optimizer.py 72.2 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
S
rename  
sneaxiy 已提交
18
from .wrapped_decorator import signature_safe_contextmanager
19

20
from paddle.fluid.framework import Program, Variable, name_scope, default_main_program, default_startup_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
33 34
from .dygraph import base as imperative_base
from .dygraph.learning_rate_scheduler import LearningRateDecay
35 36 37 38
from paddle.fluid import core
from paddle.fluid.layers import tensor
from functools import reduce
import copy
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'
46
]
Q
Qiao Longfei 已提交
47 48 49 50 51 52


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

    Define the common interface of an optimizer.
53 54
    User should not use this class directly,
    but need to use one of it's implementation.
Q
Qiao Longfei 已提交
55 56
    """

X
Xin Pan 已提交
57
    def __init__(self, learning_rate, regularization=None, name=None):
M
minqiyang 已提交
58
        if framework._in_dygraph_mode():
M
minqiyang 已提交
59 60 61 62 63 64 65 66 67 68 69 70
            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))
        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))

W
whs 已提交
71
        self._name = name
D
dzhwinter 已提交
72
        self.regularization = regularization
73
        self._learning_rate = learning_rate
D
dzhwinter 已提交
74 75
        # the learning rate type should be inferenced from loss
        self._dtype = None
76
        # each program should have a independent learning rate
77
        # program -> Variable(learning_rate)
Q
qiaolongfei 已提交
78
        self._learning_rate_map = dict()
79
        if isinstance(self._learning_rate, framework.Variable):
80 81
            self._learning_rate_map[framework.default_main_program(
            )] = self._learning_rate
82 83 84 85 86
        # 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
87
        self.helper = None
88 89 90 91
        self._opti_name_list = []

    def get_opti_var_name_list(self):
        return self._opti_name_list
Q
Qiao Longfei 已提交
92

Q
Qiao Longfei 已提交
93
    def _create_global_learning_rate(self):
94 95 96
        if imperative_base.enabled():
            # create learning rate Variable
            if isinstance(self._learning_rate, float):
M
minqiyang 已提交
97 98 99 100 101 102 103 104 105 106 107 108
                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)
109
            # get learning rate Variable from LearningRateDecay
M
minqiyang 已提交
110
            elif isinstance(self._learning_rate, LearningRateDecay):
111 112 113 114 115 116
                self._learning_rate_map[framework.default_main_program(
                )] = self._learning_rate()
            else:
                raise TypeError(
                    "optimizer's learning rate must be float or LearningRateDecay"
                )
117
        else:
118 119 120 121
            lr = self._global_learning_rate()

            if isinstance(lr, framework.Variable):
                return
M
minqiyang 已提交
122 123 124 125 126 127
            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
128

129 130 131 132 133 134 135 136
            # 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)
137

Y
yuyang18 已提交
138
    def _global_learning_rate(self, program=None):
Q
Qiao Longfei 已提交
139 140 141 142
        """
        get global decayed learning rate
        :return:
        """
143 144
        if program is None:
            program = framework.default_main_program()
Q
qiaolongfei 已提交
145
        return self._learning_rate_map.get(program, None)
Q
Qiao Longfei 已提交
146

Q
Qiao Longfei 已提交
147 148 149 150 151
    def _append_optimize_op(self, block, param_and_grad):
        """ append optimize operator to block and return all the added optimize_op
        """
        raise NotImplementedError()

152 153 154 155
    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
156 157
        if type(param_lr) == Variable:
            return param_lr
Q
qiaolongfei 已提交
158
        else:
W
Wu Yi 已提交
159
            if param_lr == 1.0:
Y
yuyang18 已提交
160
                return self._global_learning_rate()
W
Wu Yi 已提交
161
            else:
X
Xin Pan 已提交
162 163 164
                with default_main_program()._lr_schedule_guard(
                        is_with_opt=True), framework.name_scope(
                            'scale_with_param_lr'):
165
                    return self._global_learning_rate() * param_lr
166 167 168 169 170 171 172

    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
173
        """
174 175
        pass

176
    def _finish_update(self, block, parameters_and_grads):
177 178 179 180 181 182 183 184
        """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
185
            None
186 187 188
        """
        pass

