optimizer.py 59.5 KB
Newer Older
1
# Copyright (c) 2018 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
from collections import defaultdict
18 19
from contextlib import contextmanager

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

23 24
from . import framework
from . import layers
25
from . import unique_name
26
from .backward import append_backward
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
M
minqiyang 已提交
33
from .imperative import base as imperative_base
34

35
__all__ = [
Q
qiaolongfei 已提交
36
    'SGD', 'Momentum', 'Adagrad', 'Adam', 'Adamax', 'DecayedAdagrad', 'Ftrl',
37
    'SGDOptimizer', 'MomentumOptimizer', 'AdagradOptimizer', 'AdamOptimizer',
W
weixing02 已提交
38
    'AdamaxOptimizer', 'DecayedAdagradOptimizer', 'RMSPropOptimizer',
39 40
    'FtrlOptimizer', 'Adadelta', 'ModelAverage', 'LarsMomentum',
    'LarsMomentumOptimizer'
41
]
Q
Qiao Longfei 已提交
42 43 44 45 46 47


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

    Define the common interface of an optimizer.
48 49
    User should not use this class directly,
    but need to use one of it's implementation.
Q
Qiao Longfei 已提交
50 51
    """

X
Xin Pan 已提交
52
    def __init__(self, learning_rate, regularization=None, name=None):
53
        if not isinstance(learning_rate, float) and \
54 55
                not isinstance(learning_rate, framework.Variable):
            raise TypeError("learning rate should be float or Variable")
W
whs 已提交
56
        self._name = name
D
dzhwinter 已提交
57
        self.regularization = regularization
58
        self._learning_rate = learning_rate
D
dzhwinter 已提交
59 60
        # the learning rate type should be inferenced from loss
        self._dtype = None
61
        # each program should have a independent learning rate
62
        # program -> Variable(learning_rate)
Q
qiaolongfei 已提交
63
        self._learning_rate_map = dict()
64
        if isinstance(self._learning_rate, framework.Variable):
65 66
            self._learning_rate_map[framework.default_main_program(
            )] = self._learning_rate
67 68 69 70 71
        # 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
72
        self.helper = None
Q
Qiao Longfei 已提交
73

Q
Qiao Longfei 已提交
74
    def _create_global_learning_rate(self):
Y
yuyang18 已提交
75
        lr = self._global_learning_rate()
Q
Qiao Longfei 已提交
76

77 78 79 80
        if isinstance(lr, framework.Variable):
            return
        else:
            if not isinstance(self._learning_rate, float):
Q
qiaolongfei 已提交
81
                raise TypeError(
82 83
                    "learning rate variable is create outside optimizer,"
                    "can not create new learning rate variable for new program")
Q
Qiao Longfei 已提交
84

85 86 87 88 89 90
        # 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),
Q
Qiao Longfei 已提交
91
            dtype='float32' if self._dtype is None else self._dtype,
92 93
            persistable=True)

Y
yuyang18 已提交
94
    def _global_learning_rate(self, program=None):
Q
Qiao Longfei 已提交
95 96 97 98
        """
        get global decayed learning rate
        :return:
        """
99 100
        if program is None:
            program = framework.default_main_program()
Q
qiaolongfei 已提交
101
        return self._learning_rate_map.get(program, None)
Q
Qiao Longfei 已提交
102

Q
Qiao Longfei 已提交
103 104 105 106 107
    def _append_optimize_op(self, block, param_and_grad):
        """ append optimize operator to block and return all the added optimize_op
        """
        raise NotImplementedError()

108 109 110 111
    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
112 113
        if type(param_lr) == Variable:
            return param_lr
Q
qiaolongfei 已提交
114
        else:
W
Wu Yi 已提交
115
            if param_lr == 1.0:
Y
yuyang18 已提交
116
                return self._global_learning_rate()
W
Wu Yi 已提交
117
            else:
X
Xin Pan 已提交
118 119 120
                with default_main_program()._lr_schedule_guard(
                        is_with_opt=True), framework.name_scope(
                            'scale_with_param_lr'):
121
                    return self._global_learning_rate() * param_lr
122 123 124 125 126 127 128

    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
129
        """
130 131
        pass

