterov,
                "rampup_begin_step": float(self._rampup_begin_step),
                "rampup_step": float(self._rampup_step)
            },
            stop_gradient=True)

        backward = op_maker.OpRole.Backward
        dgc_op._set_attr(op_maker.kOpRoleAttrName(), backward)
        dgc_op._set_attr(op_maker.kOpRoleVarAttrName(),
                         [param_var.name, grad_var.name])


1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067
class LarsMomentumOptimizer(Optimizer):
    """
    Momentum optimizer with LARS support

    The update equations are as follows:

    .. math::

        & local\_learning\_rate = learning\_rate * lars\_coeff * \\
          \\frac{||param||}{||gradient|| + lars\_weight\_decay * ||param||}

        & velocity = mu * velocity + local\_learning\_rate * (gradient + lars\_weight\_decay * param)

        & param = param - velocity

    Args:
        learning_rate (float|Variable): the learning rate used to update parameters. \
        Can be a float value or a Variable with one float value as data element.
        momentum (float): momentum factor
        lars_coeff (float): defines how much we trust the layer to change its weights.
        lars_weight_decay (float): weight decay coefficient for decaying using LARS.
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
M
minqiyang 已提交
1068

1069 1070 1071 1072

    Examples:
        .. code-block:: python

1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088
            import paddle.fluid as fluid
            import numpy as np

            np_inp = np.array([[1.0, 2.0], [3.0, 4.0]], dtype=np.float32)
            inp = fluid.layers.data(
                name="inp", shape=[2, 2], append_batch_size=False)
            out = fluid.layers.fc(inp, size=3)
            out = fluid.layers.reduce_sum(out)
            optimizer = fluid.optimizer.LarsMomentumOptimizer(learning_rate=0.001, momentum=0.9)
            optimizer.minimize(out)

            exe = fluid.Executor(fluid.CPUPlace())
            exe.run(fluid.default_startup_program())
            exe.run(
                feed={"inp": np_inp},
                fetch_list=[out.name])
1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137
    """
    _velocity_acc_str = "velocity"

    def __init__(self,
                 learning_rate,
                 momentum,
                 lars_coeff=0.001,
                 lars_weight_decay=0.0005,
                 regularization=None,
                 name=None):
        assert learning_rate is not None
        assert momentum is not None
        super(LarsMomentumOptimizer, self).__init__(
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
        self.type = "lars_momentum"
        self._momentum = momentum
        self._lars_coeff = float(lars_coeff)
        self._lars_weight_decay = float(lars_weight_decay)

    def _create_accumulators(self, block, parameters):
        assert isinstance(block, framework.Block)

        for p in parameters:
            self._add_accumulator(self._velocity_acc_str, p)

    def _append_optimize_op(self, block, param_and_grad):
        assert isinstance(block, framework.Block)

        velocity_acc = self._get_accumulator(self._velocity_acc_str,
                                             param_and_grad[0])
        # create the momentum optimize op
        momentum_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
                "Velocity": velocity_acc,
                "LearningRate": self._create_param_lr(param_and_grad)
            },
            outputs={
                "ParamOut": param_and_grad[0],
                "VelocityOut": velocity_acc
            },
            attrs={
                "mu": self._momentum,
                "lars_coeff": self._lars_coeff,
                "lars_weight_decay": self._lars_weight_decay
M
minqiyang 已提交
1138 1139
            },
            stop_gradient=True)
1140 1141 1142 1143

        return momentum_op


1144
class AdagradOptimizer(Optimizer):
Q
qiaolongfei 已提交
1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164
    """
    **Adaptive Gradient Algorithm (Adagrad)**

    The update is done as follows:

    .. math::

        moment\_out &= moment + grad * grad

        param\_out &= param - \\frac{learning\_rate * grad}{\sqrt{moment\_out} + \epsilon}

    The original paper(http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf)
    does not have the epsilon attribute. It is added here in our implementation
    as also proposed here: http://cs231n.github.io/neural-networks-3/#ada
    for numerical stability to avoid the division by zero error.

    Args:
        learning_rate (float|Variable): the learning rate used to update parameters. \
        Can be a float value or a Variable with one float value as data element.
        epsilon (float): a small float value for numerical stability.
X
Xin Pan 已提交
1165 1166 1167
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
X
xuezhong 已提交
1168
        initial_accumulator_value (float): Initial value for moment accumulator.
Q
qiaolongfei 已提交
1169 1170 1171 1172

    Examples:
        .. code-block:: python

1173 1174 1175 1176 1177 1178 1179 1180
            import paddle.fluid as fluid
            import numpy as np

            np_inp = np.array([[1.0, 2.0], [3.0, 4.0]], dtype=np.float32)
            inp = fluid.layers.data(
                name="inp", shape=[2, 2], append_batch_size=False)
            out = fluid.layers.fc(inp, size=3)
            out = fluid.layers.reduce_sum(out)
Q
qiaolongfei 已提交
1181
            optimizer = fluid.optimizer.Adagrad(learning_rate=0.2)
1182 1183 1184 1185 1186 1187 1188
            optimizer.minimize(out)

            exe = fluid.Executor(fluid.CPUPlace())
            exe.run(fluid.default_startup_program())
            exe.run(
                feed={"inp": np_inp},
                fetch_list=[out.name])
1189 1190 1191
    """
    _moment_acc_str = "moment"

X
Xin Pan 已提交
1192 1193 1194 1195
    def __init__(self,
                 learning_rate,
                 epsilon=1.0e-6,
                 regularization=None,
1196
                 name=None,
X
xuezhong 已提交
1197
                 initial_accumulator_value=0.0):
1198 1199
        assert learning_rate is not None
        assert epsilon is not None
Q
Qiao Longfei 已提交
1200
        super(AdagradOptimizer, self).__init__(
X
Xin Pan 已提交
1201 1202 1203
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
1204 1205
        self.type = "adagrad"
        self._epsilon = epsilon
1206
        self.initial_accumulator_value = initial_accumulator_value
1207 1208 1209 1210 1211

    def _create_accumulators(self, block, parameters):
        assert isinstance(block, framework.Block)

        for p in parameters:
Q
Qiao Longfei 已提交
1212
            self._add_accumulator(self._moment_acc_str, p)
1213 1214 1215 1216 1217 1218

    def _append_optimize_op(self, block, param_and_grad):
        assert isinstance(block, framework.Block)

        moment_acc = self._get_accumulator(self._moment_acc_str,
                                           param_and_grad[0])
1219 1220 1221 1222 1223 1224 1225 1226 1227 1228
        startup_block = framework.default_startup_program().global_block()
        startup_block.append_op(
            type='fill_constant',
            inputs={},
            outputs={'Out': [moment_acc]},
            attrs={
                'dtype': moment_acc.dtype,
                'value': self.initial_accumulator_value,
                'shape': moment_acc.shape,
            })
1229

1230
        # Create the adagrad optimizer op
1231 1232 1233 1234 1235 1236
        adagrad_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
                "Moment": moment_acc,
1237
                "LearningRate": self._create_param_lr(param_and_grad)
1238 1239 1240
            },
            outputs={"ParamOut": param_and_grad[0],
                     "MomentOut": moment_acc},
M
minqiyang 已提交
1241 1242
            attrs={"epsilon": self._epsilon},
            stop_gradient=True)
1243 1244

        return adagrad_op
1245 1246 1247


class AdamOptimizer(Optimizer):
Q
qiaolongfei 已提交
1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274
    """
    This implements the Adam optimizer from Section 2 of the Adam
    paper : https://arxiv.org/abs/1412.6980.
    Adam is a first-order gradient-based optimization method based on
    adaptive estimates of lower-order moments.

    Adam updates:

    .. math::

        t & = t + 1

        moment\_1\_out & = {\\beta}_1 * moment\_1 + (1 - {\\beta}_1) * grad

        moment\_2\_out & = {\\beta}_2 * moment\_2 + (1 - {\\beta}_2) * grad * grad

        learning\_rate & = learning\_rate * \\
                          \\frac{\sqrt{1 - {\\beta}_2^t}}{1 - {\\beta}_1^t}

        param\_out & = param - learning\_rate * \\frac{moment\_1}{\sqrt{moment\_2} + \epsilon}

    Args:
        learning_rate (float|Variable): the learning rate used to update parameters. \
        Can be a float value or a Variable with one float value as data element.
        beta1 (float): The exponential decay rate for the 1st moment estimates.
        beta2 (float): The exponential decay rate for the 2nd moment estimates.
        epsilon (float): a small float value for numerical stability.
1275
        regularization: A Regularizer, such as fluid.regularizer.L2DecayRegularizer.
X
Xin Pan 已提交
1276
        name: A optional name prefix.
1277 1278 1279 1280 1281 1282
        lazy_mode(bool: false): The official Adam algorithm has two moving-average accumulators
        the accumulators are updated at every step. Every element of the two moving-average is updated
        in both dense mode and sparse mode. If the size of parameter is very large, then the update
        may be very slow. The lazy mode only update the element that has gradient is the current
        mini-batch, so it will be much more faster. But this mode has different semantics with the
        original Adam algorithm and may lead to different result.
Q
qiaolongfei 已提交
1283 1284 1285 1286