189 190 191 192 193 194
    def _add_accumulator(self,
                         name,
                         param,
                         dtype=None,
                         fill_value=0.0,
                         shape=None):
195 196 197 198 199 200 201 202 203
        """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
204 205
        if self._name is not None:
            name = self._name + "_" + name
206 207
        if (name in self._accumulators and
                param.name in self._accumulators[name]):
L
lujun 已提交
208
            if framework._in_dygraph_mode():
M
minqiyang 已提交
209
                return self._accumulators[name][param.name]
210
            raise Exception("Accumulator {} already exists for parameter {}".
211
                            format(name, param.name))
212 213
        if shape == None:
            shape = param.shape
Q
Qiao Longfei 已提交
214
        assert isinstance(self.helper, LayerHelper)
215 216 217 218 219

        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
220
        var = self.helper.create_global_variable(
221
            name=var_name,
Q
Qiao Longfei 已提交
222
            persistable=True,
F
fengjiayi 已提交
223
            dtype=dtype or param.dtype,
Q
Qiao Longfei 已提交
224
            type=param.type,
225
            shape=shape)
Q
Qiao Longfei 已提交
226
        self.helper.set_variable_initializer(
227
            var, initializer=Constant(value=float(fill_value)))
Q
Qiao Longfei 已提交
228
        self._accumulators[name][param.name] = var
229
        return var
230 231 232 233 234 235 236 237 238 239 240

    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
241 242
        if self._name is not None:
            name = self._name + "_" + name
243 244 245 246 247 248
        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]

249
    def _create_optimization_pass(self, parameters_and_grads):
Q
Qiao Longfei 已提交
250 251 252
        """Add optimization operators to update gradients to variables.

        Args:
Q
qiaolongfei 已提交
253
          parameters_and_grads(list(tuple(Variable, Variable))):
254
            a list of (variable, gradient) pair to update.
Q
Qiao Longfei 已提交
255 256

        Returns:
257
          return_op_list: a list of operators that will complete one step of
258 259 260
            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
261
        """
262 263 264 265 266
        # 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
267
        # for parameters and extend _finish_update method to add custom ops.
268

269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295
        # 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
296 297 298 299 300 301 302 303 304
        """
        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
        """
305 306
        program = framework.default_main_program()
        global_block = framework.default_main_program().global_block()
Q
Qiao Longfei 已提交
307 308 309 310 311 312 313 314 315 316 317 318 319 320 321
        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:
322 323 324 325 326 327 328 329 330 331 332 333 334
            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
335 336
        return new_param_grads, (table_param, table_grad), sgd_op

337 338 339
    def _append_dgc_ops(self, param_and_grad):
        pass

340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357
    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
358

359 360
        Return:
            list: list of (param, grad) pair, grad is the output of backward.
M
minqiyang 已提交
361

362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378
        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
379

380 381
        Returns:
            list: A list of operators appended to the current program.
M
minqiyang 已提交
382

383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410
        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
411 412
    def minimize(self,
                 loss,
413
                 startup_program=None,
Q
Qiao Longfei 已提交
414 415
                 parameter_list=None,
                 no_grad_set=None):
416 417 418 419 420
        """
        Add operations to minimize `loss` by updating `parameter_list`.

        This method combines interface `backward()` and
        `apply_gradients()` into one.
M
minqiyang 已提交
421

422 423 424 425 426 427
        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
428

429 430 431
        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
432
        """
433 434
        self._dtype = loss.dtype
        optimize_ops = []
L
lujun 已提交
435
        if framework._in_dygraph_mode():
M
minqiyang 已提交
436
            if parameter_list is not None:
M
minqiyang 已提交
437
                parameters = parameter_list
M
minqiyang 已提交
438
            else:
L
lujun 已提交
439
                parameters = framework._dygraph_tracer().all_parameters()
M
minqiyang 已提交
440 441 442

            params_grads = []
            for param in parameters:
443
                if not param.trainable:
444
                    continue
445 446 447 448 449 450 451 452
                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))
453 454
            with program_guard(framework.default_main_program(),
                               framework.default_startup_program()):
455
                optimize_ops = self._create_optimization_pass(params_grads)
M
minqiyang 已提交
456
        else:
457
            program = loss.block.program
458 459 460
            with program_guard(program, startup_program):
                params_grads = self.backward(loss, startup_program,
                                             parameter_list, no_grad_set)
461 462 463
                # 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)
464
                optimize_ops = self.apply_gradients(params_grads)
M
minqiyang 已提交
465