132
    def _finish_update(self, block, parameters_and_grads):
133 134 135 136 137 138 139 140
        """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
141
            None
142 143 144
        """
        pass

145 146 147 148 149 150
    def _add_accumulator(self,
                         name,
                         param,
                         dtype=None,
                         fill_value=0.0,
                         shape=None):
151 152 153 154 155 156 157 158 159
        """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
160 161
        if self._name is not None:
            name = self._name + "_" + name
162 163
        if (name in self._accumulators and
                param.name in self._accumulators[name]):
164
            raise Exception("Accumulator {} already exists for parameter {}".
165
                            format(name, param.name))
166 167
        if shape == None:
            shape = param.shape
Q
Qiao Longfei 已提交
168 169
        assert isinstance(self.helper, LayerHelper)
        var = self.helper.create_global_variable(
Y
Yu Yang 已提交
170
            name=unique_name.generate(name),
Q
Qiao Longfei 已提交
171
            persistable=True,
F
fengjiayi 已提交
172
            dtype=dtype or param.dtype,
Q
Qiao Longfei 已提交
173
            type=param.type,
174
            shape=shape)
Q
Qiao Longfei 已提交
175
        self.helper.set_variable_initializer(
176
            var, initializer=Constant(value=float(fill_value)))
Q
Qiao Longfei 已提交
177
        self._accumulators[name][param.name] = var
178
        return var
179 180 181 182 183 184 185 186 187 188 189

    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
190 191
        if self._name is not None:
            name = self._name + "_" + name
192 193 194 195 196 197
        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]

198
    def _create_optimization_pass(self, parameters_and_grads):
Q
Qiao Longfei 已提交
199 200 201
        """Add optimization operators to update gradients to variables.

        Args:
Q
qiaolongfei 已提交
202
          parameters_and_grads(list(tuple(Variable, Variable))):
203
            a list of (variable, gradient) pair to update.
Q
Qiao Longfei 已提交
204 205

        Returns:
206
          return_op_list: a list of operators that will complete one step of
207 208 209
            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
210
        """
211 212 213 214 215
        # 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
216
        # for parameters and extend _finish_update method to add custom ops.
217

218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244
        # 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 = []
        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)

        # 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
245 246 247 248 249 250 251 252 253
        """
        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
        """
254 255
        program = framework.default_main_program()
        global_block = framework.default_main_program().global_block()
Q
Qiao Longfei 已提交
256 257 258 259 260 261 262 263 264 265 266 267 268 269 270
        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:
271 272 273 274 275 276 277 278 279 280 281 282 283
            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
284 285
        return new_param_grads, (table_param, table_grad), sgd_op

286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303
    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
304

305 306
        Return:
            list: list of (param, grad) pair, grad is the output of backward.
M
minqiyang 已提交
307

308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324
        Examples:
            See examples in `apply_gradients`.
        """
        if callbacks is None:
            callbacks = [error_clip_callback]
        else:
            assert (isinstance(callbacks, list))
            callbacks.append(error_clip_callback)
        return append_backward(loss, parameter_list, no_grad_set, callbacks)

    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
325

326 327
        Returns:
            list: A list of operators appended to the current program.
M
minqiyang 已提交
328

329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356
        Examples:
            .. code-block:: python

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

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

        params_grads = append_gradient_clip_ops(params_grads)

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

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

        return optimize_ops

Q
Qiao Longfei 已提交
357 358
    def minimize(self,
                 loss,
359
                 startup_program=None,
Q
Qiao Longfei 已提交
360 361
                 parameter_list=None,
                 no_grad_set=None):
362 363 364 365 366
        """
        Add operations to minimize `loss` by updating `parameter_list`.