    Examples:
        .. code-block:: python

1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309
            import paddle
            import paddle.fluid as fluid

            place = fluid.CPUPlace()
            main = fluid.Program()
            with fluid.program_guard(main):
                x = fluid.layers.data(name='x', shape=[13], dtype='float32')
                y = fluid.layers.data(name='y', shape=[1], dtype='float32')
                y_predict = fluid.layers.fc(input=x, size=1, act=None)
                cost = fluid.layers.square_error_cost(input=y_predict, label=y)
                avg_cost = fluid.layers.mean(cost)

                adam_optimizer = fluid.optimizer.AdamOptimizer(0.01)
                adam_optimizer.minimize(avg_cost)

                fetch_list = [avg_cost]
                train_reader = paddle.batch(
                    paddle.dataset.uci_housing.train(), batch_size=1)
                feeder = fluid.DataFeeder(place=place, feed_list=[x, y])
                exe = fluid.Executor(place)
                exe.run(fluid.default_startup_program())
                for data in train_reader():
                    exe.run(main, feed=feeder.feed(data), fetch_list=fetch_list)
Q
qiaolongfei 已提交
1310

1311 1312 1313
    """
    _moment1_acc_str = "moment1"
    _moment2_acc_str = "moment2"
Q
qiaolongfei 已提交
1314 1315
    _beta1_pow_acc_str = "beta1_pow_acc"
    _beta2_pow_acc_str = "beta2_pow_acc"
1316 1317 1318 1319 1320

    def __init__(self,
                 learning_rate=0.001,
                 beta1=0.9,
                 beta2=0.999,
1321
                 epsilon=1e-8,
X
Xin Pan 已提交
1322
                 regularization=None,
Q
Qiao Longfei 已提交
1323
                 name=None,
Q
Qiao Longfei 已提交
1324
                 lazy_mode=False):
1325 1326 1327 1328
        assert learning_rate is not None
        assert beta1 is not None
        assert beta2 is not None
        assert epsilon is not None
Q
Qiao Longfei 已提交
1329
        super(AdamOptimizer, self).__init__(
X
Xin Pan 已提交
1330 1331 1332
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
1333 1334 1335 1336
        self.type = "adam"
        self._beta1 = beta1
        self._beta2 = beta2
        self._epsilon = epsilon
Q
Qiao Longfei 已提交
1337
        self._lazy_mode = lazy_mode
1338 1339 1340 1341 1342 1343

    def _create_accumulators(self, block, parameters):
        assert isinstance(block, framework.Block)

        # Create accumulator tensors for first and second moments
        for p in parameters:
Q
Qiao Longfei 已提交
1344 1345
            self._add_accumulator(self._moment1_acc_str, p)
            self._add_accumulator(self._moment2_acc_str, p)
Q
qiaolongfei 已提交
1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357
            self._add_accumulator(
                name=self._beta1_pow_acc_str,
                param=p,
                dtype='float32',
                fill_value=self._beta1,
                shape=[1])
            self._add_accumulator(
                name=self._beta2_pow_acc_str,
                param=p,
                dtype='float32',
                fill_value=self._beta2,
                shape=[1])
1358 1359 1360 1361 1362 1363 1364 1365

    def _append_optimize_op(self, block, param_and_grad):
        assert isinstance(block, framework.Block)

        moment1 = self._get_accumulator(self._moment1_acc_str,
                                        param_and_grad[0])
        moment2 = self._get_accumulator(self._moment2_acc_str,
                                        param_and_grad[0])
Q
qiaolongfei 已提交
1366 1367 1368 1369 1370
        beta1_pow_acc = self._get_accumulator(self._beta1_pow_acc_str,
                                              param_and_grad[0])
        beta2_pow_acc = self._get_accumulator(self._beta2_pow_acc_str,
                                              param_and_grad[0])

1371
        # create the adam optimize op
1372 1373 1374 1375 1376
        adam_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
1377
                "LearningRate": self._create_param_lr(param_and_grad),
1378 1379
                "Moment1": moment1,
                "Moment2": moment2,
Q
qiaolongfei 已提交
1380 1381
                "Beta1Pow": beta1_pow_acc,
                "Beta2Pow": beta2_pow_acc
1382 1383 1384 1385 1386 1387 1388 1389 1390
            },
            outputs={
                "ParamOut": param_and_grad[0],
                "Moment1Out": moment1,
                "Moment2Out": moment2
            },
            attrs={
                "beta1": self._beta1,
                "beta2": self._beta2,
Q
Qiao Longfei 已提交
1391
                "epsilon": self._epsilon,
1392 1393
                "lazy_mode": self._lazy_mode,
                "min_row_size_to_use_multithread": 1000
M
minqiyang 已提交
1394 1395
            },
            stop_gradient=True)
1396 1397 1398

        return adam_op

1399
    def _finish_update(self, block, param_and_grads):
1400 1401 1402
        """Update Beta1 and Beta2 Power accumulators
        """
        assert isinstance(block, framework.Block)
Q
Qiao Longfei 已提交
1403
        main_block = block.program.global_block()
1404 1405 1406
        for param, grad in param_and_grads:
            if grad is None:
                continue
X
Xin Pan 已提交
1407 1408
            with param.block.program._optimized_guard(
                [param, grad]), name_scope("optimizer"):
1409 1410 1411 1412 1413 1414 1415 1416
                beta1_pow_acc = self._get_accumulator(self._beta1_pow_acc_str,
                                                      param)
                beta2_pow_acc = self._get_accumulator(self._beta2_pow_acc_str,
                                                      param)
                main_block.append_op(
                    type="scale",
                    inputs={"X": beta1_pow_acc},
                    outputs={"Out": beta1_pow_acc},
M
minqiyang 已提交
1417 1418
                    attrs={"scale": self._beta1},
                    stop_gradient=True)
1419 1420 1421 1422 1423

                main_block.append_op(
                    type="scale",
                    inputs={"X": beta2_pow_acc},
                    outputs={"Out": beta2_pow_acc},
M
minqiyang 已提交
1424 1425
                    attrs={"scale": self._beta2},
                    stop_gradient=True)
1426 1427 1428


class AdamaxOptimizer(Optimizer):
Q
qiaolongfei 已提交
1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452
    """
    We implement the Adamax optimizer from Section 7 of the Adam
    paper: https://arxiv.org/abs/1412.6980. Adamax is a variant of the
    Adam algorithm based on the infinity norm.

    Adamax updates:

    .. math::

        t & = t + 1

        moment\_out & = {\\beta}_1 * moment + (1 - {\\beta}_1) * grad

        inf\_norm\_out & = max({\\beta}_2 * inf\_norm + \epsilon, |grad|)

        learning\_rate & = \\frac{learning\_rate}{1 - {\\beta}_1^t}

        param\_out & = param - learning\_rate * \\frac{moment\_out}{inf\_norm\_out}


    The original paper does not have an epsilon attribute.
    However, it is added here for numerical stability to prevent the
    division by 0 error.

1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479
    Examples:
        .. code-block:: python

          import paddle.fluid as fluid
          import numpy

          # First create the Executor.
          place = fluid.CPUPlace() # fluid.CUDAPlace(0)
          exe = fluid.Executor(place)

          train_program = fluid.Program()
          startup_program = fluid.Program()
          with fluid.program_guard(train_program, startup_program):
              data = fluid.layers.data(name='X', shape=[1], dtype='float32')
              hidden = fluid.layers.fc(input=data, size=10)
              loss = fluid.layers.mean(hidden)
              adam = fluid.optimizer.Adamax(learning_rate=0.2)
              adam.minimize(loss)

          # Run the startup program once and only once.
          exe.run(startup_program)

          x = numpy.random.random(size=(10, 1)).astype('float32')
          outs = exe.run(program=train_program,
                        feed={'X': x},
                         fetch_list=[loss.name])

Q
qiaolongfei 已提交
1480 1481 1482 1483 1484 1485
    Args:
        learning_rate (float|Variable): the learning rate used to update parameters. \
        Can be a float value or a Variable with one float value as data element.
        beta1 (float): The exponential decay rate for the 1st moment estimates.
        beta2 (float): The exponential decay rate for the 2nd moment estimates.
        epsilon (float): a small float value for numerical stability.
X
Xin Pan 已提交
1486 1487 1488
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
Q
qiaolongfei 已提交
1489

C
chengduo 已提交
1490 1491
    Notes:
       Currently, AdamaxOptimizer doesn't support sparse parameter optimization.
1492 1493 1494
    """
    _moment_acc_str = "moment"
    _inf_norm_acc_str = "inf_norm"
Q
qiaolongfei 已提交
1495
    _beta1_pow_acc_str = "beta1_pow_acc"
1496 1497 1498 1499 1500

    def __init__(self,
                 learning_rate=0.001,
                 beta1=0.9,
                 beta2=0.999,
1501
                 epsilon=1e-8,
X
Xin Pan 已提交
1502 1503
                 regularization=None,
                 name=None):
1504 1505 1506 1507
        assert learning_rate is not None
        assert beta1 is not None
        assert beta2 is not None
        assert epsilon is not None
Q
Qiao Longfei 已提交
1508
        super(AdamaxOptimizer, self).__init__(
X
Xin Pan 已提交
1509 1510 1511
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
1512 1513 1514 1515 1516 1517 1518 1519
        self.type = "adamax"
        self._beta1 = beta1
        self._beta2 = beta2
        self._epsilon = epsilon

    def _create_accumulators(self, block, parameters):
        # Create accumulator tensors for first moment and infinity norm
        for p in parameters:
Q
Qiao Longfei 已提交
1520 1521
            self._add_accumulator(self._moment_acc_str, p)
            self._add_accumulator(self._inf_norm_acc_str, p)
Q
qiaolongfei 已提交
1522 1523 1524 1525 1526 1527
            self._add_accumulator(
                name=self._beta1_pow_acc_str,
                param=p,
                dtype='float32',
                fill_value=self._beta1,
                shape=[1])
1528 1529 1530 1531 1532 1533 1534