Q
Qiao Longfei 已提交
466
        return optimize_ops, params_grads
Q
Qiao Longfei 已提交
467 468 469


class SGDOptimizer(Optimizer):
Q
qiaolongfei 已提交
470 471 472 473 474 475 476 477 478 479
    """
    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
480 481 482
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
Q
qiaolongfei 已提交
483 484 485 486

    Examples:
        .. code-block:: python

Q
qiaolongfei 已提交
487
            sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.2)
Q
qiaolongfei 已提交
488
            sgd_optimizer.minimize(cost)
Q
Qiao Longfei 已提交
489 490
    """

X
Xin Pan 已提交
491
    def __init__(self, learning_rate, regularization=None, name=None):
Q
Qiao Longfei 已提交
492
        assert learning_rate is not None
Q
Qiao Longfei 已提交
493
        super(SGDOptimizer, self).__init__(
X
Xin Pan 已提交
494 495 496
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
Q
Qiao Longfei 已提交
497 498
        self.type = "sgd"

499 500
    def _append_optimize_op(self, block, param_and_grad):
        assert isinstance(block, framework.Block)
501

Q
Qiao Longfei 已提交
502 503 504 505 506 507
        # create the optimize op
        sgd_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
508
                "LearningRate": self._create_param_lr(param_and_grad)
Q
Qiao Longfei 已提交
509
            },
M
minqiyang 已提交
510 511
            outputs={"ParamOut": param_and_grad[0]},
            stop_gradient=True)
Q
Qiao Longfei 已提交
512 513

        return sgd_op
514 515 516


class MomentumOptimizer(Optimizer):
Q
qiaolongfei 已提交
517 518 519 520 521 522 523 524 525 526 527 528 529 530
    """

    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):

531
        &\quad   param = param - (gradient + mu * velocity) * learning\_rate
Q
qiaolongfei 已提交
532 533 534

        & else:

Q
qiaolongfei 已提交
535
        &\quad   param = param - learning\_rate * velocity
Q
qiaolongfei 已提交
536 537 538 539 540 541

    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
542 543 544
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
Q
qiaolongfei 已提交
545 546 547 548

    Examples:
        .. code-block:: python

Q
qiaolongfei 已提交
549
            optimizer = fluid.optimizer.Momentum(learning_rate=0.2, momentum=0.1)
Q
qiaolongfei 已提交
550
            optimizer.minimize(cost)
551 552 553
    """
    _velocity_acc_str = "velocity"

X
Xin Pan 已提交
554 555 556 557 558 559
    def __init__(self,
                 learning_rate,
                 momentum,
                 use_nesterov=False,
                 regularization=None,
                 name=None):
560 561
        assert learning_rate is not None
        assert momentum is not None
Q
Qiao Longfei 已提交
562
        super(MomentumOptimizer, self).__init__(
X
Xin Pan 已提交
563 564 565
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
566 567
        self.type = "momentum"
        self._momentum = momentum
568
        self._use_nesterov = bool(use_nesterov)
569 570 571 572 573

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

        for p in parameters:
Q
Qiao Longfei 已提交
574
            self._add_accumulator(self._velocity_acc_str, p)
575 576 577 578 579 580 581 582 583 584 585 586 587

    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,
588
                "LearningRate": self._create_param_lr(param_and_grad)
589 590 591 592 593
            },
            outputs={
                "ParamOut": param_and_grad[0],
                "VelocityOut": velocity_acc
            },
594
            attrs={"mu": self._momentum,
M
minqiyang 已提交
595 596
                   "use_nesterov": self._use_nesterov},
            stop_gradient=True)
597 598

        return momentum_op
599 600


601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619
class DGCMomentumOptimizer(MomentumOptimizer):
    """

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

    DGC reduce the communication bandwidth by sending only the important gradients (sparse update):\
        only gradients larger than a threshold are transmitted.

    To avoid losing information, DGC accumulate the rest of the gradients locally.

    Eventually, these gradients become large enough to be transmitted.

    Thus, DGC send the large gradients immediately but eventually send all of the gradients over time.

    To ensure no loss of accuracy, DGC employs momentum correc-tionandlocal gradient clipping on top of the gradient sparsification to maintain model performance.