        This method combines interface `backward()` and
        `apply_gradients()` into one.
M
minqiyang 已提交
367

368 369 370 371 372 373
        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.
Q
Qiao Longfei 已提交
374

375 376 377
        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
378
        """
379 380 381
        self._dtype = loss.dtype
        program = loss.block.program
        optimize_ops = []
382
        if imperative_base.enabled():
M
minqiyang 已提交
383
            if parameter_list is not None:
M
minqiyang 已提交
384
                parameters = parameter_list
M
minqiyang 已提交
385 386
            else:
                parameters = program.global_block().all_parameters()
M
minqiyang 已提交
387 388 389

            params_grads = []
            for param in parameters:
M
minqiyang 已提交
390
                if param.stop_gradient or not param.trainable:
391
                    continue
M
minqiyang 已提交
392 393 394 395 396 397 398
                # 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))
399 400
            with program_guard(program, startup_program):
                optimize_ops = self._create_optimization_pass(params_grads)
M
minqiyang 已提交
401
        else:
402 403 404 405
            with program_guard(program, startup_program):
                params_grads = self.backward(loss, startup_program,
                                             parameter_list, no_grad_set)
                optimize_ops = self.apply_gradients(params_grads)
M
minqiyang 已提交
406

Q
Qiao Longfei 已提交
407
        return optimize_ops, params_grads
Q
Qiao Longfei 已提交
408 409 410


class SGDOptimizer(Optimizer):
Q
qiaolongfei 已提交
411 412 413 414 415 416 417 418 419 420
    """
    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
421 422 423
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
Q
qiaolongfei 已提交
424 425 426 427

    Examples:
        .. code-block:: python

Q
qiaolongfei 已提交
428
            sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.2)
Q
qiaolongfei 已提交
429
            sgd_optimizer.minimize(cost)
Q
Qiao Longfei 已提交
430 431
    """

X
Xin Pan 已提交
432
    def __init__(self, learning_rate, regularization=None, name=None):
Q
Qiao Longfei 已提交
433
        assert learning_rate is not None
Q
Qiao Longfei 已提交
434
        super(SGDOptimizer, self).__init__(
X
Xin Pan 已提交
435 436 437
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
Q
Qiao Longfei 已提交
438 439
        self.type = "sgd"

440 441
    def _append_optimize_op(self, block, param_and_grad):
        assert isinstance(block, framework.Block)
442

Q
Qiao Longfei 已提交
443 444 445 446 447 448
        # create the optimize op
        sgd_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
449
                "LearningRate": self._create_param_lr(param_and_grad)
Q
Qiao Longfei 已提交
450
            },
M
minqiyang 已提交
451 452
            outputs={"ParamOut": param_and_grad[0]},
            stop_gradient=True)
Q
Qiao Longfei 已提交
453 454

        return sgd_op
455 456 457


class MomentumOptimizer(Optimizer):
Q
qiaolongfei 已提交
458 459 460 461 462 463 464 465 466 467 468 469 470 471
    """

    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):

472
        &\quad   param = param - (gradient + mu * velocity) * learning\_rate
Q
qiaolongfei 已提交
473 474 475

        & else:

Q
qiaolongfei 已提交
476
        &\quad   param = param - learning\_rate * velocity
Q
qiaolongfei 已提交
477 478 479 480 481 482

    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
483 484 485
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
Q
qiaolongfei 已提交
486 487 488 489

    Examples:
        .. code-block:: python

Q
qiaolongfei 已提交
490
            optimizer = fluid.optimizer.Momentum(learning_rate=0.2, momentum=0.1)
Q
qiaolongfei 已提交
491
            optimizer.minimize(cost)
492 493 494
    """
    _velocity_acc_str = "velocity"

X
Xin Pan 已提交
495 496 497 498 499 500
    def __init__(self,
                 learning_rate,
                 momentum,
                 use_nesterov=False,
                 regularization=None,
                 name=None):
501 502
        assert learning_rate is not None
        assert momentum is not None
Q
Qiao Longfei 已提交
503
        super(MomentumOptimizer, self).__init__(
X
Xin Pan 已提交
504 505 506
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
507 508
        self.type = "momentum"
        self._momentum = momentum
509
        self._use_nesterov = bool(use_nesterov)
510 511 512 513 514

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

        for p in parameters:
Q
Qiao Longfei 已提交
515
            self._add_accumulator(self._velocity_acc_str, p)
516 517 518 519 520 521 522 523 524 525 526 527 528

    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,
529
                "LearningRate": self._create_param_lr(param_and_grad)
530 531 532 533 534
            },
            outputs={
                "ParamOut": param_and_grad[0],
                "VelocityOut": velocity_acc
            },
535
            attrs={"mu": self._momentum,
M
minqiyang 已提交
536 537
                   "use_nesterov": self._use_nesterov},
            stop_gradient=True)
538 539

        return momentum_op
540 541


542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565
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
566