    def _append_optimize_op(self, block, param_and_grad):
        assert isinstance(block, framework.Block)

        moment = self._get_accumulator(self._moment_acc_str, param_and_grad[0])
        inf_norm = self._get_accumulator(self._inf_norm_acc_str,
                                         param_and_grad[0])
Q
qiaolongfei 已提交
1535 1536
        beta1_pow_acc = self._get_accumulator(self._beta1_pow_acc_str,
                                              param_and_grad[0])
1537 1538 1539 1540 1541 1542
        # create the adamax optimize op
        adamax_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
1543
                "LearningRate": self._create_param_lr(param_and_grad),
1544 1545
                "Moment": moment,
                "InfNorm": inf_norm,
Q
qiaolongfei 已提交
1546
                "Beta1Pow": beta1_pow_acc
1547 1548 1549 1550 1551 1552 1553 1554 1555 1556
            },
            outputs={
                "ParamOut": param_and_grad[0],
                "MomentOut": moment,
                "InfNormOut": inf_norm
            },
            attrs={
                "beta1": self._beta1,
                "beta2": self._beta2,
                "epsilon": self._epsilon
M
minqiyang 已提交
1557 1558
            },
            stop_gradient=True)
1559 1560 1561

        return adamax_op

1562
    def _finish_update(self, block, parameters_and_grads):
1563 1564 1565
        """Update Beta1 Power accumulator
        """
        assert isinstance(block, framework.Block)
Q
Qiao Longfei 已提交
1566
        main_block = block.program.global_block()
1567 1568 1569
        for param, grad in parameters_and_grads:
            if grad is None:
                continue
X
Xin Pan 已提交
1570 1571
            with param.block.program._optimized_guard(
                [param, grad]), name_scope('adamx'):
1572 1573 1574 1575 1576 1577
                beta1_pow_acc = self._get_accumulator(self._beta1_pow_acc_str,
                                                      param)
                main_block.append_op(
                    type="scale",
                    inputs={"X": beta1_pow_acc},
                    outputs={"Out": beta1_pow_acc},
M
minqiyang 已提交
1578 1579
                    attrs={"scale": self._beta1},
                    stop_gradient=True)
1580 1581 1582


class DecayedAdagradOptimizer(Optimizer):
1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604
    """
    **Decayed Adagrad Optimizer**

    The original paper(http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf)

    The update is done as follows:

    .. math::

        moment\_out & = decay * moment + (1 - decay) * grad * grad

        param\_out & = param - \\frac{learning\_rate * grad}{\sqrt{moment\_out} + \epsilon}

    The original paper(http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf)
    does not have an epsilon attribute. It is added here for numerical
    stability to avoid the division by zero error.

    Args:
        learning_rate (float|Variable): the learning rate used to update parameters. \
        Can be a float value or a Variable with one float value as data element.
        decay (float): decay rate.
        epsilon (float): a small float value for numerical stability.
X
Xin Pan 已提交
1605 1606 1607
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
1608 1609 1610 1611

    Examples:
        .. code-block:: python

1612 1613 1614 1615 1616 1617 1618
            import paddle.fluid as fluid
            import paddle.fluid.layers as layers
            from paddle.fluid.optimizer import DecayedAdagrad

            x = layers.data( name='x', shape=[-1, 10], dtype='float32' )
            trans = layers.fc( x, 100 )
            cost = layers.reduce_mean( trans )
1619 1620
            optimizer = fluid.optimizer.DecayedAdagrad(learning_rate=0.2)
            optimizer.minimize(cost)
C
chengduo 已提交
1621 1622 1623

    Notes:
       Currently, DecayedAdagradOptimizer doesn't support sparse parameter optimization.
1624 1625 1626
    """
    _moment_acc_str = "moment"

X
Xin Pan 已提交
1627 1628 1629 1630 1631 1632
    def __init__(self,
                 learning_rate,
                 decay=0.95,
                 epsilon=1.0e-6,
                 regularization=None,
                 name=None):
1633 1634 1635 1636
        assert learning_rate is not None
        assert decay is not None
        assert epsilon is not None

Q
Qiao Longfei 已提交
1637
        super(DecayedAdagradOptimizer, self).__init__(
X
Xin Pan 已提交
1638 1639 1640
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667
        self.type = "decayed_adagrad"
        self._decay = decay
        self._epsilon = epsilon

    def _create_accumulators(self, block, parameters):
        assert isinstance(block, framework.Block)

        for p in parameters:
            self._add_accumulator(self._moment_acc_str, p)

    def _append_optimize_op(self, block, param_and_grad):
        assert isinstance(block, framework.Block)

        moment_acc = self._get_accumulator(self._moment_acc_str,
                                           param_and_grad[0])

        # Create the decayed adagrad optimizer op
        decayed_adagrad_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
                "Moment": moment_acc,
                "LearningRate": self._create_param_lr(param_and_grad)
            },
            outputs={"ParamOut": param_and_grad[0],
                     "MomentOut": moment_acc},
M
minqiyang 已提交
1668 1669
            attrs={"epsilon": self._epsilon},
            stop_gradient=True)
1670 1671

        return decayed_adagrad_op
1672 1673


1674
class AdadeltaOptimizer(Optimizer):
1675 1676
    """
    **Adadelta Optimizer**
Q
qiaolongfei 已提交
1677

1678
    Simple Adadelta optimizer with average squared grad state and
1679
    average squared update state.
1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691
    The details of adadelta please refer to this
    `ADADELTA: AN ADAPTIVE LEARNING RATE METHOD
    <http://www.matthewzeiler.com/pubs/googleTR2012/googleTR2012.pdf>`_.

    ..  math::

        E(g_t^2) &= \\rho * E(g_{t-1}^2) + (1-\\rho) * g^2 \\\\
        learning\\_rate &= sqrt( ( E(dx_{t-1}^2) + \\epsilon ) / ( \\
                          E(g_t^2) + \\epsilon ) ) \\\\
        E(dx_t^2) &= \\rho * E(dx_{t-1}^2) + (1-\\rho) * (-g*learning\\_rate)^2

    Args:
Q
qiaolongfei 已提交
1692
        learning_rate(float): global learning rate
1693 1694
        rho(float): rho in equation
        epsilon(float): epsilon in equation
X
Xin Pan 已提交
1695 1696 1697
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
1698 1699 1700 1701 1702 1703 1704

    Examples:
        .. code-block:: python

            optimizer = fluid.optimizer.Adadelta(
                learning_rate=0.0003, epsilon=1.0e-6, rho=0.95)
            _, params_grads = optimizer.minimize(cost)
C
chengduo 已提交
1705 1706 1707

    Notes:
       Currently, AdadeltaOptimizer doesn't support sparse parameter optimization.
1708
    """
1709

1710 1711 1712
    _avg_squared_grad_acc_str = "_avg_squared_grad"
    _avg_squared_update_acc_str = "_avg_squared_update"

X
Xin Pan 已提交
1713 1714 1715 1716 1717 1718
    def __init__(self,
                 learning_rate,
                 epsilon=1.0e-6,
                 rho=0.95,
                 regularization=None,
                 name=None):
1719 1720 1721 1722 1723 1724
        if learning_rate is None:
            raise ValueError("learning_rate is not set.")
        if epsilon is None:
            raise ValueError("epsilon is not set.")
        if rho is None:
            raise ValueError("rho is not set.")
1725
        super(AdadeltaOptimizer, self).__init__(
X
Xin Pan 已提交
1726 1727 1728
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
1729 1730 1731 1732 1733
        self.type = "adadelta"
        self._epsilon = epsilon
        self._rho = rho

    def _create_accumulators(self, block, parameters):
1734 1735
        if not isinstance(block, framework.Block):
            raise TypeError("block is not instance of framework.Block.")
1736 1737 1738 1739 1740 1741

        for p in parameters:
            self._add_accumulator(self._avg_squared_grad_acc_str, p)
            self._add_accumulator(self._avg_squared_update_acc_str, p)

    def _append_optimize_op(self, block, param_and_grad):
1742 1743
        if not isinstance(block, framework.Block):
            raise TypeError("block is not instance of framework.Block.")
1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764

        avg_squared_grad_acc = self._get_accumulator(
            self._avg_squared_grad_acc_str, param_and_grad[0])
        avg_squared_update_acc = self._get_accumulator(
            self._avg_squared_update_acc_str, param_and_grad[0])

        # Create the adadelta optimizer op
        adadelta_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
                "AvgSquaredGrad": avg_squared_grad_acc,
                "AvgSquaredUpdate": avg_squared_update_acc
            },
            outputs={
                "ParamOut": param_and_grad[0],
                "AvgSquaredGradOut": avg_squared_grad_acc,
                "AvgSquaredUpdateOut": avg_squared_update_acc
            },
            attrs={"epsilon": self._epsilon,
M
minqiyang 已提交
1765 1766
                   "rho": self._rho},
            stop_gradient=True)
1767 1768 1769 1770

        return adadelta_op


Q
qingqing01 已提交
1771 1772 1773 1774 1775 1776 1777 1778 1779 1780
class RMSPropOptimizer(Optimizer):
    """
    Root Mean Squared Propagation (RMSProp) is an unpublished, adaptive learning
    rate method. The original slides proposed RMSProp: Slide 29 of
    http://www.cs.toronto.edu/~tijmen/csc321/slides/lecture_slides_lec6.pdf .

