optimizer.py 117.0 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',
M
mapingshuo 已提交
46
    'ExponentialMovingAverage', 'PipelineOptimizer', 'LookaheadOptimizer'
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
        Examples:
            .. code-block:: python

513
                import paddle.fluid as fluid
514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538
                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
539 540 541 542 543 544 545 546 547 548 549 550 551 552
    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
553
        if framework.in_dygraph_mode():
C
chengduo 已提交
554 555
            with program_guard(framework.default_main_program(),
                               framework.default_startup_program()):
556 557
                params_grads = append_regularization_ops(params_grads,
                                                         self.regularization)
C
chengduo 已提交
558 559 560 561 562 563 564
                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

565
    @imperative_base.no_grad
Q
Qiao Longfei 已提交
566 567
    def minimize(self,
                 loss,
568
                 startup_program=None,
Q
Qiao Longfei 已提交
569
                 parameter_list=None,
570 571
                 no_grad_set=None,
                 grad_clip=None):
572 573 574 575 576
        """
        Add operations to minimize `loss` by updating `parameter_list`.

        This method combines interface `backward()` and
        `apply_gradients()` into one.
M
minqiyang 已提交
577

578 579 580 581 582 583
        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.
584
            grad_clip (GradClipBase|None) : Gradient clip strategy
Q
Qiao Longfei 已提交
585

586 587 588
        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
589
        """
C
chengduo 已提交
590 591 592 593 594
        params_grads = self.backward(
            loss,
            startup_program=startup_program,
            parameter_list=parameter_list,
            no_grad_set=no_grad_set)
595 596 597 598 599

        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
600 601
        optimize_ops = self.apply_optimize(
            loss, startup_program=startup_program, params_grads=params_grads)
M
minqiyang 已提交
602

603 604 605
        if framework.in_dygraph_mode():
            framework._dygraph_tracer()._clear_ops()

Q
Qiao Longfei 已提交
606
        return optimize_ops, params_grads
Q
Qiao Longfei 已提交
607 608 609


class SGDOptimizer(Optimizer):
Q
qiaolongfei 已提交
610 611 612 613 614 615 616 617 618 619
    """
    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
620 621 622
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
Q
qiaolongfei 已提交
623 624 625 626

    Examples:
        .. code-block:: python

627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651
            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
652 653
    """

X
Xin Pan 已提交
654
    def __init__(self, learning_rate, regularization=None, name=None):
Q
Qiao Longfei 已提交
655
        assert learning_rate is not None
Q
Qiao Longfei 已提交
656
        super(SGDOptimizer, self).__init__(
X
Xin Pan 已提交
657 658 659
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
Q
Qiao Longfei 已提交
660 661
        self.type = "sgd"

662 663
    def _append_optimize_op(self, block, param_and_grad):
        assert isinstance(block, framework.Block)
664

Q
Qiao Longfei 已提交
665 666 667 668 669 670
        # create the optimize op
        sgd_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
671
                "LearningRate": self._create_param_lr(param_and_grad)
Q
Qiao Longfei 已提交
672
            },
M
minqiyang 已提交
673 674
            outputs={"ParamOut": param_and_grad[0]},
            stop_gradient=True)
Q
Qiao Longfei 已提交
675 676

        return sgd_op
677 678 679


class MomentumOptimizer(Optimizer):
Q
qiaolongfei 已提交
680 681 682 683 684 685 686 687 688 689 690 691 692 693
    """

    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):

694
        &\quad   param = param - (gradient + mu * velocity) * learning\_rate
Q
qiaolongfei 已提交
695 696 697

        & else:

Q
qiaolongfei 已提交
698
        &\quad   param = param - learning\_rate * velocity
Q
qiaolongfei 已提交
699 700 701 702 703 704

    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
705 706 707
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
Q
qiaolongfei 已提交
708 709 710 711

    Examples:
        .. code-block:: python

712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736
            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)

737 738 739
    """
    _velocity_acc_str = "velocity"

X
Xin Pan 已提交
740 741 742 743 744 745
    def __init__(self,
                 learning_rate,
                 momentum,
                 use_nesterov=False,
                 regularization=None,
                 name=None):
746 747
        assert learning_rate is not None
        assert momentum is not None
Q
Qiao Longfei 已提交
748
        super(MomentumOptimizer, self).__init__(
X
Xin Pan 已提交
749 750 751
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
752 753
        self.type = "momentum"
        self._momentum = momentum
754
        self._use_nesterov = bool(use_nesterov)
755 756 757 758 759

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

        for p in parameters:
Q
Qiao Longfei 已提交
760
            self._add_accumulator(self._velocity_acc_str, p)
761 762 763 764 765 766 767 768 769 770 771 772 773

    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,
774
                "LearningRate": self._create_param_lr(param_and_grad)
775 776 777 778 779
            },
            outputs={
                "ParamOut": param_and_grad[0],
                "VelocityOut": velocity_acc
            },
780
            attrs={"mu": self._momentum,
M
minqiyang 已提交
781 782
                   "use_nesterov": self._use_nesterov},
            stop_gradient=True)
783 784

        return momentum_op
785 786


787 788 789 790 791
class DGCMomentumOptimizer(MomentumOptimizer):
    """

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

G
gongweibao 已提交
792
    DGC reduces the communication bandwidth by sending only the important gradients (sparse update):\
793 794
        only gradients larger than a threshold are transmitted.

G
gongweibao 已提交
795
    To avoid losing information, DGC accumulates the rest of the gradients locally.
796 797 798

    Eventually, these gradients become large enough to be transmitted.