567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621

    Examples:
        .. code-block:: python

            optimizer = fluid.optimizer.LarsMomentum(learning_rate=0.2, momentum=0.1, lars_weight_decay=0.001)
            optimizer.minimize(cost)
    """
    _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
622 623
            },
            stop_gradient=True)
624 625 626 627

        return momentum_op


628
class AdagradOptimizer(Optimizer):
Q
qiaolongfei 已提交
629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648
    """
    **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
649 650 651
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
Q
qiaolongfei 已提交
652 653 654 655 656 657

    Examples:
        .. code-block:: python

            optimizer = fluid.optimizer.Adagrad(learning_rate=0.2)
            optimizer.minimize(cost)
658 659 660
    """
    _moment_acc_str = "moment"

X
Xin Pan 已提交
661 662 663 664
    def __init__(self,
                 learning_rate,
                 epsilon=1.0e-6,
                 regularization=None,
665 666
                 name=None,
                 initial_accumulator_value=0.1):
667 668
        assert learning_rate is not None
        assert epsilon is not None
Q
Qiao Longfei 已提交
669
        super(AdagradOptimizer, self).__init__(
X
Xin Pan 已提交
670 671 672
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
673 674
        self.type = "adagrad"
        self._epsilon = epsilon
675
        self.initial_accumulator_value = initial_accumulator_value
676 677 678 679 680

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

        for p in parameters:
Q
Qiao Longfei 已提交
681
            self._add_accumulator(self._moment_acc_str, p)
682 683 684 685 686 687

    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])
688 689 690 691 692 693 694 695 696 697
        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,
            })
698

699
        # Create the adagrad optimizer op
700 701 702 703 704 705
        adagrad_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
                "Moment": moment_acc,
706
                "LearningRate": self._create_param_lr(param_and_grad)
707 708 709
            },
            outputs={"ParamOut": param_and_grad[0],
                     "MomentOut": moment_acc},
M
minqiyang 已提交
710 711
            attrs={"epsilon": self._epsilon},
            stop_gradient=True)
712 713

        return adagrad_op
714 715 716


class AdamOptimizer(Optimizer):
Q
qiaolongfei 已提交
717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743
    """
    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.
744
        regularization: A Regularizer, such as fluid.regularizer.L2DecayRegularizer.
X
Xin Pan 已提交
745
        name: A optional name prefix.
746 747 748 749 750 751
        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
752 753 754 755 756 757 758

    Examples:
        .. code-block:: python

            optimizer = fluid.optimizer.Adam(learning_rate=0.2)
            optimizer.minimize(cost)

759 760 761
    """
    _moment1_acc_str = "moment1"
    _moment2_acc_str = "moment2"
Q
qiaolongfei 已提交
762 763
    _beta1_pow_acc_str = "beta1_pow_acc"
    _beta2_pow_acc_str = "beta2_pow_acc"
764 765 766 767 768

    def __init__(self,
                 learning_rate=0.001,
                 beta1=0.9,
                 beta2=0.999,
769
                 epsilon=1e-8,
X
Xin Pan 已提交
770
                 regularization=None,
Q
Qiao Longfei 已提交
771
                 name=None,
Q
Qiao Longfei 已提交
772
                 lazy_mode=False):
773 774 775 776
        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
777
        super(AdamOptimizer, self).__init__(
X
Xin Pan 已提交
778 779 780
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
781 782 783 784
        self.type = "adam"
        self._beta1 = beta1
        self._beta2 = beta2
        self._epsilon = epsilon
Q
Qiao Longfei 已提交
785
        self._lazy_mode = lazy_mode
786 787 788 789 790 791

    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
792 793
            self._add_accumulator(self._moment1_acc_str, p)
            self._add_accumulator(self._moment2_acc_str, p)
Q
qiaolongfei 已提交
794 795 796 797 798 799 800 801 802 803 804 805
            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])
806 807 808 809 810 811 812 813

    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
814 815 816 817 818
        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])