    The original equation is as follows:

    ..  math::

Q
qiaolongfei 已提交
1781
        r(w, t) & = \\rho r(w, t-1) + (1 - \\rho)(\\nabla Q_{i}(w))^2
Q
qingqing01 已提交
1782 1783 1784 1785

        w & = w - \\frac{\\eta} {\\sqrt{r(w,t) + \\epsilon}} \\nabla Q_{i}(w)

    The first equation calculates moving average of the squared gradient for
Q
qiaolongfei 已提交
1786
    each weight. Then dividing the gradient by :math:`sqrt{v(w,t)}`.
Q
qingqing01 已提交
1787 1788 1789 1790 1791 1792

    In some cases, adding a momentum term :math: `\\beta` is beneficial.
    In our implementation, Nesterov momentum is used:

    ..  math::

Q
qiaolongfei 已提交
1793
        r(w, t) & = \\rho r(w, t-1) + (1 - \\rho)(\\nabla Q_{i}(w))^2
Q
qingqing01 已提交
1794

1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808
        v(w, t) & = \\beta v(w, t-1) + \\frac{\\eta} {\\sqrt{r(w,t) +
            \\epsilon}} \\nabla Q_{i}(w)

        w & = w - v(w, t)

    if centered is True:

    ..  math::

        r(w, t) & = \\rho r(w, t-1) + (1 - \\rho)(\\nabla Q_{i}(w))^2

        g(w, t) & = \\rho g(w, t-1) + (1 - \\rho)\\nabla Q_{i}(w)

        v(w, t) & = \\beta v(w, t-1) + \\frac{\\eta} {\\sqrt{r(w,t) - (g(w, t))^2 +
Q
qingqing01 已提交
1809 1810 1811 1812
            \\epsilon}} \\nabla Q_{i}(w)

        w & = w - v(w, t)

Q
qiaolongfei 已提交
1813
    where, :math:`\\rho` is a hyperparameter and typical values are 0.9, 0.95
Q
qingqing01 已提交
1814 1815 1816 1817 1818 1819
    and so on. :math: `beta` is the momentum term. :math: `\\epsilon` is a
    smoothing term to avoid division by zero, usually set somewhere in range
    from 1e-4 to 1e-8.


    Args:
Q
qiaolongfei 已提交
1820
        learning_rate(float): global learning rate.
Q
qingqing01 已提交
1821 1822 1823
        rho(float): rho is :math: `\\rho` in equation, set 0.95 by default.
        epsilon(float): :math: `\\epsilon` in equation is smoothing term to
            avoid division by zero, set 1e-6 by default.
Q
qiaolongfei 已提交
1824
        momentum(float): :math:`\\beta` in equation is the momentum term,
Q
qingqing01 已提交
1825
            set 0.0 by default.
1826 1827 1828 1829
        centered(bool): If True, gradients are normalized by the estimated variance of
            the gradient; if False, by the uncentered second moment. Setting this to
            True may help with training, but is slightly more expensive in terms of
            computation and memory. Defaults to False.
X
Xin Pan 已提交
1830 1831 1832
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
Q
qingqing01 已提交
1833 1834 1835 1836 1837 1838 1839

    Raises:
        ValueError: If learning_rate, rho, epsilon, momentum are None.

    Examples:
          .. code-block:: python

1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864
            import paddle
            import paddle.fluid as fluid
            import numpy as np

            place = fluid.CPUPlace()
            main = fluid.Program()
            with fluid.program_guard(main):
                x = fluid.layers.data(name='x', shape=[13], dtype='float32')
                y = fluid.layers.data(name='y', shape=[1], dtype='float32')
                y_predict = fluid.layers.fc(input=x, size=1, act=None)
                cost = fluid.layers.square_error_cost(input=y_predict, label=y)
                avg_cost = fluid.layers.mean(cost)

                rms_optimizer = fluid.optimizer.RMSProp(learning_rate=0.1)
                rms_optimizer.minimize(avg_cost)

                fetch_list = [avg_cost]
                train_reader = paddle.batch(
                    paddle.dataset.uci_housing.train(), batch_size=1)
                feeder = fluid.DataFeeder(place=place, feed_list=[x, y])
                exe = fluid.Executor(place)
                exe.run(fluid.default_startup_program())
                for data in train_reader():
                    exe.run(main, feed=feeder.feed(data), fetch_list=fetch_list)

Q
qingqing01 已提交
1865 1866 1867 1868
    """

    _momentum_acc_str = "momentum"
    _mean_square_acc_str = "mean_square"
1869
    _mean_grad_acc_str = "mean_grad"
Q
qingqing01 已提交
1870 1871 1872 1873 1874 1875

    def __init__(self,
                 learning_rate,
                 rho=0.95,
                 epsilon=1.0e-6,
                 momentum=0.0,
1876
                 centered=False,
X
Xin Pan 已提交
1877 1878
                 regularization=None,
                 name=None):
Q
qingqing01 已提交
1879
        super(RMSPropOptimizer, self).__init__(
X
Xin Pan 已提交
1880 1881 1882
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
Q
qingqing01 已提交
1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895
        if learning_rate is None:
            raise ValueError("learning_rate is not set.")
        if rho is None:
            raise ValueError("rho is not set.")
        if epsilon is None:
            raise ValueError("epsilon is not set.")
        if momentum is None:
            raise ValueError("momentum is not set.")

        self.type = "rmsprop"
        self._rho = rho
        self._epsilon = epsilon
        self._momentum = momentum
1896
        self._centered = centered
Q
qingqing01 已提交
1897 1898 1899 1900 1901 1902 1903 1904

    def _create_accumulators(self, block, parameters):
        if not isinstance(block, framework.Block):
            raise TypeError("block is not instance of framework.Block.")

        for p in parameters:
            self._add_accumulator(self._momentum_acc_str, p)
            self._add_accumulator(self._mean_square_acc_str, p)
1905
            self._add_accumulator(self._mean_grad_acc_str, p)
Q
qingqing01 已提交
1906 1907 1908 1909 1910 1911 1912 1913 1914

    def _append_optimize_op(self, block, param_and_grad):
        if not isinstance(block, framework.Block):
            raise TypeError("block is not instance of framework.Block.")

        momentum_acc = self._get_accumulator(self._momentum_acc_str,
                                             param_and_grad[0])
        mean_square_acc = self._get_accumulator(self._mean_square_acc_str,
                                                param_and_grad[0])
1915 1916
        mean_grad_acc = self._get_accumulator(self._mean_grad_acc_str,
                                              param_and_grad[0])
Q
qingqing01 已提交
1917 1918 1919 1920 1921 1922 1923
        rmsprop_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
                "Moment": momentum_acc,
                "MeanSquare": mean_square_acc,
1924
                "MeanGrad": mean_grad_acc,
Q
qingqing01 已提交
1925 1926 1927 1928 1929
                "LearningRate": self._create_param_lr(param_and_grad),
            },
            outputs={
                "ParamOut": param_and_grad[0],
                "MomentOut": momentum_acc,
1930 1931
                "MeanSquareOut": mean_square_acc,
                "MeanGradOut": mean_grad_acc
Q
qingqing01 已提交
1932 1933 1934 1935
            },
            attrs={
                "epsilon": self._epsilon,
                "decay": self._rho,
1936 1937
                "momentum": self._momentum,
                "centered": self._centered
M
minqiyang 已提交
1938 1939
            },
            stop_gradient=True)
Q
qingqing01 已提交
1940 1941 1942 1943

        return rmsprop_op


Q
qiaolongfei 已提交
1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985
class FtrlOptimizer(Optimizer):
    """
    FTRL (Follow The Regularized Leader) Optimizer.

    The paper that proposed Follow The Regularized Leader (FTRL):
    (https://www.eecs.tufts.edu/~dsculley/papers/ad-click-prediction.pdf)

    ..  math::

        &new\_accum = squared\_accum + grad^2

        &if (lr\_power == -0.5):

        &\quad  linear\_accum += grad - \\frac{\\sqrt{new\_accum} - \\sqrt{squared\_accum}}{learning\_rate * param}

        &else:

        &\quad   linear\_accum += grad - \\frac{new\_accum^{-lr\_power} - accum^{-lr\_power}}{learning\_rate * param}


        &x = l1 * sign(linear\_accum) - linear\_accum

        &if (lr\_power == -0.5):

        &\quad   y = \\frac{\\sqrt{new\_accum}}{learning\_rate} + (2 * l2)

        &\quad   pre\_shrink = \\frac{x}{y}

        &\quad   param = (abs(linear\_accum) > l1).select(pre\_shrink, 0.0)

        &else:

        &\quad   y = \\frac{new\_accum^{-lr\_power}}{learning\_rate} + (2 * l2)

        &\quad   pre\_shrink = \\frac{x}{y}

        &\quad   param = (abs(linear\_accum) > l1).select(pre\_shrink, 0.0)

        &squared\_accum += grad^2

    Args:
        learning_rate (float|Variable): global learning rate.
M
minqiyang 已提交
1986 1987 1988
        l1 (float): L1 regularization strength.
        l2 (float): L2 regularization strength.
        lr_power (float): Learning Rate Power.
X
Xin Pan 已提交
1989 1990 1991
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
Q
qiaolongfei 已提交
1992 1993 1994 1995 1996 1997 1998

    Raises:
        ValueError: If learning_rate, rho, epsilon, momentum are None.