G
gongweibao 已提交
799
    Thus, DGC sends the large gradients immediately but eventually send all of the gradients over time.
800

G
gongweibao 已提交
801
    To ensure no loss of accuracy, DGC employs momentum correction and local gradient clipping on top of the gradient sparsification to maintain model performance.
802 803 804 805

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

    This optimizer will do two things:
806

807 808
        1. Compress the gradient by get TopK import value from tensor \
            and use it for allreduce to reduce network bandwidth.
809

810 811 812 813 814 815
        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
816
        rampup_begin_step (int): The beginning step from which gradient compression is implemented.
817 818 819 820 821 822 823
        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
824
        num_trainers: The number of training nodes.
825
        regularization: A Regularizer, such as fluid.regularizer.L2DecayRegularizer.
G
gongweibao 已提交
826
        name: An optional name prefix.
827 828 829 830

    Examples:
        .. code-block:: python

831
            import paddle.fluid as fluid
832
            optimizer = fluid.optimizer.DGCMomentumOptimizer(
G
gongweibao 已提交
833 834 835 836 837
                        learning_rate=0.0001,
                        momentum=0.9,
                        rampup_step=1000,
                        rampup_begin_step=1252,
                        sparsity=[0.999, 0.999])
838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903

    """

    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
904
            counter_name=core.dgc.kDGCCounterName(), begin=0)
905 906 907 908 909 910

        # 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
911
            name=core.dgc.kDGCRampUpBeginStepName(),
912 913 914 915
            value=self._rampup_begin_step * 1.0,
            force_cpu=True)

        for param_var, grad_var in param_and_grads:
G
gongweibao 已提交
916
            var_numel = abs(reduce(lambda x, y: x * y, param_var.shape))
917 918 919 920 921 922 923 924 925 926
            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
927
                name=param_var.name + core.dgc.kDGCUName(),
928 929 930 931 932
                value=0.0)
            v_var = tensor.create_global_var(
                shape=param_var.shape,
                dtype=param_var.dtype,
                persistable=True,
G
gongweibao 已提交
933
                name=param_var.name + core.dgc.kDGCVName(),
934 935 936 937 938 939
                value=0.0)

            k_var = tensor.create_global_var(
                shape=[1],
                dtype=param_var.dtype,
                persistable=True,
G
gongweibao 已提交
940
                name=param_var.name + core.dgc.kDGCKName(),
941 942 943 944 945 946 947
                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
948
                name=param_var.name + core.dgc.kDGCEncodedName(),
949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989
                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:
990 991
            name = unique_name.generate_with_ignorable_key(".".join(
                [helper.name, 'tmp']))
992 993 994 995 996

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

        helper.append_op(
G
gongweibao 已提交
997
            type="dgc_clip_by_norm",
998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009
            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
1010
                x=grad_var, max_norm=clip_norm, name=grad_var.name)
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 1044 1045

    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])


1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069
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
1070

1071 1072 1073 1074

    Examples:
        .. code-block:: python

1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090
            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])
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 1138 1139
    """
    _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
1140 1141
            },
            stop_gradient=True)
1142 1143 1144 1145

        return momentum_op


1146
class AdagradOptimizer(Optimizer):
Q
qiaolongfei 已提交
1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166
    """
    **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
1167 1168 1169
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
X
xuezhong 已提交
1170
        initial_accumulator_value (float): Initial value for moment accumulator.
Q
qiaolongfei 已提交
1171 1172 1173 1174

    Examples:
        .. code-block:: python

1175 1176 1177 1178 1179 1180 1181 1182
            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
1183
            optimizer = fluid.optimizer.Adagrad(learning_rate=0.2)
1184 1185 1186 1187 1188 1189 1190
            optimizer.minimize(out)

            exe = fluid.Executor(fluid.CPUPlace())
            exe.run(fluid.default_startup_program())
            exe.run(
                feed={"inp": np_inp},
                fetch_list=[out.name])
1191 1192 1193
    """
    _moment_acc_str = "moment"

X
Xin Pan 已提交
1194 1195 1196 1197
    def __init__(self,
                 learning_rate,
                 epsilon=1.0e-6,
                 regularization=None,
1198
                 name=None,
X
xuezhong 已提交
1199
                 initial_accumulator_value=0.0):
1200 1201
        assert learning_rate is not None
        assert epsilon is not None
Q
Qiao Longfei 已提交
1202
        super(AdagradOptimizer, self).__init__(
X
Xin Pan 已提交
1203 1204 1205
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
1206 1207
        self.type = "adagrad"
        self._epsilon = epsilon
1208
        self.initial_accumulator_value = initial_accumulator_value
1209 1210 1211 1212 1213

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

        for p in parameters:
Q
Qiao Longfei 已提交
1214
            self._add_accumulator(self._moment_acc_str, p)
1215 1216 1217 1218 1219 1220

    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])
1221 1222 1223 1224 1225 1226 1227 1228 1229 1230
        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,
            })
1231

1232
        # Create the adagrad optimizer op
1233 1234 1235 1236 1237 1238
        adagrad_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
                "Moment": moment_acc,
1239
                "LearningRate": self._create_param_lr(param_and_grad)
1240 1241 1242
            },
            outputs={"ParamOut": param_and_grad[0],
                     "MomentOut": moment_acc},
M
minqiyang 已提交
1243 1244
            attrs={"epsilon": self._epsilon},
            stop_gradient=True)
1245 1246

        return adagrad_op
1247 1248 1249


class AdamOptimizer(Optimizer):
Q
qiaolongfei 已提交
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 1275 1276
    """
    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.
1277
        regularization: A Regularizer, such as fluid.regularizer.L2DecayRegularizer.
X
Xin Pan 已提交
1278
        name: A optional name prefix.
1279 1280 1281 1282 1283 1284
        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
1285 1286 1287 1288

    Examples:
        .. code-block:: python

1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311
            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
1312

1313 1314 1315
    """
    _moment1_acc_str = "moment1"
    _moment2_acc_str = "moment2"
Q
qiaolongfei 已提交
1316 1317
    _beta1_pow_acc_str = "beta1_pow_acc"
    _beta2_pow_acc_str = "beta2_pow_acc"
1318 1319 1320 1321 1322

    def __init__(self,
                 learning_rate=0.001,
                 beta1=0.9,
                 beta2=0.999,
1323
                 epsilon=1e-8,
X
Xin Pan 已提交
1324
                 regularization=None,
Q
Qiao Longfei 已提交
1325
                 name=None,
Q
Qiao Longfei 已提交
1326
                 lazy_mode=False):
1327 1328 1329 1330
        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
1331
        super(AdamOptimizer, self).__init__(
X
Xin Pan 已提交
1332 1333 1334
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
1335 1336 1337 1338
        self.type = "adam"
        self._beta1 = beta1
        self._beta2 = beta2
        self._epsilon = epsilon
Q
Qiao Longfei 已提交
1339
        self._lazy_mode = lazy_mode
1340 1341 1342 1343 1344 1345