819
        # create the adam optimize op
820 821 822 823 824
        adam_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
825
                "LearningRate": self._create_param_lr(param_and_grad),
826 827
                "Moment1": moment1,
                "Moment2": moment2,
Q
qiaolongfei 已提交
828 829
                "Beta1Pow": beta1_pow_acc,
                "Beta2Pow": beta2_pow_acc
830 831 832 833 834 835 836 837 838
            },
            outputs={
                "ParamOut": param_and_grad[0],
                "Moment1Out": moment1,
                "Moment2Out": moment2
            },
            attrs={
                "beta1": self._beta1,
                "beta2": self._beta2,
Q
Qiao Longfei 已提交
839
                "epsilon": self._epsilon,
840 841
                "lazy_mode": self._lazy_mode,
                "min_row_size_to_use_multithread": 1000
M
minqiyang 已提交
842 843
            },
            stop_gradient=True)
844 845 846

        return adam_op

847
    def _finish_update(self, block, param_and_grads):
848 849 850
        """Update Beta1 and Beta2 Power accumulators
        """
        assert isinstance(block, framework.Block)
Q
Qiao Longfei 已提交
851
        main_block = block.program.global_block()
852 853 854
        for param, grad in param_and_grads:
            if grad is None:
                continue
X
Xin Pan 已提交
855 856
            with param.block.program._optimized_guard(
                [param, grad]), name_scope("optimizer"):
857 858 859 860 861 862 863 864
                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
865 866
                    attrs={"scale": self._beta1},
                    stop_gradient=True)
867 868 869 870 871

                main_block.append_op(
                    type="scale",
                    inputs={"X": beta2_pow_acc},
                    outputs={"Out": beta2_pow_acc},
M
minqiyang 已提交
872 873
                    attrs={"scale": self._beta2},
                    stop_gradient=True)
874 875 876


class AdamaxOptimizer(Optimizer):
Q
qiaolongfei 已提交
877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906
    """
    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.

    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
907 908 909
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
Q
qiaolongfei 已提交
910 911 912 913 914 915

    Examples:
        .. code-block:: python

            optimizer = fluid.optimizer.Adamax(learning_rate=0.2)
            optimizer.minimize(cost)
C
chengduo 已提交
916 917 918

    Notes:
       Currently, AdamaxOptimizer doesn't support sparse parameter optimization.
919 920 921
    """
    _moment_acc_str = "moment"
    _inf_norm_acc_str = "inf_norm"
Q
qiaolongfei 已提交
922
    _beta1_pow_acc_str = "beta1_pow_acc"
923 924 925 926 927

    def __init__(self,
                 learning_rate=0.001,
                 beta1=0.9,
                 beta2=0.999,
928
                 epsilon=1e-8,
X
Xin Pan 已提交
929 930
                 regularization=None,
                 name=None):
931 932 933 934
        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
935
        super(AdamaxOptimizer, self).__init__(
X
Xin Pan 已提交
936 937 938
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
939 940 941 942 943 944 945 946
        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
947 948
            self._add_accumulator(self._moment_acc_str, p)
            self._add_accumulator(self._inf_norm_acc_str, p)
Q
qiaolongfei 已提交
949 950 951 952 953 954
            self._add_accumulator(
                name=self._beta1_pow_acc_str,
                param=p,
                dtype='float32',
                fill_value=self._beta1,
                shape=[1])
955 956 957 958 959 960 961

    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
962 963
        beta1_pow_acc = self._get_accumulator(self._beta1_pow_acc_str,
                                              param_and_grad[0])
964 965 966 967 968 969
        # create the adamax optimize op
        adamax_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
970
                "LearningRate": self._create_param_lr(param_and_grad),
971 972
                "Moment": moment,
                "InfNorm": inf_norm,
Q
qiaolongfei 已提交
973
                "Beta1Pow": beta1_pow_acc
974 975 976 977 978 979 980 981 982 983
            },
            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
984 985
            },
            stop_gradient=True)
986 987 988

        return adamax_op

989
    def _finish_update(self, block, parameters_and_grads):
990 991 992
        """Update Beta1 Power accumulator
        """
        assert isinstance(block, framework.Block)
Q
Qiao Longfei 已提交
993
        main_block = block.program.global_block()
994 995 996
        for param, grad in parameters_and_grads:
            if grad is None:
                continue
X
Xin Pan 已提交
997 998
            with param.block.program._optimized_guard(
                [param, grad]), name_scope('adamx'):
999 1000 1001 1002 1003 1004
                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
1005 1006
                    attrs={"scale": self._beta1},
                    stop_gradient=True)
1007 1008 1009


class DecayedAdagradOptimizer(Optimizer):
1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031
    """
    **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
1032 1033 1034
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
1035 1036 1037 1038 1039 1040