    Examples:
          .. code-block:: python

1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022
            import paddle
            import paddle.fluid as fluid
            import numpy as np

            place = fluid.CPUPlace()
            main = fluid.Program()
            with fluid.program_guard(main):
                x = fluid.layers.data(name='x', shape=[13], dtype='float32')
                y = fluid.layers.data(name='y', shape=[1], dtype='float32')
                y_predict = fluid.layers.fc(input=x, size=1, act=None)
                cost = fluid.layers.square_error_cost(input=y_predict, label=y)
                avg_cost = fluid.layers.mean(cost)

                ftrl_optimizer = fluid.optimizer.Ftrl(learning_rate=0.1)
                ftrl_optimizer.minimize(avg_cost)

                fetch_list = [avg_cost]
                train_reader = paddle.batch(
                    paddle.dataset.uci_housing.train(), batch_size=1)
                feeder = fluid.DataFeeder(place=place, feed_list=[x, y])
                exe = fluid.Executor(place)
                exe.run(fluid.default_startup_program())
                for data in train_reader():
                    exe.run(main, feed=feeder.feed(data), fetch_list=fetch_list)
C
chengduo 已提交
2023 2024 2025

    Notes:
       Currently, FtrlOptimizer doesn't support sparse parameter optimization.
Q
qiaolongfei 已提交
2026 2027 2028 2029 2030
    """

    _squared_acc_str = "squared"
    _linear_acc_str = "linear"

X
Xin Pan 已提交
2031 2032 2033 2034 2035 2036 2037
    def __init__(self,
                 learning_rate,
                 l1=0.0,
                 l2=0.0,
                 lr_power=-0.5,
                 regularization=None,
                 name=None):
Q
qiaolongfei 已提交
2038
        super(FtrlOptimizer, self).__init__(
X
Xin Pan 已提交
2039 2040 2041
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
Q
qiaolongfei 已提交
2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081
        if learning_rate is None:
            raise ValueError("learning_rate is not set.")

        self.type = "ftrl"
        self._l1 = l1
        self._l2 = l2
        self._lr_power = lr_power

    def _create_accumulators(self, block, parameters):
        if not isinstance(block, framework.Block):
            raise TypeError("block is not instance of framework.Block.")

        for p in parameters:
            self._add_accumulator(self._squared_acc_str, p)
            self._add_accumulator(self._linear_acc_str, p)

    def _append_optimize_op(self, block, param_and_grad):
        if not isinstance(block, framework.Block):
            raise TypeError("block is not instance of framework.Block.")

        squared_acc = self._get_accumulator(self._squared_acc_str,
                                            param_and_grad[0])
        linear_acc = self._get_accumulator(self._linear_acc_str,
                                           param_and_grad[0])
        ftrl_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
                "SquaredAccumulator": squared_acc,
                "LinearAccumulator": linear_acc,
                "LearningRate": self._create_param_lr(param_and_grad),
            },
            outputs={
                "ParamOut": param_and_grad[0],
                "SquaredAccumOut": squared_acc,
                "LinearAccumOut": linear_acc
            },
            attrs={"l1": self._l1,
                   "l2": self._l1,
M
minqiyang 已提交
2082 2083
                   "lr_power": self._lr_power},
            stop_gradient=True)
Q
qiaolongfei 已提交
2084 2085 2086 2087

        return ftrl_op


Y
Yibing Liu 已提交
2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213 2214
class LambOptimizer(AdamOptimizer):
    """
    LAMB (Layer-wise Adaptive Moments optimizer for Batching training) Optimizer.

    LAMB Optimizer is designed to scale up the batch size of training without losing 
    accuracy, which supports adaptive element-wise updating and accurate layer-wise 
    correction. For more information, please refer to `Reducing BERT Pre-Training 
    Time from 3 Days to 76 Minutes <https://arxiv.org/abs/1904.00962>`_ .

    The updating of parameters follows:

    ..  math::

	m_t^l & = \\beta_1 m_{t - 1}^l + (1 - \\beta_1)g_t^l

	v_t^l & = \\beta_2 v_{t - 1}^l + (1 - \\beta_2)g_t^l \odot g_t^l

	\\widehat{m}_t^l & = m_t^l/(1 - \\beta_1^t)

	\\widehat{v}_t^l & = v_t^l/(1 - \\beta_2^t)
	
        r_1 & = \\left \| w_{t-1}^l \\right \|_2
	
        r_2 & = \\left \|  \\frac{\\widehat{m}_t^l}{\\sqrt{\\widehat{v}_t^l+\\epsilon}} + \\lambda w_{t-1}^l \\right \|_2

	r & = r_1 / r_2

	\\eta^l & = r \\times \\eta

	w_t^l & = w_{t-1}^l -\\eta ^l \\times (\\frac{\\widehat{m}_t^l}{\\sqrt{\\widehat{v}_t^l+\\epsilon}} + \\lambda w_{t-1}^l)


    where :math:`m` is the 1st moment, and :math:`v` the 2nd moment, :math:`\\eta` the 
    learning rate, :math:`\\lambda` the LAMB weight decay rate.

    Args:
        learning_rate (float|Variable): the learning rate used to update parameters. \
                                        Can be a float value or a Variable with one \
                                        float value as data element.
        lamb_weight_decay (float): The LAMB weight decay rate.
        beta1 (float): The exponential decay rate for the 1st moment estimates.
        beta2 (float): The exponential decay rate for the 2nd moment estimates.
        epsilon (float): A small float value for numerical stability.
        regularization: A Regularizer, such as
                        fluid.regularizer.L1DecayRegularizer.
        name (str|None): An optional name prefix.

    Examples:
        .. code-block:: python
            
            import paddle.fluid as fluid 

            data = fluid.layers.data(name='x', shape=[5], dtype='float32')
            hidden = fluid.layers.fc(input=data, size=10)
            cost = fluid.layers.mean(hidden)

            optimizer = fluid.optimizer.Lamb(learning_rate=0.002)
            optimizer.minimize(cost)
    """
    _moment1_acc_str = "moment1"
    _moment2_acc_str = "moment2"
    _beta1_pow_acc_str = "beta1_pow_acc"
    _beta2_pow_acc_str = "beta2_pow_acc"

    def __init__(self,
                 learning_rate=0.001,
                 lamb_weight_decay=0.01,
                 beta1=0.9,
                 beta2=0.999,
                 epsilon=1e-6,
                 regularization=None,
                 name=None):
        assert learning_rate is not None
        assert lamb_weight_decay is not None
        assert beta1 is not None
        assert beta2 is not None
        assert epsilon is not None
        super(LambOptimizer, self).__init__(
            learning_rate=learning_rate,
            regularization=regularization,
            beta1=beta1,
            beta2=beta2,
            epsilon=epsilon,
            name=name)
        self.type = "lamb"
        self._weight_decay = lamb_weight_decay

    def _append_optimize_op(self, block, param_and_grad):
        assert isinstance(block, framework.Block)

        moment1 = self._get_accumulator(self._moment1_acc_str,
                                        param_and_grad[0])
        moment2 = self._get_accumulator(self._moment2_acc_str,
                                        param_and_grad[0])
        beta1_pow_acc = self._get_accumulator(self._beta1_pow_acc_str,
                                              param_and_grad[0])
        beta2_pow_acc = self._get_accumulator(self._beta2_pow_acc_str,
                                              param_and_grad[0])

        # create the lamb optimize op
        lamb_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
                "LearningRate": self._create_param_lr(param_and_grad),
                "Moment1": moment1,
                "Moment2": moment2,
                "Beta1Pow": beta1_pow_acc,
                "Beta2Pow": beta2_pow_acc
            },
            outputs={
                "ParamOut": param_and_grad[0],
                "Moment1Out": moment1,
                "Moment2Out": moment2
            },
            attrs={
                "beta1": self._beta1,
                "beta2": self._beta2,
                "epsilon": self._epsilon,
                "weight_decay": self._weight_decay
            },
            stop_gradient=True)

        return lamb_op


2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228
# We short the class name, since users will use the optimizer with the package
# name. The sample code:
#
# import paddle.fluid as fluid
#
# sgd = fluid.optimizer.SGD(...)
#
# It is no need to add an `Optimizer` as the class suffix
SGD = SGDOptimizer
Momentum = MomentumOptimizer
Adagrad = AdagradOptimizer
Adam = AdamOptimizer
Adamax = AdamaxOptimizer
DecayedAdagrad = DecayedAdagradOptimizer
2229
Adadelta = AdadeltaOptimizer
Q
qingqing01 已提交
2230
RMSProp = RMSPropOptimizer
Q
qiaolongfei 已提交
2231
Ftrl = FtrlOptimizer
2232
LarsMomentum = LarsMomentumOptimizer
Y
Yibing Liu 已提交
2233
Lamb = LambOptimizer
2234 2235 2236


class ModelAverage(Optimizer):
2237
    """Accumulate the average of parameters within sliding window. The average
2238 2239
    result will be saved in temporary variables which can be applied to
    parameter variables of current model by calling 'apply()' method. And the
2240
    'restore()' method is used to restore the parameter values of current model.
2241 2242 2243 2244 2245 2246 2247 2248

    The size of average window is determined by average_window_rate,
    min_average_window, max_average_window and current update times.