    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
1346 1347
            self._add_accumulator(self._moment1_acc_str, p)
            self._add_accumulator(self._moment2_acc_str, p)
Q
qiaolongfei 已提交
1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359
            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])
1360 1361 1362 1363 1364 1365 1366 1367

    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
1368 1369 1370 1371 1372
        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])

1373
        # create the adam optimize op
1374 1375 1376 1377 1378
        adam_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
1379
                "LearningRate": self._create_param_lr(param_and_grad),
1380 1381
                "Moment1": moment1,
                "Moment2": moment2,
Q
qiaolongfei 已提交
1382 1383
                "Beta1Pow": beta1_pow_acc,
                "Beta2Pow": beta2_pow_acc
1384 1385 1386 1387 1388 1389 1390 1391 1392
            },
            outputs={
                "ParamOut": param_and_grad[0],
                "Moment1Out": moment1,
                "Moment2Out": moment2
            },
            attrs={
                "beta1": self._beta1,
                "beta2": self._beta2,
Q
Qiao Longfei 已提交
1393
                "epsilon": self._epsilon,
1394 1395
                "lazy_mode": self._lazy_mode,
                "min_row_size_to_use_multithread": 1000
M
minqiyang 已提交
1396 1397
            },
            stop_gradient=True)
1398 1399 1400

        return adam_op

1401
    def _finish_update(self, block, param_and_grads):
1402 1403 1404
        """Update Beta1 and Beta2 Power accumulators
        """
        assert isinstance(block, framework.Block)
Q
Qiao Longfei 已提交
1405
        main_block = block.program.global_block()
1406 1407 1408
        for param, grad in param_and_grads:
            if grad is None:
                continue
X
Xin Pan 已提交
1409 1410
            with param.block.program._optimized_guard(
                [param, grad]), name_scope("optimizer"):
1411 1412 1413 1414 1415 1416 1417 1418
                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
1419 1420
                    attrs={"scale": self._beta1},
                    stop_gradient=True)
1421 1422 1423 1424 1425

                main_block.append_op(
                    type="scale",
                    inputs={"X": beta2_pow_acc},
                    outputs={"Out": beta2_pow_acc},
M
minqiyang 已提交
1426 1427
                    attrs={"scale": self._beta2},
                    stop_gradient=True)
1428 1429 1430


class AdamaxOptimizer(Optimizer):
Q
qiaolongfei 已提交
1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454
    """
    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.

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 1480 1481
    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
1482 1483 1484 1485 1486 1487
    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
1488 1489 1490
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
Q
qiaolongfei 已提交
1491

C
chengduo 已提交
1492 1493
    Notes:
       Currently, AdamaxOptimizer doesn't support sparse parameter optimization.
1494 1495 1496
    """
    _moment_acc_str = "moment"
    _inf_norm_acc_str = "inf_norm"
Q
qiaolongfei 已提交
1497
    _beta1_pow_acc_str = "beta1_pow_acc"
1498 1499 1500 1501 1502

    def __init__(self,
                 learning_rate=0.001,
                 beta1=0.9,
                 beta2=0.999,
1503
                 epsilon=1e-8,
X
Xin Pan 已提交
1504 1505
                 regularization=None,
                 name=None):
1506 1507 1508 1509
        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
1510
        super(AdamaxOptimizer, self).__init__(
X
Xin Pan 已提交
1511 1512 1513
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
1514 1515 1516 1517 1518 1519 1520 1521
        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
1522 1523
            self._add_accumulator(self._moment_acc_str, p)
            self._add_accumulator(self._inf_norm_acc_str, p)
Q
qiaolongfei 已提交
1524 1525 1526 1527 1528 1529
            self._add_accumulator(
                name=self._beta1_pow_acc_str,
                param=p,
                dtype='float32',
                fill_value=self._beta1,
                shape=[1])
1530 1531 1532 1533 1534 1535 1536

    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
1537 1538
        beta1_pow_acc = self._get_accumulator(self._beta1_pow_acc_str,
                                              param_and_grad[0])
1539 1540 1541 1542 1543 1544
        # create the adamax optimize op
        adamax_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
1545
                "LearningRate": self._create_param_lr(param_and_grad),
1546 1547
                "Moment": moment,
                "InfNorm": inf_norm,
Q
qiaolongfei 已提交
1548
                "Beta1Pow": beta1_pow_acc
1549 1550 1551 1552 1553 1554 1555 1556 1557 1558
            },
            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
1559 1560
            },
            stop_gradient=True)
1561 1562 1563

        return adamax_op

1564
    def _finish_update(self, block, parameters_and_grads):
1565 1566 1567
        """Update Beta1 Power accumulator
        """
        assert isinstance(block, framework.Block)
Q
Qiao Longfei 已提交
1568
        main_block = block.program.global_block()
1569 1570 1571
        for param, grad in parameters_and_grads:
            if grad is None:
                continue
X
Xin Pan 已提交
1572 1573
            with param.block.program._optimized_guard(
                [param, grad]), name_scope('adamx'):
1574 1575 1576 1577 1578 1579
                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
1580 1581
                    attrs={"scale": self._beta1},
                    stop_gradient=True)
1582 1583 1584


class DecayedAdagradOptimizer(Optimizer):
1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606
    """
    **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
1607 1608 1609
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
1610 1611 1612 1613

    Examples:
        .. code-block:: python

1614 1615 1616 1617 1618 1619 1620
            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 )
1621 1622
            optimizer = fluid.optimizer.DecayedAdagrad(learning_rate=0.2)
            optimizer.minimize(cost)
C
chengduo 已提交
1623 1624 1625