    Examples:
        .. code-block:: python

            optimizer = fluid.optimizer.DecayedAdagrad(learning_rate=0.2)
            optimizer.minimize(cost)
C
chengduo 已提交
1041 1042 1043

    Notes:
       Currently, DecayedAdagradOptimizer doesn't support sparse parameter optimization.
1044 1045 1046
    """
    _moment_acc_str = "moment"

X
Xin Pan 已提交
1047 1048 1049 1050 1051 1052
    def __init__(self,
                 learning_rate,
                 decay=0.95,
                 epsilon=1.0e-6,
                 regularization=None,
                 name=None):
1053 1054 1055 1056
        assert learning_rate is not None
        assert decay is not None
        assert epsilon is not None

Q
Qiao Longfei 已提交
1057
        super(DecayedAdagradOptimizer, self).__init__(
X
Xin Pan 已提交
1058 1059 1060
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087
        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
1088 1089
            attrs={"epsilon": self._epsilon},
            stop_gradient=True)
1090 1091

        return decayed_adagrad_op
1092 1093


1094
class AdadeltaOptimizer(Optimizer):
1095 1096
    """
    **Adadelta Optimizer**
Q
qiaolongfei 已提交
1097

1098
    Simple Adadelta optimizer with average squared grad state and
1099
    average squared update state.
1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111
    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
1112
        learning_rate(float): global learning rate
1113 1114
        rho(float): rho in equation
        epsilon(float): epsilon in equation
X
Xin Pan 已提交
1115 1116 1117
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
1118 1119 1120 1121 1122 1123 1124

    Examples:
        .. code-block:: python

            optimizer = fluid.optimizer.Adadelta(
                learning_rate=0.0003, epsilon=1.0e-6, rho=0.95)
            _, params_grads = optimizer.minimize(cost)
C
chengduo 已提交
1125 1126 1127

    Notes:
       Currently, AdadeltaOptimizer doesn't support sparse parameter optimization.
1128
    """
1129

1130 1131 1132
    _avg_squared_grad_acc_str = "_avg_squared_grad"
    _avg_squared_update_acc_str = "_avg_squared_update"

X
Xin Pan 已提交
1133 1134 1135 1136 1137 1138
    def __init__(self,
                 learning_rate,
                 epsilon=1.0e-6,
                 rho=0.95,
                 regularization=None,
                 name=None):
1139 1140 1141 1142 1143 1144
        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.")
1145
        super(AdadeltaOptimizer, self).__init__(
X
Xin Pan 已提交
1146 1147 1148
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
1149 1150 1151 1152 1153
        self.type = "adadelta"
        self._epsilon = epsilon
        self._rho = rho

    def _create_accumulators(self, block, parameters):
1154 1155
        if not isinstance(block, framework.Block):
            raise TypeError("block is not instance of framework.Block.")
1156 1157 1158 1159 1160 1161

        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):
1162 1163
        if not isinstance(block, framework.Block):
            raise TypeError("block is not instance of framework.Block.")
1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184

        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
1185 1186
                   "rho": self._rho},
            stop_gradient=True)
1187 1188 1189 1190

        return adadelta_op


Q
qingqing01 已提交
1191 1192 1193 1194 1195 1196 1197 1198 1199 1200
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
1201
        r(w, t) & = \\rho r(w, t-1) + (1 - \\rho)(\\nabla Q_{i}(w))^2
Q
qingqing01 已提交
1202 1203 1204 1205

        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
1206
    each weight. Then dividing the gradient by :math:`sqrt{v(w,t)}`.
Q
qingqing01 已提交
1207 1208 1209 1210 1211 1212

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

    ..  math::

Q
qiaolongfei 已提交
1213
        r(w, t) & = \\rho r(w, t-1) + (1 - \\rho)(\\nabla Q_{i}(w))^2
Q
qingqing01 已提交
1214