    Args:
        average_window_rate: The rate of average window.
        min_average_window: The minimum size of average window.
        max_average_window: The maximum size of average window.
X
Xin Pan 已提交
2249 2250 2251
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
2252

2253
    Examples:
Q
qiaolongfei 已提交
2254 2255 2256

      .. code-block:: python

2257 2258 2259 2260 2261 2262
        import paddle.fluid as fluid
        import numpy

        # First create the Executor.
        place = fluid.CPUPlace()  # fluid.CUDAPlace(0)
        exe = fluid.Executor(place)
2263

2264 2265 2266 2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285
        train_program = fluid.Program()
        startup_program = fluid.Program()
        with fluid.program_guard(train_program, startup_program):
            # build net
            data = fluid.layers.data(name='X', shape=[1], dtype='float32')
            hidden = fluid.layers.fc(input=data, size=10)
            loss = fluid.layers.mean(hidden)
            optimizer = fluid.optimizer.Momentum(learning_rate=0.2, momentum=0.1)
            optimizer.minimize(loss)

            # build ModelAverage optimizer
            model_average = fluid.optimizer.ModelAverage(0.15,
                                                         min_average_window=10000,
                                                         max_average_window=20000)

            exe.run(startup_program)
            x = numpy.random.random(size=(10, 1)).astype('float32')
            outs = exe.run(program=train_program,
                           feed={'X': x},
                           fetch_list=[loss.name])

            # apply ModelAverage
2286
            with model_average.apply(exe):
2287 2288 2289 2290
                x = numpy.random.random(size=(10, 1)).astype('float32')
                exe.run(program=train_program,
                        feed={'X': x},
                        fetch_list=[loss.name])
2291 2292 2293
    """

    def __init__(self,
W
wanghaoshuang 已提交
2294
                 average_window_rate,
2295 2296
                 min_average_window=10000,
                 max_average_window=10000,
X
Xin Pan 已提交
2297 2298 2299 2300
                 regularization=None,
                 name=None):
        super(ModelAverage, self).__init__(
            0.0, regularization=regularization, name=name)
2301 2302 2303
        self.average_window = average_window_rate
        self.min_average_window = min_average_window
        self.max_average_window = max_average_window
2304

2305
        self.params_grads = []
2306 2307
        for param in framework.default_main_program().global_block(
        ).all_parameters():
2308
            if param.do_model_average != False:
2309
                grad = param.block.create_var(
2310 2311
                    name=unique_name.generate_with_ignorable_key(".".join(
                        [param.name, 'tmp'])),
2312 2313
                    dtype=param.dtype,
                    persistable=False,
W
wanghaoshuang 已提交
2314
                    stop_gradient=True)
2315
                self.params_grads.append((param, grad))
2316

2317
        for param, grad in self.params_grads:
2318 2319
            if grad is None:
                continue
X
Xin Pan 已提交
2320 2321
            with param.block.program._optimized_guard(
                [param, grad]), name_scope('move_average'):
2322
                self._append_average_accumulate_op(param)
2323

2324 2325 2326 2327
        self.apply_program = Program()
        block = self.apply_program.global_block()
        with program_guard(main_program=self.apply_program):
            for param_grad in self.params_grads:
2328
                self._add_average_apply_op(block, param_grad)
2329 2330 2331 2332 2333

        self.restore_program = Program()
        block = self.restore_program.global_block()
        with program_guard(main_program=self.restore_program):
            for param_grad in self.params_grads:
2334
                self._add_average_restore_op(block, param_grad)
2335

2336
    def _add_average_apply_op(self, block, param_grad):
L
Luo Tao 已提交
2337 2338 2339 2340 2341 2342
        param = block._clone_variable(param_grad[0])
        grad = block._clone_variable(param_grad[1])
        sum_1 = block._clone_variable(self._get_accumulator('sum_1', param))
        sum_2 = block._clone_variable(self._get_accumulator('sum_2', param))
        sum_3 = block._clone_variable(self._get_accumulator('sum_3', param))
        num_accumulates = block._clone_variable(
2343
            self._get_accumulator('num_accumulates', param))
L
Luo Tao 已提交
2344
        old_num_accumulates = block._clone_variable(
2345
            self._get_accumulator('old_num_accumulates', param))
L
Luo Tao 已提交
2346
        num_updates = block._clone_variable(
2347 2348 2349 2350 2351 2352
            self._get_accumulator('num_updates', param))
        # backup param value to grad
        layers.assign(input=param, output=grad)
        # param = (sum_1 + sum_2 + sum_3) / (num_accumulates + old_num_accumulates)
        tmp = layers.sum(x=[num_accumulates, old_num_accumulates])
        sum = layers.sum(x=[sum_1, sum_2, sum_3])
D
dzhwinter 已提交
2353 2354 2355 2356
        tmp = layers.cast(
            x=tmp, dtype='float32' if self._dtype == None else self._dtype)
        sum = layers.cast(
            x=sum, dtype='float32' if self._dtype == None else self._dtype)
S
sneaxiy 已提交
2357
        ops._elementwise_div(x=sum, y=tmp, out=param)
2358 2359

    def _add_average_restore_op(self, block, param_grad):
L
Luo Tao 已提交
2360 2361
        param = block._clone_variable(param_grad[0])
        grad = block._clone_variable(param_grad[1])
2362 2363 2364 2365 2366 2367 2368 2369 2370 2371 2372 2373 2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395 2396 2397 2398
        layers.assign(input=grad, output=param)

    def _append_average_accumulate_op(self, param):
        self.helper = LayerHelper("average_accumulate")
        sum_1 = self._add_accumulator('sum_1', param)
        sum_2 = self._add_accumulator('sum_2', param)
        sum_3 = self._add_accumulator('sum_3', param)
        num_accumulates = self._add_accumulator(
            'num_accumulates', param, dtype='int64', shape=[1])
        old_num_accumulates = self._add_accumulator(
            'old_num_accumulates', param, dtype='int64', shape=[1])
        num_updates = self._add_accumulator(
            'num_updates', param, dtype='int64', shape=[1])

        self.helper.append_op(
            type='average_accumulates',
            inputs={
                "param": param,
                "in_sum_1": sum_1,
                "in_sum_2": sum_2,
                "in_sum_3": sum_3,
                "in_num_accumulates": num_accumulates,
                "in_old_num_accumulates": old_num_accumulates,
                "in_num_updates": num_updates
            },
            outputs={
                "out_sum_1": sum_1,
                "out_sum_2": sum_2,
                "out_sum_3": sum_3,
                "out_num_accumulates": num_accumulates,
                "out_old_num_accumulates": old_num_accumulates,
                "out_num_updates": num_updates,
            },
            attrs={
                "average_window": self.average_window,
                "min_average_window": self.min_average_window,
                "max_average_window": self.max_average_window,
M
minqiyang 已提交
2399 2400
            },
            stop_gradient=True)
2401

S
rename  
sneaxiy 已提交
2402
    @signature_safe_contextmanager
2403
    def apply(self, executor, need_restore=True):
2404
        """Apply average values to parameters of current model.
2405 2406 2407 2408

        Args:
            executor(fluid.Executor): current executor.
            need_restore(bool): If you finally need to do restore, set it to True. Default is True.
2409
        """
2410 2411 2412 2413 2414 2415
        executor.run(self.apply_program)
        try:
            yield
        finally:
            if need_restore:
                self.restore(executor)
2416 2417 2418

    def restore(self, executor):
        """Restore parameter values of current model.
2419 2420 2421
        
        Args:
            executor(fluid.Executor): current executor.
2422
        """
2423
        executor.run(self.restore_program)
2424 2425 2426 2427 2428 2429 2430 2431 2432 2433


class ExponentialMovingAverage(object):
    """
    Compute the moving average of parameters with exponential decay.
    Given a parameter :math:`\\theta`, its exponential moving average (EMA)
    will be

    ..  math::

2434
        \\text{EMA}_0 & = 0
2435

2436 2437
	\\text{EMA}_t & = \\text{decay} * \\text{EMA}_{t-1} + (1 - \\text{decay}) * \\theta_t

Y
Yibing Liu 已提交
2438 2439 2440 2441
    The average results calculated by **update()** method will be saved in 
    temporary variables which are created and maintained by the object, and can 
    be applied to parameters of current model by calling **apply()** method. And 
    the **restore()** method is used to restore the parameters.
2442 2443 2444 2445 2446 2447 2448 2449 2450 2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462

    **Bias correction**. All EMAs are initialized to :math:`0` and hence they will be 
    zero biased, which can be corrected by divided by a factor 
    :math:`(1 - \\text{decay}^t)` , i.e., the actual EMAs applied to parameters 
    when calling **apply()** method would be 

    ..  math::
    
        \\widehat{\\text{EMA}}_t = \\frac{\\text{EMA}_t}{1 - \\text{decay}^t}

    **Decay rate scheduling**. A large decay rate very close to 1 would result 
    in that the averages move very slowly. And a better strategy is to set a 
    relative smaller decay rate in the very beginning. The argument **thres_steps**
    allows users to pass a Variable to schedule the decay rate, in this case, 
    the actual decay rate becomes
     
    ..  math::
    
        \\min(\\text{decay}, \\frac{1 + \\text{thres_steps}}{10 + \\text{thres_steps}})

    Usually **thres_steps** can be the global training steps.
2463 2464 2465


    Args:
2466 2467 2468
	decay (float): The exponential decay rate, usually close to 1, such as 
                       0.999, 0.9999, ... .
        thres_steps (Variable|None): If not `None`, schedule the decay rate.
2469 2470 2471 2472 2473 2474
	name (str|None): An optional name prefix.