    Notes:
       Currently, DecayedAdagradOptimizer doesn't support sparse parameter optimization.
1626 1627 1628
    """
    _moment_acc_str = "moment"

X
Xin Pan 已提交
1629 1630 1631 1632 1633 1634
    def __init__(self,
                 learning_rate,
                 decay=0.95,
                 epsilon=1.0e-6,
                 regularization=None,
                 name=None):
1635 1636 1637 1638
        assert learning_rate is not None
        assert decay is not None
        assert epsilon is not None

Q
Qiao Longfei 已提交
1639
        super(DecayedAdagradOptimizer, self).__init__(
X
Xin Pan 已提交
1640 1641 1642
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
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 1668 1669
        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
1670 1671
            attrs={"epsilon": self._epsilon},
            stop_gradient=True)
1672 1673

        return decayed_adagrad_op
1674 1675


1676
class AdadeltaOptimizer(Optimizer):
1677 1678
    """
    **Adadelta Optimizer**
Q
qiaolongfei 已提交
1679

1680
    Simple Adadelta optimizer with average squared grad state and
1681
    average squared update state.
1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693
    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
1694
        learning_rate(float): global learning rate
1695 1696
        rho(float): rho in equation
        epsilon(float): epsilon in equation
X
Xin Pan 已提交
1697 1698 1699
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
1700 1701 1702 1703

    Examples:
        .. code-block:: python

1704
            import paddle.fluid as fluid
1705 1706 1707
            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
1708 1709 1710

    Notes:
       Currently, AdadeltaOptimizer doesn't support sparse parameter optimization.
1711
    """
1712

1713 1714 1715
    _avg_squared_grad_acc_str = "_avg_squared_grad"
    _avg_squared_update_acc_str = "_avg_squared_update"

X
Xin Pan 已提交
1716 1717 1718 1719 1720 1721
    def __init__(self,
                 learning_rate,
                 epsilon=1.0e-6,
                 rho=0.95,
                 regularization=None,
                 name=None):
1722 1723 1724 1725 1726 1727
        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.")
1728
        super(AdadeltaOptimizer, self).__init__(
X
Xin Pan 已提交
1729 1730 1731
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
1732 1733 1734 1735 1736
        self.type = "adadelta"
        self._epsilon = epsilon
        self._rho = rho

    def _create_accumulators(self, block, parameters):
1737 1738
        if not isinstance(block, framework.Block):
            raise TypeError("block is not instance of framework.Block.")
1739 1740 1741 1742 1743 1744

        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):
1745 1746
        if not isinstance(block, framework.Block):
            raise TypeError("block is not instance of framework.Block.")
1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767

        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
1768 1769
                   "rho": self._rho},
            stop_gradient=True)
1770 1771 1772 1773

        return adadelta_op


Q
qingqing01 已提交
1774 1775 1776 1777 1778 1779 1780 1781 1782 1783
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
1784
        r(w, t) & = \\rho r(w, t-1) + (1 - \\rho)(\\nabla Q_{i}(w))^2
Q
qingqing01 已提交
1785 1786 1787 1788

        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
1789
    each weight. Then dividing the gradient by :math:`sqrt{v(w,t)}`.
Q
qingqing01 已提交
1790 1791 1792 1793 1794 1795

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

    ..  math::

Q
qiaolongfei 已提交
1796
        r(w, t) & = \\rho r(w, t-1) + (1 - \\rho)(\\nabla Q_{i}(w))^2
Q
qingqing01 已提交
1797

1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811
        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
1812 1813 1814 1815
            \\epsilon}} \\nabla Q_{i}(w)

        w & = w - v(w, t)

Q
qiaolongfei 已提交
1816
    where, :math:`\\rho` is a hyperparameter and typical values are 0.9, 0.95
Q
qingqing01 已提交
1817 1818 1819 1820 1821 1822
    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
1823
        learning_rate(float): global learning rate.
Q
qingqing01 已提交
1824 1825 1826
        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
1827
        momentum(float): :math:`\\beta` in equation is the momentum term,
Q
qingqing01 已提交
1828
            set 0.0 by default.
1829 1830 1831 1832
        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
1833 1834 1835
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
Q
qingqing01 已提交
1836 1837 1838 1839 1840 1841 1842

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

    Examples:
          .. code-block:: python

1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867
            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
1868 1869 1870 1871
    """

    _momentum_acc_str = "momentum"
    _mean_square_acc_str = "mean_square"
1872
    _mean_grad_acc_str = "mean_grad"
Q
qingqing01 已提交
1873 1874 1875 1876 1877 1878

    def __init__(self,
                 learning_rate,
                 rho=0.95,
                 epsilon=1.0e-6,
                 momentum=0.0,
1879
                 centered=False,
X
Xin Pan 已提交
1880 1881
                 regularization=None,
                 name=None):
Q
qingqing01 已提交
1882
        super(RMSPropOptimizer, self).__init__(
X
Xin Pan 已提交
1883 1884 1885
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
Q
qingqing01 已提交
1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898
        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
1899
        self._centered = centered
Q
qingqing01 已提交
1900 1901 1902 1903 1904 1905 1906 1907

    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)
1908
            self._add_accumulator(self._mean_grad_acc_str, p)
Q
qingqing01 已提交
1909 1910 1911 1912 1913 1914 1915 1916 1917