1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228
        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
1229 1230 1231 1232
            \\epsilon}} \\nabla Q_{i}(w)

        w & = w - v(w, t)

Q
qiaolongfei 已提交
1233
    where, :math:`\\rho` is a hyperparameter and typical values are 0.9, 0.95
Q
qingqing01 已提交
1234 1235 1236 1237 1238 1239
    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
1240
        learning_rate(float): global learning rate.
Q
qingqing01 已提交
1241 1242 1243
        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
1244
        momentum(float): :math:`\\beta` in equation is the momentum term,
Q
qingqing01 已提交
1245
            set 0.0 by default.
1246 1247 1248 1249
        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
1250 1251 1252
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
Q
qingqing01 已提交
1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265

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

    Examples:
          .. code-block:: python

              optimizer = fluid.optimizer.RMSProp(0.0001)
              _, params_grads = optimizer.minimize(cost)
    """

    _momentum_acc_str = "momentum"
    _mean_square_acc_str = "mean_square"
1266
    _mean_grad_acc_str = "mean_grad"
Q
qingqing01 已提交
1267 1268 1269 1270 1271 1272

    def __init__(self,
                 learning_rate,
                 rho=0.95,
                 epsilon=1.0e-6,
                 momentum=0.0,
1273
                 centered=False,
X
Xin Pan 已提交
1274 1275
                 regularization=None,
                 name=None):
Q
qingqing01 已提交
1276
        super(RMSPropOptimizer, self).__init__(
X
Xin Pan 已提交
1277 1278 1279
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
Q
qingqing01 已提交
1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292
        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
1293
        self._centered = centered
Q
qingqing01 已提交
1294 1295 1296 1297 1298 1299 1300 1301

    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)
1302
            self._add_accumulator(self._mean_grad_acc_str, p)
Q
qingqing01 已提交
1303 1304 1305 1306 1307 1308 1309 1310 1311

    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])
1312 1313
        mean_grad_acc = self._get_accumulator(self._mean_grad_acc_str,
                                              param_and_grad[0])
Q
qingqing01 已提交
1314 1315 1316 1317 1318 1319 1320
        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,
1321
                "MeanGrad": mean_grad_acc,
Q
qingqing01 已提交
1322 1323 1324 1325 1326
                "LearningRate": self._create_param_lr(param_and_grad),
            },
            outputs={
                "ParamOut": param_and_grad[0],
                "MomentOut": momentum_acc,
1327 1328
                "MeanSquareOut": mean_square_acc,
                "MeanGradOut": mean_grad_acc
Q
qingqing01 已提交
1329 1330 1331 1332
            },
            attrs={
                "epsilon": self._epsilon,
                "decay": self._rho,
1333 1334
                "momentum": self._momentum,
                "centered": self._centered
M
minqiyang 已提交
1335 1336
            },
            stop_gradient=True)
Q
qingqing01 已提交
1337 1338 1339 1340

        return rmsprop_op


Q
qiaolongfei 已提交
1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385
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.
        l1 (float):
        l2 (float):
        lr_power (float):
X
Xin Pan 已提交
1386 1387 1388
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
Q
qiaolongfei 已提交
1389 1390 1391 1392 1393 1394 1395 1396 1397

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

    Examples:
          .. code-block:: python

              optimizer = fluid.optimizer.Ftrl(0.0001)
              _, params_grads = optimizer.minimize(cost)
C
chengduo 已提交
1398 1399 1400

    Notes:
       Currently, FtrlOptimizer doesn't support sparse parameter optimization.
Q
qiaolongfei 已提交
1401 1402 1403 1404 1405
    """

    _squared_acc_str = "squared"
    _linear_acc_str = "linear"