    Examples:

	.. code-block:: python
2475 2476 2477 2478 2479 2480 2481 2482 2483 2484 2485 2486 2487 2488 2489 2490 2491 2492 2493 2494 2495 2496 2497 2498 2499 2500 2501 2502 2503 2504 2505 2506 2507 2508 2509 2510 2511 2512 2513 2514 2515 2516 2517 2518

	    import numpy
	    import paddle
	    import paddle.fluid as fluid

	    data = fluid.layers.data(name='x', shape=[5], dtype='float32')
	    hidden = fluid.layers.fc(input=data, size=10)
	    cost = fluid.layers.mean(hidden)

	    test_program = fluid.default_main_program().clone(for_test=True)

	    optimizer = fluid.optimizer.Adam(learning_rate=0.001)
	    optimizer.minimize(cost)

	    global_steps = fluid.layers.learning_rate_scheduler._decay_step_counter()
	    ema = fluid.optimizer.ExponentialMovingAverage(0.999, thres_steps=global_steps)
	    ema.update()

	    place = fluid.CPUPlace()
	    exe = fluid.Executor(place)
	    exe.run(fluid.default_startup_program())

	    for pass_id in range(3):
		for batch_id in range(6):
		    data = numpy.random.random(size=(10, 5)).astype('float32')
		    exe.run(program=fluid.default_main_program(),
			feed={'x': data}, 
			fetch_list=[cost.name])

		# usage 1
		with ema.apply(exe):
		    data = numpy.random.random(size=(10, 5)).astype('float32')
		    exe.run(program=test_program,
			    feed={'x': data}, 
			    fetch_list=[hidden.name])
			    

		 # usage 2
		with ema.apply(exe, need_restore=False):
		    data = numpy.random.random(size=(10, 5)).astype('float32')
		    exe.run(program=test_program,
			    feed={'x': data}, 
			    fetch_list=[hidden.name])
		ema.restore(exe)
2519 2520
    """

2521
    def __init__(self, decay=0.999, thres_steps=None, name=None):
2522
        self._decay = decay
2523
        self._thres_steps = thres_steps
2524
        self._name = name if name is not None else ''
2525 2526
        self._decay_var = self._get_ema_decay()

Y
Yibing Liu 已提交
2527
        self._params_tmps = []
2528
        for param in default_main_program().global_block().all_parameters():
2529 2530 2531 2532 2533 2534 2535
            if param.do_model_average != False:
                tmp = param.block.create_var(
                    name=unique_name.generate(".".join(
                        [self._name + param.name, 'ema_tmp'])),
                    dtype=param.dtype,
                    persistable=False,
                    stop_gradient=True)
Y
Yibing Liu 已提交
2536
                self._params_tmps.append((param, tmp))
2537

Y
Yibing Liu 已提交
2538 2539
        self._ema_vars = {}
        for param, tmp in self._params_tmps:
2540 2541
            with param.block.program._optimized_guard(
                [param, tmp]), name_scope('moving_average'):
Y
Yibing Liu 已提交
2542
                self._ema_vars[param.name] = self._create_ema_vars(param)
2543 2544 2545 2546

        self.apply_program = Program()
        block = self.apply_program.global_block()
        with program_guard(main_program=self.apply_program):
2547
            decay_pow = self._get_decay_pow(block)
Y
Yibing Liu 已提交
2548
            for param, tmp in self._params_tmps:
2549 2550
                param = block._clone_variable(param)
                tmp = block._clone_variable(tmp)
Y
Yibing Liu 已提交
2551
                ema = block._clone_variable(self._ema_vars[param.name])
2552
                layers.assign(input=param, output=tmp)
2553 2554
                # bias correction
                ema = ema / (1.0 - decay_pow)
2555 2556 2557 2558 2559
                layers.assign(input=ema, output=param)

        self.restore_program = Program()
        block = self.restore_program.global_block()
        with program_guard(main_program=self.restore_program):
Y
Yibing Liu 已提交
2560
            for param, tmp in self._params_tmps:
2561 2562 2563 2564
                tmp = block._clone_variable(tmp)
                param = block._clone_variable(param)
                layers.assign(input=tmp, output=param)

2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591
    def _get_ema_decay(self):
        with default_main_program()._lr_schedule_guard():
            decay_var = layers.tensor.create_global_var(
                shape=[1],
                value=self._decay,
                dtype='float32',
                persistable=True,
                name="scheduled_ema_decay_rate")

            if self._thres_steps is not None:
                decay_t = (self._thres_steps + 1.0) / (self._thres_steps + 10.0)
                with layers.control_flow.Switch() as switch:
                    with switch.case(decay_t < self._decay):
                        layers.tensor.assign(decay_t, decay_var)
                    with switch.default():
                        layers.tensor.assign(
                            np.array(
                                [self._decay], dtype=np.float32),
                            decay_var)
        return decay_var

    def _get_decay_pow(self, block):
        global_steps = layers.learning_rate_scheduler._decay_step_counter()
        decay_var = block._clone_variable(self._decay_var)
        decay_pow_acc = layers.elementwise_pow(decay_var, global_steps + 1)
        return decay_pow_acc

Y
Yibing Liu 已提交
2592
    def _create_ema_vars(self, param):
2593 2594 2595 2596 2597 2598 2599 2600 2601
        param_ema = layers.create_global_var(
            name=unique_name.generate(self._name + param.name + '_ema'),
            shape=param.shape,
            value=0.0,
            dtype=param.dtype,
            persistable=True)

        return param_ema

Y
Yibing Liu 已提交
2602 2603 2604 2605 2606
    def update(self):
        """ 
        Update Exponential Moving Average. Should only call this method in 
        train program.
        """
2607
        param_master_emas = []
Y
Yibing Liu 已提交
2608 2609 2610 2611
        for param, tmp in self._params_tmps:
            with param.block.program._optimized_guard(
                [param, tmp]), name_scope('moving_average'):
                param_ema = self._ema_vars[param.name]
2612 2613 2614 2615 2616 2617 2618 2619 2620 2621 2622 2623 2624 2625 2626 2627 2628 2629
                if self._ema_vars.has_key(param.name + '.master'):
                    master_ema = self._ema_vars[param.name + '.master']
                    param_master_emas.append([param_ema, master_ema])
                else:
                    ema_t = param_ema * self._decay_var + param * (
                        1 - self._decay_var)
                    layers.assign(input=ema_t, output=param_ema)

        # for fp16 params
        for param_ema, master_ema in param_master_emas:
            default_main_program().global_block().append_op(
                type="cast",
                inputs={"X": master_ema},
                outputs={"Out": param_ema},
                attrs={
                    "in_dtype": master_ema.dtype,
                    "out_dtype": param_ema.dtype
                })
Y
Yibing Liu 已提交
2630

2631 2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651 2652 2653
    @signature_safe_contextmanager
    def apply(self, executor, need_restore=True):
        """
        Apply moving average to parameters for evaluation.
        
        Args:
            executor (Executor): The Executor to execute applying.
            need_restore (bool): Whether to restore parameters after applying.
        """
        executor.run(self.apply_program)
        try:
            yield
        finally:
            if need_restore:
                self.restore(executor)

    def restore(self, executor):
        """Restore parameters.
        
        Args:
            executor (Executor): The Executor to execute restoring.
        """
        executor.run(self.restore_program)
H
hutuxian 已提交
2654 2655 2656


class PipelineOptimizer(object):
2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711
    """
    Pipeline Optimizer
    Train with pipeline mode. The program will be splited by cut_list.
    If the len of cut_list is k, then the whole program (including 
    backward part) will be splited to 2*k-1 sections. So the length of place_list
    and concurrency_list must be also 2*k-1.
    Note: Though the asynchronous mode is applied in pipeline training to speed up,
    the final performance depends on the training progress of each pipeline heavily.
    And we will try the synchronous mode in the future
    Args:
        optimizer (Optimizer): The based optimizer, such as SGD
        cut_list (list of Variable list): The cut variable of the main_program
        place_list (list of Place): The place where the section will run on
        concurrency_list (list of int): The concurrency degree
        queue_size (int): Each section will consume scopes from its in-scope queue 
                        and produce scopes to out-scope queue. And this parameter 
                        specify the scope queue size. [Optional. Default: 30]
        sync_steps (int): The synchronization steps between different cards. [Optional. Default: 1]
        start_cpu_core_id (int): specify the first cpu core id. [Optional. Default:0]
    Examples:
        .. code-block:: python
        x = fluid.layers.data(name='x', shape=[1], dtype='int64', lod_level=0)
        y = fluid.layers.data(name='y', shape=[1], dtype='int64', lod_level=0)
        emb_x = layers.embedding(input=x, param_attr=fluid.ParamAttr(name="embx"), size=[10,2], is_sparse=False)
        emb_y = layers.embedding(input=y, param_attr=fluid.ParamAttr(name="emby",learning_rate=0.9), size=[10,2], is_sparse=False)
        concat = layers.concat([emb_x, emb_y], axis=1)
        fc = layers.fc(input=concat, name="fc", size=1, num_flatten_dims=1, bias_attr=False)
        loss = layers.reduce_mean(fc)
        optimizer = fluid.optimizer.SGD(learning_rate=0.5)
        optimizer = fluid.optimizer.PipelineOptimizer(optimizer,
                cut_list=[[emb_x, emb_y], [loss]],
                place_list=[fluid.CPUPlace(), fluid.CUDAPlace(0), fluid.CPUPlace()],
                concurrency_list=[1, 1, 4],
                queue_size=2,
                sync_steps=1,
                )
        optimizer.minimize(loss)
        place = fluid.CPUPlace()
        exe = fluid.Executor(place)
        exe.run(fluid.default_startup_program())
        filelist = [] # you should set your own filelist, e.g. filelist = ["dataA.txt"]
        dataset = fluid.DatasetFactory().create_dataset("FileInstantDataset")
        dataset.set_use_var([x,y])
        dataset.set_batch_size(batch_size)
        dataset.set_filelist(filelist)
        exe.train_from_dataset(
                    fluid.default_main_program(),
                    dataset,
                    thread=2,
                    debug=False,
                    fetch_list=[],
                    fetch_info=[],
                    print_period=1)
    """