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

    This optimizer will do two things:
620

621 622
        1. Compress the gradient by get TopK import value from tensor \
            and use it for allreduce to reduce network bandwidth.
623

624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858
        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.
        rampup_begin_step (int): The begining step from which gradient compression is implemented.
        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.
        num_trainers: The number of training node.
        regularization: A Regularizer, such as fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.

    Examples:
        .. code-block:: python

            optimizer = fluid.optimizer.DGCMomentumOptimizer(
                learning_rate=fluid.layers.piecewise_decay(
                    boundaries=bd, values=lr),
                momentum=0.9,
                rampup_begin_step=1252,
                regularization=fluid.regularizer.L2Decay(1e-4))
            optimizer.minimize(cost)

    """

    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(
            counter_name='__g_dgc_counter__', begin=0)

        # 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,
            name='__g_rampup_begin_step__',
            value=self._rampup_begin_step * 1.0,
            force_cpu=True)

        for param_var, grad_var in param_and_grads:
            var_numel = reduce(lambda x, y: x * y, param_var.shape)
            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,
                name=param_var.name + "__dgc_u__",
                value=0.0)
            v_var = tensor.create_global_var(
                shape=param_var.shape,
                dtype=param_var.dtype,
                persistable=True,
                name=param_var.name + "__dgc_v__",
                value=0.0)

            k_var = tensor.create_global_var(
                shape=[1],
                dtype=param_var.dtype,
                persistable=True,
                name=param_var.name + "__dgc_k__",
                value=0.0,
                force_cpu=True)

            encoded_var = tensor.create_global_var(
                shape=[1],
                dtype=param_var.dtype,
                persistable=True,
                name=param_var.name + "__dgc_encoded__",
                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:
            name = unique_name.generate(".".join([helper.name, 'tmp']))

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

        helper.append_op(
            type="clip_by_norm",
            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(
                x=grad_var, max_norm=clip_norm, name=grad_var.name + "@DGC")

    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])


859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882
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
883

884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938

    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
939 940
            },
            stop_gradient=True)
941 942 943 944

        return momentum_op


945
class AdagradOptimizer(Optimizer):
Q
qiaolongfei 已提交
946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965
    """
    **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
966 967 968
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
X
xuezhong 已提交
969
        initial_accumulator_value (float): Initial value for moment accumulator.
Q
qiaolongfei 已提交
970 971 972 973 974 975

    Examples:
        .. code-block:: python

            optimizer = fluid.optimizer.Adagrad(learning_rate=0.2)
            optimizer.minimize(cost)
976 977 978
    """
    _moment_acc_str = "moment"

X
Xin Pan 已提交
979 980 981 982
    def __init__(self,
                 learning_rate,
                 epsilon=1.0e-6,
                 regularization=None,
983
                 name=None,
X
xuezhong 已提交
984
                 initial_accumulator_value=0.0):
985 986
        assert learning_rate is not None
        assert epsilon is not None
Q
Qiao Longfei 已提交
987
        super(AdagradOptimizer, self).__init__(
X
Xin Pan 已提交
988 989 990
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
991 992
        self.type = "adagrad"
        self._epsilon = epsilon
993
        self.initial_accumulator_value = initial_accumulator_value
994 995 996 997 998

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

        for p in parameters:
Q
Qiao Longfei 已提交
999
            self._add_accumulator(self._moment_acc_str, p)
1000 1001 1002 1003 1004 1005

    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])
1006 1007 1008 1009 1010 1011 1012 1013 1014 1015
        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,
            })
1016

1017
        # Create the adagrad optimizer op
1018 1019 1020 1021 1022 1023
        adagrad_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
                "Moment": moment_acc,
1024
                "LearningRate": self._create_param_lr(param_and_grad)
1025 1026 1027
            },
            outputs={"ParamOut": param_and_grad[0],
                     "MomentOut": moment_acc},
M
minqiyang 已提交
1028 1029
            attrs={"epsilon": self._epsilon},
            stop_gradient=True)
1030 1031

        return adagrad_op
1032 1033 1034


class AdamOptimizer(Optimizer):
Q
qiaolongfei 已提交
1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061
    """
    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.
1062
        regularization: A Regularizer, such as fluid.regularizer.L2DecayRegularizer.
X
Xin Pan 已提交
1063
        name: A optional name prefix.
1064 1065 1066 1067 1068 1069
        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
1070 1071 1072 1073 1074 1075 1076