    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])
1918 1919
        mean_grad_acc = self._get_accumulator(self._mean_grad_acc_str,
                                              param_and_grad[0])
Q
qingqing01 已提交
1920 1921 1922 1923 1924 1925 1926
        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,
1927
                "MeanGrad": mean_grad_acc,
Q
qingqing01 已提交
1928 1929 1930 1931 1932
                "LearningRate": self._create_param_lr(param_and_grad),
            },
            outputs={
                "ParamOut": param_and_grad[0],
                "MomentOut": momentum_acc,
1933 1934
                "MeanSquareOut": mean_square_acc,
                "MeanGradOut": mean_grad_acc
Q
qingqing01 已提交
1935 1936 1937 1938
            },
            attrs={
                "epsilon": self._epsilon,
                "decay": self._rho,
1939 1940
                "momentum": self._momentum,
                "centered": self._centered
M
minqiyang 已提交
1941 1942
            },
            stop_gradient=True)
Q
qingqing01 已提交
1943 1944 1945 1946

        return rmsprop_op


Q
qiaolongfei 已提交
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 1986 1987 1988
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
1989 1990 1991
        l1 (float): L1 regularization strength.
        l2 (float): L2 regularization strength.
        lr_power (float): Learning Rate Power.
X
Xin Pan 已提交
1992 1993 1994
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
Q
qiaolongfei 已提交
1995 1996 1997 1998 1999 2000 2001

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

    Examples:
          .. code-block:: python

2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025
            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
2026 2027 2028

    Notes:
       Currently, FtrlOptimizer doesn't support sparse parameter optimization.
Q
qiaolongfei 已提交
2029 2030 2031 2032 2033
    """

    _squared_acc_str = "squared"
    _linear_acc_str = "linear"

X
Xin Pan 已提交
2034 2035 2036 2037 2038 2039 2040
    def __init__(self,
                 learning_rate,
                 l1=0.0,
                 l2=0.0,
                 lr_power=-0.5,
                 regularization=None,
                 name=None):
Q
qiaolongfei 已提交
2041
        super(FtrlOptimizer, self).__init__(
X
Xin Pan 已提交
2042 2043 2044
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
Q
qiaolongfei 已提交
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 2082 2083 2084
        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
2085 2086
                   "lr_power": self._lr_power},
            stop_gradient=True)
Q
qiaolongfei 已提交
2087 2088 2089 2090

        return ftrl_op


Y
Yibing Liu 已提交
2091 2092 2093 2094 2095 2096
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 
Y
Yibing Liu 已提交
2097 2098
    correction. For more information, please refer to `Large Batch Optimization for 
    Deep Learning: Training BERT in 76 minutes <https://arxiv.org/abs/1904.00962>`_ .
Y
Yibing Liu 已提交
2099 2100 2101 2102 2103

    The updating of parameters follows:

    ..  math::

Y
Yibing Liu 已提交
2104
        m_t &= \\beta_1 m_{t - 1}+ (1 - \\beta_1)g_t \\
Y
Yibing Liu 已提交
2105

Y
Yibing Liu 已提交
2106
        v_t &= \\beta_2 v_{t - 1}  + (1 - \\beta_2)g_t^2 \\
Y
Yibing Liu 已提交
2107

Y
Yibing Liu 已提交
2108
        r_t &= \\frac{m_t}{\\sqrt{v_t}+\\epsilon} \\
Y
Yibing Liu 已提交
2109

Y
Yibing Liu 已提交
2110
        w_t &= w_{t-1} -\\eta_t \\frac{\\left \| w_{t-1}\\right \|}{\\left \| r_t + \\lambda w_{t-1}\\right \|} (r_t + \\lambda w_{t-1})
Y
Yibing Liu 已提交
2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123


    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.
Y
Yibing Liu 已提交
2124
        regularization (Regularizer): A Regularizer, such as
Y
Yibing Liu 已提交
2125
                        fluid.regularizer.L1DecayRegularizer.
Y
Yibing Liu 已提交
2126 2127
        exclude_from_weight_decay_fn (function): Exclude a parameter from weight 
            decay when **exclude_from_weight_decay_fn(parameter)** returns true.
Y
Yibing Liu 已提交
2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138
        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)

Y
Yibing Liu 已提交
2139 2140 2141 2142 2143
            def exclude_fn(param):
                return param.name.endswith('.b_0')

            optimizer = fluid.optimizer.Lamb(learning_rate=0.002,
                                             exclude_from_weight_decay_fn=exclude_fn)
Y
Yibing Liu 已提交
2144 2145 2146 2147
            optimizer.minimize(cost)
    """
    _moment1_acc_str = "moment1"
    _moment2_acc_str = "moment2"
Y
Yibing Liu 已提交
2148
    # these two not used in op temporarily
Y
Yibing Liu 已提交
2149 2150 2151 2152 2153 2154 2155 2156 2157 2158
    _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,
Y
Yibing Liu 已提交
2159
                 exclude_from_weight_decay_fn=None,
Y
Yibing Liu 已提交
2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174
                 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
Y
Yibing Liu 已提交
2175
        self._exclude_from_weight_decay_fn = exclude_from_weight_decay_fn
Y
Yibing Liu 已提交
2176 2177 2178

    def _append_optimize_op(self, block, param_and_grad):
        assert isinstance(block, framework.Block)
2179
        block.program._use_lamb = True
Y
Yibing Liu 已提交
2180 2181 2182 2183 2184 2185 2186 2187 2188 2189

        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])

Y
Yibing Liu 已提交
2190 2191 2192 2193 2194 2195
        if self._exclude_from_weight_decay_fn is not None \
            and self._exclude_from_weight_decay_fn(param_and_grad[0]):
            weight_decay = 0.0
        else:
            weight_decay = self._weight_decay

Y
Yibing Liu 已提交
2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216
        # 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,
Y
Yibing Liu 已提交
2217
                "weight_decay": weight_decay
Y
Yibing Liu 已提交
2218 2219 2220 2221 2222 2223
            },
            stop_gradient=True)

        return lamb_op


2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234 2235 2236 2237
# 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
2238
Adadelta = AdadeltaOptimizer
Q
qingqing01 已提交
2239
RMSProp = RMSPropOptimizer
Q
qiaolongfei 已提交
2240
Ftrl = FtrlOptimizer
2241
LarsMomentum = LarsMomentumOptimizer
Y
Yibing Liu 已提交
2242
Lamb = LambOptimizer
2243 2244 2245


class ModelAverage(Optimizer):
2246
    """Accumulate the average of parameters within sliding window. The average
2247 2248
    result will be saved in temporary variables which can be applied to
    parameter variables of current model by calling 'apply()' method. And the
2249
    'restore()' method is used to restore the parameter values of current model.
2250 2251 2252 2253 2254 2255 2256 2257