X
Xin Pan 已提交
1406 1407 1408 1409 1410 1411 1412
    def __init__(self,
                 learning_rate,
                 l1=0.0,
                 l2=0.0,
                 lr_power=-0.5,
                 regularization=None,
                 name=None):
Q
qiaolongfei 已提交
1413
        super(FtrlOptimizer, self).__init__(
X
Xin Pan 已提交
1414 1415 1416
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
Q
qiaolongfei 已提交
1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456
        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
1457 1458
                   "lr_power": self._lr_power},
            stop_gradient=True)
Q
qiaolongfei 已提交
1459 1460 1461 1462

        return ftrl_op


1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476
# 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
1477
Adadelta = AdadeltaOptimizer
Q
qingqing01 已提交
1478
RMSProp = RMSPropOptimizer
Q
qiaolongfei 已提交
1479
Ftrl = FtrlOptimizer
1480
LarsMomentum = LarsMomentumOptimizer
1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495


class ModelAverage(Optimizer):
    """Accumulate the average of parameters whtin sliding window. The average
    result will be saved in temporary variables which can be applied to
    parameter variables of current model by calling 'apply()' method. And the
    'restore()' method is used to restored the parameter values of current model.

    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
1496 1497 1498
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
1499
    Examples:
Q
qiaolongfei 已提交
1500 1501 1502

      .. code-block:: python

1503
        optimizer = fluid.optimizer.Momentum()
1504 1505
        optimizer.minimize(cost)
        model_average = fluid.optimizer.ModelAverage(0.15,
1506 1507 1508 1509 1510
                                                min_average_window=10000,
                                                max_average_window=20000)
        for pass_id in range(args.pass_num):
            for data in train_reader():
                exe.run(fluid.default_main_program()...)
1511 1512 1513 1514

            with model_average.apply(exe):
                for data in test_reader():
                    exe.run(inference_program...)
1515 1516 1517
    """

    def __init__(self,
W
wanghaoshuang 已提交
1518
                 average_window_rate,
1519 1520
                 min_average_window=10000,
                 max_average_window=10000,
X
Xin Pan 已提交
1521 1522 1523 1524
                 regularization=None,
                 name=None):
        super(ModelAverage, self).__init__(
            0.0, regularization=regularization, name=name)
1525 1526 1527
        self.average_window = average_window_rate
        self.min_average_window = min_average_window
        self.max_average_window = max_average_window
1528

1529
        self.params_grads = []
1530 1531
        for param in framework.default_main_program().global_block(
        ).all_parameters():
1532
            if param.do_model_average != False:
1533 1534 1535 1536
                grad = param.block.create_var(
                    name=unique_name.generate(".".join([param.name, 'tmp'])),
                    dtype=param.dtype,
                    persistable=False,
W
wanghaoshuang 已提交
1537
                    stop_gradient=True)
1538
                self.params_grads.append((param, grad))
1539

1540
        for param, grad in self.params_grads:
1541 1542
            if grad is None:
                continue
X
Xin Pan 已提交
1543 1544
            with param.block.program._optimized_guard(
                [param, grad]), name_scope('move_average'):
1545
                self._append_average_accumulate_op(param)
1546

1547 1548 1549 1550
        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:
1551
                self._add_average_apply_op(block, param_grad)
1552 1553 1554 1555 1556

        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:
1557
                self._add_average_restore_op(block, param_grad)
1558

1559
    def _add_average_apply_op(self, block, param_grad):
L
Luo Tao 已提交
1560 1561 1562 1563 1564 1565
        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(
1566
            self._get_accumulator('num_accumulates', param))
L
Luo Tao 已提交
1567
        old_num_accumulates = block._clone_variable(
1568
            self._get_accumulator('old_num_accumulates', param))
L
Luo Tao 已提交
1569
        num_updates = block._clone_variable(
1570 1571 1572 1573 1574 1575
            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
1576 1577 1578 1579
        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
1580
        ops._elementwise_div(x=sum, y=tmp, out=param)
1581 1582

    def _add_average_restore_op(self, block, param_grad):
L
Luo Tao 已提交
1583 1584
        param = block._clone_variable(param_grad[0])
        grad = block._clone_variable(param_grad[1])
1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621
        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
1622 1623
            },
            stop_gradient=True)
1624

1625 1626
    @contextmanager
    def apply(self, executor, need_restore=True):
1627 1628
        """Apply average values to parameters of current model.
        """
1629 1630 1631 1632 1633 1634
        executor.run(self.apply_program)
        try:
            yield
        finally:
            if need_restore:
                self.restore(executor)
1635 1636 1637 1638

    def restore(self, executor):
        """Restore parameter values of current model.
        """
1639
        executor.run(self.restore_program)