H
hutuxian 已提交
2712 2713 2714 2715 2716 2717 2718 2719 2720 2721 2722 2723 2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748 2749 2750 2751 2752 2753 2754 2755 2756 2757 2758 2759 2760 2761 2762 2763 2764 2765 2766 2767 2768 2769 2770 2771 2772 2773 2774 2775 2776 2777 2778 2779 2780 2781 2782 2783 2784 2785 2786 2787 2788 2789 2790 2791 2792 2793 2794 2795 2796 2797 2798 2799 2800 2801 2802 2803 2804 2805 2806 2807 2808 2809 2810 2811 2812 2813 2814 2815 2816 2817 2818 2819 2820 2821 2822 2823 2824 2825 2826 2827 2828 2829 2830 2831 2832 2833 2834 2835 2836 2837 2838 2839 2840 2841 2842 2843 2844 2845 2846 2847 2848 2849 2850 2851 2852 2853 2854 2855 2856 2857 2858 2859 2860 2861 2862 2863 2864 2865 2866 2867 2868 2869 2870 2871 2872 2873 2874 2875 2876 2877 2878 2879 2880 2881 2882 2883 2884 2885 2886 2887 2888 2889 2890 2891 2892 2893 2894 2895 2896 2897 2898 2899 2900 2901 2902 2903 2904 2905 2906 2907 2908 2909 2910 2911 2912 2913 2914 2915 2916 2917 2918 2919 2920 2921 2922 2923 2924 2925 2926 2927 2928 2929 2930 2931 2932 2933 2934 2935
    def __init__(self,
                 optimizer,
                 cut_list=None,
                 place_list=None,
                 concurrency_list=None,
                 queue_size=30,
                 sync_steps=1,
                 start_cpu_core_id=0):
        # TODO: check properties
        self._optimizer = optimizer
        self._cut_list = cut_list
        self._place_list = place_list
        self._concurrency_list = concurrency_list
        self._queue_size = queue_size
        self._sync_steps = sync_steps
        self._start_cpu_core_id = start_cpu_core_id

    def create_vars(self, block, main_program):
        used_var_set = set()
        for op_idx in range(block.desc.op_size()):
            op_desc = block.desc.op(op_idx)
            vars = op_desc.input_arg_names() + op_desc.output_arg_names()
            for var in vars:
                if var in used_var_set:
                    continue
                used_var_set.add(var)
                source_var = main_program.block(0).var(str(var))
                block._clone_variable(source_var, False)

    def extract_section_opt_ops(self, ops, cut_point_name):
        """
        Extract opt ops in the given section
        """
        output_names = set(cut_point_name)
        relevant_op_flags = [True] * len(ops)
        for i, op in reversed(list(enumerate(ops))):
            if _some_in_set_(op.desc.output_arg_names(), output_names):
                for name in op.desc.input_arg_names():
                    output_names.add(name)
            else:
                relevant_op_flags[i] = False

        op_path = [ops[i] for i in range(len(ops)) if relevant_op_flags[i]]
        return op_path

    def find_input_output(self, ops, name, is_forward=True):
        """
        Find the inputs or outputs of a section
        """
        all_set = set()
        part_set = set()
        for op in ops:
            if is_forward:
                part_set.update(op.desc.output_arg_names())
            else:
                part_set.update(op.desc.input_arg_names())
            all_set.update(op.desc.output_arg_names())
            all_set.update(op.desc.input_arg_names())
        return all_set - part_set

    def find_persistable_vars(self, ops, whole_parameters):
        """
        find the persistable input vars in current section
        """
        res = set()
        for op in ops:
            vars = op.desc.input_arg_names()
            for var in vars:
                if var in whole_parameters:
                    res.add(var)
        return res

    def _is_opt_role_op(self, op):
        op_maker = core.op_proto_and_checker_maker
        optimize_role = core.op_proto_and_checker_maker.OpRole.Optimize
        if op_maker.kOpRoleAttrName() in op.attr_names and \
                int(op.all_attrs()[op_maker.kOpRoleAttrName()]) & int(optimize_role) != 0:
            return True
        return False

    def _is_lr_role_op(self, op):
        op_maker = core.op_proto_and_checker_maker
        optimize_role = core.op_proto_and_checker_maker.OpRole.LRSched
        if op_maker.kOpRoleAttrName() in op.attr_names and \
                int(op.all_attrs()[op_maker.kOpRoleAttrName()]) == int(optimize_role):
            return True
        return False

    def extract_section_ops(self, ops, cut_point_name):
        """
        Extract ops in the given section 
        """
        output_names = set(cut_point_name)
        relevant_op_flags = [True] * len(ops)
        for i, op in reversed(list(enumerate(ops))):
            if not self._is_opt_role_op(op) and _some_in_set_(
                    op.desc.output_arg_names(), output_names):
                for name in op.desc.input_arg_names():
                    output_names.add(name)
            elif op.desc.type() == "print" and op.desc.input_arg_names()[
                    0] in output_names:
                continue
            else:
                relevant_op_flags[i] = False

        op_path = [ops[i] for i in range(len(ops)) if relevant_op_flags[i]]
        return op_path

    def find_section_opt(self, ops, params):
        res = self.extract_section_opt_ops(ops, params)
        return res

    def split_program(self, main_program, cut_list):
        programs = []
        block = main_program.block(0)
        whole_parameters = [e.name for e in block.all_parameters()]
        cut_var_names = []
        cut_len = len(cut_list)
        sec_params = []
        for i, cut_vars in enumerate(cut_list[:-1]):
            cut_var_names.append([cut_var.name for cut_var in cut_vars])
        for i, cut_vars in reversed(list(enumerate(cut_list[:-1]))):
            cut_var_names.append(
                [_append_grad_suffix_(cut_var.name) for cut_var in cut_vars])
            if i == 0:
                cut_var_names[-1] += [var.name for var in cut_list[-1]]
        ops = block.ops[:]
        for i, cut_vars in enumerate(cut_var_names):
            program = {
                "program": Program(),
                "input_set": set(),
                "output_set": set()
            }
            cur_ops = self.extract_section_ops(ops, cut_vars)
            if i == 0:
                for op in ops:
                    if self._is_lr_role_op(op):
                        cur_ops.append(op)
            #prevent inplace in/out
            program["input_set"].update(
                self.find_input_output(
                    cur_ops, [], is_forward=True))
            for e in cur_ops:
                ops.remove(e)

            if i < cut_len:
                sec_params.append(
                    self.find_persistable_vars(cur_ops, whole_parameters))
            if i >= cut_len - 1:
                opt_ops = self.find_section_opt(ops,
                                                sec_params[2 * cut_len - 2 - i])

                for e in opt_ops:
                    ops.remove(e)
                cur_ops += opt_ops

            op_descs = [op.desc for op in cur_ops]
            for op_desc in op_descs:
                ap_op = program["program"].block(0).desc.append_op()
                ap_op.copy_from(op_desc)
            program["input_set"].update(
                self.find_input_output(
                    cur_ops, cut_vars, is_forward=True))
            program["input_set"].update(sec_params[min(i, 2 * cut_len - 2 - i)])
            program["output_set"].update(
                self.find_input_output(
                    cur_ops, cut_vars, is_forward=False))
            programs.append(program)
        program = {
            "program": Program(),
            "input_set": set(),
            "output_set": set()
        }
        op_descs = [op.desc for op in ops]
        for op_desc in op_descs:
            ap_op = program["program"].block(0).desc.append_op()
            ap_op.copy_from(op_desc)
        program["input_set"].update(
            [cut_var.name + "@GRAD" for cut_var in cut_list[0]])
        program["input_set"].update(
            self.find_input_output(
                ops, [], is_forward=True))
        program["input_set"].update(sec_params[0])
        programs.append(program)
        inputs = set()
        for program in reversed(list(programs)):
            output_list = list(program["output_set"])
            for output in output_list:
                if output not in inputs:
                    program["output_set"].remove(output)
            inputs.update(program["input_set"])
        return programs

    def minimize(self,
                 loss,
                 startup_program=None,
                 parameter_list=None,
                 no_grad_set=None):
        self._optimizer.minimize(loss, startup_program, parameter_list,
                                 no_grad_set)
        program = loss.block.program
        program_list = self.split_program(program, self._cut_list)
        for p in program_list:
            self.create_vars(p["program"].block(0), program)
        whole_parameters = [e.name for e in program.block(0).all_parameters()]
        param_need_sync = []
        for i, section_p in enumerate(program_list):
            if not isinstance(self._place_list[i], core.CUDAPlace):
                continue
            section_var = [e for e in section_p["program"].block(0).vars]
            for p in section_var:
                if p in whole_parameters:
                    param_need_sync.append(p)
        program._pipeline_opt = {
            "trainer": "PipelineTrainer",
            "device_worker": "Section",
            "section_program_list": program_list,
            "place_list": self._place_list,
            "concurrency_list": self._concurrency_list,
            "queue_size": self._queue_size,
            "start_cpu_core_id": self._start_cpu_core_id,
            "sync_steps": self._sync_steps,
            "param_need_sync": param_need_sync
        }