    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
2258 2259 2260
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
2261

2262
    Examples:
Q
qiaolongfei 已提交
2263 2264 2265

      .. code-block:: python

2266 2267 2268 2269 2270 2271
        import paddle.fluid as fluid
        import numpy

        # First create the Executor.
        place = fluid.CPUPlace()  # fluid.CUDAPlace(0)
        exe = fluid.Executor(place)
2272

2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294
        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
2295
            with model_average.apply(exe):
2296 2297 2298 2299
                x = numpy.random.random(size=(10, 1)).astype('float32')
                exe.run(program=train_program,
                        feed={'X': x},
                        fetch_list=[loss.name])
2300 2301 2302
    """

    def __init__(self,
W
wanghaoshuang 已提交
2303
                 average_window_rate,
2304 2305
                 min_average_window=10000,
                 max_average_window=10000,
X
Xin Pan 已提交
2306 2307 2308 2309
                 regularization=None,
                 name=None):
        super(ModelAverage, self).__init__(
            0.0, regularization=regularization, name=name)
2310 2311 2312
        self.average_window = average_window_rate
        self.min_average_window = min_average_window
        self.max_average_window = max_average_window
2313

2314
        self.params_grads = []
2315 2316
        for param in framework.default_main_program().global_block(
        ).all_parameters():
2317
            if param.do_model_average != False:
2318
                grad = param.block.create_var(
2319 2320
                    name=unique_name.generate_with_ignorable_key(".".join(
                        [param.name, 'tmp'])),
2321 2322
                    dtype=param.dtype,
                    persistable=False,
W
wanghaoshuang 已提交
2323
                    stop_gradient=True)
2324
                self.params_grads.append((param, grad))
2325

2326
        for param, grad in self.params_grads:
2327 2328
            if grad is None:
                continue
X
Xin Pan 已提交
2329 2330
            with param.block.program._optimized_guard(
                [param, grad]), name_scope('move_average'):
2331
                self._append_average_accumulate_op(param)
2332

2333 2334 2335 2336
        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:
2337
                self._add_average_apply_op(block, param_grad)
2338 2339 2340 2341 2342

        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:
2343
                self._add_average_restore_op(block, param_grad)
2344

2345
    def _add_average_apply_op(self, block, param_grad):
L
Luo Tao 已提交
2346 2347 2348 2349 2350 2351
        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(
2352
            self._get_accumulator('num_accumulates', param))
L
Luo Tao 已提交
2353
        old_num_accumulates = block._clone_variable(
2354
            self._get_accumulator('old_num_accumulates', param))
L
Luo Tao 已提交
2355
        num_updates = block._clone_variable(
2356 2357 2358 2359 2360 2361
            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
2362 2363 2364 2365
        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
2366
        ops._elementwise_div(x=sum, y=tmp, out=param)
2367 2368

    def _add_average_restore_op(self, block, param_grad):
L
Luo Tao 已提交
2369 2370
        param = block._clone_variable(param_grad[0])
        grad = block._clone_variable(param_grad[1])
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 2399 2400 2401 2402 2403 2404 2405 2406 2407
        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
2408 2409
            },
            stop_gradient=True)
2410

S
rename  
sneaxiy 已提交
2411
    @signature_safe_contextmanager
2412
    def apply(self, executor, need_restore=True):
2413
        """Apply average values to parameters of current model.
2414 2415 2416 2417

        Args:
            executor(fluid.Executor): current executor.
            need_restore(bool): If you finally need to do restore, set it to True. Default is True.
2418
        """
2419 2420 2421 2422 2423 2424
        executor.run(self.apply_program)
        try:
            yield
        finally:
            if need_restore:
                self.restore(executor)
2425 2426 2427

    def restore(self, executor):
        """Restore parameter values of current model.
2428 2429 2430
        
        Args:
            executor(fluid.Executor): current executor.
2431
        """
2432
        executor.run(self.restore_program)
2433 2434 2435 2436 2437 2438 2439 2440 2441 2442


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::

2443
        \\text{EMA}_0 & = 0
2444

2445 2446
	\\text{EMA}_t & = \\text{decay} * \\text{EMA}_{t-1} + (1 - \\text{decay}) * \\theta_t

Y
Yibing Liu 已提交
2447 2448 2449 2450
    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.
2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466 2467 2468 2469 2470 2471

    **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.
2472 2473 2474


    Args:
2475 2476 2477
	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.
2478 2479 2480 2481 2482 2483
	name (str|None): An optional name prefix.


    Examples:

	.. code-block:: python
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 2519 2520 2521 2522 2523 2524 2525 2526 2527

	    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)
2528 2529
    """

2530
    def __init__(self, decay=0.999, thres_steps=None, name=None):
2531
        self._decay = decay
2532
        self._thres_steps = thres_steps
2533
        self._name = name if name is not None else ''
2534 2535
        self._decay_var = self._get_ema_decay()

Y
Yibing Liu 已提交
2536
        self._params_tmps = []
2537
        for param in default_main_program().global_block().all_parameters():
2538 2539 2540 2541 2542 2543 2544
            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
2545
                self._params_tmps.append((param, tmp))
2546

Y
Yibing Liu 已提交
2547 2548
        self._ema_vars = {}
        for param, tmp in self._params_tmps:
2549 2550
            with param.block.program._optimized_guard(
                [param, tmp]), name_scope('moving_average'):
Y
Yibing Liu 已提交
2551
                self._ema_vars[param.name] = self._create_ema_vars(param)
2552 2553 2554 2555

        self.apply_program = Program()
        block = self.apply_program.global_block()
        with program_guard(main_program=self.apply_program):
2556
            decay_pow = self._get_decay_pow(block)
Y
Yibing Liu 已提交
2557
            for param, tmp in self._params_tmps:
2558 2559
                param = block._clone_variable(param)
                tmp = block._clone_variable(tmp)
Y
Yibing Liu 已提交
2560
                ema = block._clone_variable(self._ema_vars[param.name])
2561
                layers.assign(input=param, output=tmp)
2562 2563
                # bias correction
                ema = ema / (1.0 - decay_pow)
2564 2565 2566 2567 2568
                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
2569
            for param, tmp in self._params_tmps:
2570 2571 2572 2573
                tmp = block._clone_variable(tmp)
                param = block._clone_variable(param)
                layers.assign(input=tmp, output=param)

2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595 2596 2597 2598 2599 2600
    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
2601
    def _create_ema_vars(self, param):
2602 2603 2604 2605 2606 2607 2608 2609 2610
        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
2611 2612 2613 2614 2615
    def update(self):
        """ 
        Update Exponential Moving Average. Should only call this method in 
        train program.
        """
2616
        param_master_emas = []
Y
Yibing Liu 已提交
2617 2618 2619 2620
        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]
2621
                if param.name + '.master' in self._ema_vars:
2622 2623 2624 2625 2626 2627 2628 2629 2630 2631 2632 2633 2634 2635 2636 2637 2638
                    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
2639

2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651 2652 2653 2654 2655 2656 2657 2658 2659 2660 2661 2662
    @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
2663 2664 2665


class PipelineOptimizer(object):
2666 2667
    """
    Pipeline Optimizer
H
hutuxian 已提交
2668 2669 2670 2671 2672 2673 2674 2675 2676

    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, \
2677
    the final performance depends on the training progress of each pipeline heavily.
H
hutuxian 已提交
2678 2679 2680

    And we will try the synchronous mode in the future.

2681
    Args:
H
hutuxian 已提交
2682 2683 2684 2685
        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.
2686 2687
        queue_size (int): Each section will consume scopes from its in-scope queue 
                        and produce scopes to out-scope queue. And this parameter 
H
hutuxian 已提交
2688 2689 2690 2691
                        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].

2692 2693
    Examples:
        .. code-block:: python
H
hutuxian 已提交
2694

2695
            import paddle.fluid as fluid
H
hutuxian 已提交
2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713 2714 2715 2716 2717 2718 2719 2720 2721 2722 2723 2724 2725 2726 2727 2728 2729
            import paddle.fluid.layers as layers

            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)
2730 2731
    """

H
hutuxian 已提交
2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748
    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

H
hutuxian 已提交
2749
    def _create_vars(self, block, main_program):
H
hutuxian 已提交
2750 2751 2752 2753 2754 2755 2756 2757 2758 2759 2760
        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)

H
hutuxian 已提交
2761
    def _extract_section_opt_ops(self, ops, cut_point_name):
H
hutuxian 已提交
2762 2763 2764 2765 2766 2767 2768 2769 2770 2771 2772 2773 2774 2775 2776
        """
        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

H
hutuxian 已提交
2777
    def _find_input_output(self, ops, name, is_forward=True):
H
hutuxian 已提交
2778 2779 2780 2781 2782 2783 2784 2785 2786 2787 2788 2789 2790 2791
        """
        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

H
hutuxian 已提交
2792
    def _find_persistable_vars(self, ops, whole_parameters):
H
hutuxian 已提交
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
        """
        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

H
hutuxian 已提交
2820
    def _extract_section_ops(self, ops, cut_point_name):
H
hutuxian 已提交
2821 2822 2823 2824 2825 2826 2827 2828 2829 2830 2831 2832 2833 2834 2835 2836 2837 2838 2839
        """
        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

H
hutuxian 已提交
2840 2841
    def _find_section_opt(self, ops, params):
        res = self._extract_section_opt_ops(ops, params)
H
hutuxian 已提交
2842 2843
        return res

H
hutuxian 已提交
2844
    def _split_program(self, main_program, cut_list):
H
hutuxian 已提交
2845 2846 2847 2848 2849 2850 2851 2852 2853 2854 2855 2856 2857 2858 2859 2860 2861 2862 2863 2864
        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()
            }
H
hutuxian 已提交
2865
            cur_ops = self._extract_section_ops(ops, cut_vars)
H
hutuxian 已提交
2866 2867 2868 2869 2870 2871
            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(
H
hutuxian 已提交
2872
                self._find_input_output(
H
hutuxian 已提交
2873 2874 2875 2876 2877 2878
                    cur_ops, [], is_forward=True))
            for e in cur_ops:
                ops.remove(e)

            if i < cut_len:
                sec_params.append(
H
hutuxian 已提交
2879
                    self._find_persistable_vars(cur_ops, whole_parameters))
H
hutuxian 已提交
2880
            if i >= cut_len - 1:
H
hutuxian 已提交
2881 2882
                opt_ops = self._find_section_opt(
                    ops, sec_params[2 * cut_len - 2 - i])
H
hutuxian 已提交
2883 2884 2885 2886 2887 2888 2889 2890 2891 2892

                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(
H
hutuxian 已提交
2893
                self._find_input_output(
H
hutuxian 已提交
2894 2895 2896
                    cur_ops, cut_vars, is_forward=True))
            program["input_set"].update(sec_params[min(i, 2 * cut_len - 2 - i)])
            program["output_set"].update(
H
hutuxian 已提交
2897
                self._find_input_output(
H
hutuxian 已提交
2898 2899 2900 2901 2902 2903 2904 2905 2906 2907 2908 2909 2910 2911
                    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(
H
hutuxian 已提交
2912
            self._find_input_output(
H
hutuxian 已提交
2913 2914 2915 2916 2917 2918 2919 2920 2921 2922 2923 2924 2925 2926 2927 2928 2929 2930 2931 2932
                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
H
hutuxian 已提交
2933
        program_list = self._split_program(program, self._cut_list)
H
hutuxian 已提交
2934
        for p in program_list:
H
hutuxian 已提交
2935
            self._create_vars(p["program"].block(0), program)
H
hutuxian 已提交
2936 2937 2938 2939 2940 2941 2942 2943 2944 2945 2946 2947 2948 2949 2950 2951 2952 2953 2954 2955
        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
        }
M
mapingshuo 已提交
2956 2957 2958 2959 2960 2961 2962 2963 2964 2965 2966 2967 2968 2969 2970 2971 2972 2973 2974 2975 2976 2977 2978 2979 2980 2981 2982 2983 2984 2985 2986 2987 2988 2989 2990 2991 2992 2993 2994 2995 2996 2997 2998 2999 3000 3001 3002 3003 3004 3005 3006 3007 3008 3009 3010 3011 3012 3013 3014 3015 3016 3017 3018 3019 3020 3021 3022 3023 3024 3025 3026 3027 3028 3029 3030 3031 3032 3033 3034 3035 3036 3037 3038 3039 3040 3041 3042 3043 3044 3045 3046 3047 3048 3049 3050 3051 3052 3053 3054 3055 3056 3057 3058 3059 3060 3061 3062 3063 3064 3065 3066 3067 3068 3069 3070 3071 3072 3073 3074 3075 3076 3077 3078 3079 3080 3081 3082 3083 3084 3085 3086 3087 3088 3089 3090 3091 3092 3093 3094 3095 3096 3097 3098 3099 3100 3101 3102 3103 3104 3105 3106 3107 3108


class LookaheadOptimizer(object):
    """
    This implements the Lookahead optimizer of the
    paper : https://arxiv.org/abs/1907.08610.

    Lookahead keeps two sets of params: the fast_params and
    the slow_params. inner_optimizer update fast_params every 
    training step. Lookahead updates the slow_params and fast_params 
    every k training steps as follows:

    .. math::
        
        slow\_param_t &= slow\_param_{t-1} + \\alpha * (fast\_param_{t-1} - slow\_param_{t-1})
	
	fast\_param_t &=  slow\_param_t

    Args:
        inner_optimizer (Optimizer): The optimizer that update fast params step by step. 
        alpha (float): The learning rate of Lookahead.
        k (int): The slow params is updated every k steps.

    Examples:
        .. code-block:: python

            import paddle
            import paddle.fluid as fluid
            import numpy as np

	    x = fluid.layers.data(name='x', shape=[2], dtype='float32')
	    label = fluid.layers.data(name="label", shape=[1], dtype="int64")
	    y = fluid.layers.fc(input=[x], size=2, act="softmax")
	    loss = fluid.layers.cross_entropy(input=y, label=label)
	    loss = fluid.layers.mean(x=loss)
	    sgd = fluid.optimizer.SGD(learning_rate=0.01)
	    optimizer = fluid.optimizer.LookaheadOptimizer(sgd,
                                            alpha=0.5,
                                            k=5)
	    optimizer.minimize(loss)
	    main_program = fluid.default_main_program()
	    place = fluid.CPUPlace()
	    exe = fluid.Executor(place)
	    exe.run(fluid.default_startup_program())

	    feeder = fluid.DataFeeder(feed_list=[x, label], place=place)

	    step = 0
            while(step < 10):
                step += 1
		exe.run(fluid.default_main_program(),
            	feed=feeder.feed(batch_data))

    """

    def __init__(self, inner_optimizer, alpha=0.5, k=5):

        assert (inner_optimizer is not None), "inner optimizer can not be None"
        assert (
            0.0 <= alpha <= 1.0
        ), "alpha should be larger or equal to 0.0, and less or equal than 1.0"
        assert (isinstance(k, int) and k > 0), "k should be a positive integer"

        self.inner_optimizer = inner_optimizer
        self.alpha = alpha
        self.k = k
        self.type = "lookahead"

    def minimize(self, loss, startup_program=None):

        # Apply inner optimizer to the main_program
        mini_out = self.inner_optimizer.minimize(
            loss, startup_program=startup_program)

        # Get startup_program and main_program
        if startup_program is None:
            startup_program = default_startup_program()
        main_block = loss.block

        # add some vars to the main_program
        params = [param.name for param in main_block.all_parameters()]
        param_to_slow = {}
        for param in params:
            fast_var = main_block.var(param)
            assert (fast_var is not None)
            slow_var = main_block.create_var(
                name=param + "@SLOW",
                shape=fast_var.shape,
                dtype=fast_var.dtype,
                persistable=True)
            param_to_slow[param] = slow_var

        # add some vars to the startup_program
        startup_block = startup_program.global_block()
        for param in params:
            fast_var = startup_block.var(param)
            assert (fast_var is not None)
            slow_var = startup_block.create_var(
                name=param + "@SLOW",
                shape=fast_var.shape,
                dtype=fast_var.dtype,
                persistable=True)

            startup_block.append_op(
                type="assign",
                inputs={"X": fast_var},
                outputs={"Out": slow_var})

        # Add Var k to main prog and startup prog
        k = layers.create_global_var(
            name="lookahead_k",
            shape=[1],
            value=int(self.k),
            dtype='int32',
            persistable=True)

        # Add Var alpha to main prog and startup prog
        alpha = layers.create_global_var(
            name="lookahead_alpha",
            shape=[1],
            value=float(self.alpha),
            dtype='float32',
            persistable=True)

        # Add Var step
        step = layers.create_global_var(
            name="lookahead_step",
            shape=[1],
            value=int(0),
            dtype='int32',
            persistable=True)
        layers.increment(x=step, value=1.0, in_place=True)

        # lookahead
        zero_var = layers.fill_constant(shape=[1], dtype='float32', value=0.0)

        one_var = layers.fill_constant(shape=[1], dtype='float32', value=1.0)

        mod = layers.elementwise_mod(step, k)
        with layers.control_flow.Switch() as switch:
            with switch.case(mod == zero_var):
                for param_name in params:
                    fast_var = main_block.var(param_name)
                    slow_var = param_to_slow[param_name]
                    tmp_var = layers.elementwise_add(
                        layers.elementwise_mul(fast_var, alpha),
                        layers.elementwise_mul(
                            slow_var, layers.elementwise_sub(one_var, alpha)))
                    layers.assign(input=tmp_var, output=slow_var)
                    layers.assign(input=tmp_var, output=fast_var)
            with switch.default():
                pass
        return mini_out