    Examples:
        .. code-block:: python

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

1077 1078 1079
    """
    _moment1_acc_str = "moment1"
    _moment2_acc_str = "moment2"
Q
qiaolongfei 已提交
1080 1081
    _beta1_pow_acc_str = "beta1_pow_acc"
    _beta2_pow_acc_str = "beta2_pow_acc"
1082 1083 1084 1085 1086

    def __init__(self,
                 learning_rate=0.001,
                 beta1=0.9,
                 beta2=0.999,
1087
                 epsilon=1e-8,
X
Xin Pan 已提交
1088
                 regularization=None,
Q
Qiao Longfei 已提交
1089
                 name=None,
Q
Qiao Longfei 已提交
1090
                 lazy_mode=False):
1091 1092 1093 1094
        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
1095
        super(AdamOptimizer, self).__init__(
X
Xin Pan 已提交
1096 1097 1098
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
1099 1100 1101 1102
        self.type = "adam"
        self._beta1 = beta1
        self._beta2 = beta2
        self._epsilon = epsilon
Q
Qiao Longfei 已提交
1103
        self._lazy_mode = lazy_mode
1104 1105 1106 1107 1108 1109

    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
1110 1111
            self._add_accumulator(self._moment1_acc_str, p)
            self._add_accumulator(self._moment2_acc_str, p)
Q
qiaolongfei 已提交
1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123
            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])
1124 1125 1126 1127 1128 1129 1130 1131

    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
1132 1133 1134 1135 1136
        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])

1137
        # create the adam optimize op
1138 1139 1140 1141 1142
        adam_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
1143
                "LearningRate": self._create_param_lr(param_and_grad),
1144 1145
                "Moment1": moment1,
                "Moment2": moment2,
Q
qiaolongfei 已提交
1146 1147
                "Beta1Pow": beta1_pow_acc,
                "Beta2Pow": beta2_pow_acc
1148 1149 1150 1151 1152 1153 1154 1155 1156
            },
            outputs={
                "ParamOut": param_and_grad[0],
                "Moment1Out": moment1,
                "Moment2Out": moment2
            },
            attrs={
                "beta1": self._beta1,
                "beta2": self._beta2,
Q
Qiao Longfei 已提交
1157
                "epsilon": self._epsilon,
1158 1159
                "lazy_mode": self._lazy_mode,
                "min_row_size_to_use_multithread": 1000
M
minqiyang 已提交
1160 1161
            },
            stop_gradient=True)
1162 1163 1164

        return adam_op

1165
    def _finish_update(self, block, param_and_grads):
1166 1167 1168
        """Update Beta1 and Beta2 Power accumulators
        """
        assert isinstance(block, framework.Block)
Q
Qiao Longfei 已提交
1169
        main_block = block.program.global_block()
1170 1171 1172
        for param, grad in param_and_grads:
            if grad is None:
                continue
X
Xin Pan 已提交
1173 1174
            with param.block.program._optimized_guard(
                [param, grad]), name_scope("optimizer"):
1175 1176 1177 1178 1179 1180 1181 1182
                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
1183 1184
                    attrs={"scale": self._beta1},
                    stop_gradient=True)
1185 1186 1187 1188 1189

                main_block.append_op(
                    type="scale",
                    inputs={"X": beta2_pow_acc},
                    outputs={"Out": beta2_pow_acc},
M
minqiyang 已提交
1190 1191
                    attrs={"scale": self._beta2},
                    stop_gradient=True)
1192 1193 1194


class AdamaxOptimizer(Optimizer):
Q
qiaolongfei 已提交
1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224
    """
    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
1225 1226 1227
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
Q
qiaolongfei 已提交
1228 1229 1230 1231 1232 1233

    Examples:
        .. code-block:: python

            optimizer = fluid.optimizer.Adamax(learning_rate=0.2)
            optimizer.minimize(cost)
C
chengduo 已提交
1234 1235 1236

    Notes:
       Currently, AdamaxOptimizer doesn't support sparse parameter optimization.
1237 1238 1239
    """
    _moment_acc_str = "moment"
    _inf_norm_acc_str = "inf_norm"
Q
qiaolongfei 已提交
1240
    _beta1_pow_acc_str = "beta1_pow_acc"
1241 1242 1243 1244 1245

    def __init__(self,
                 learning_rate=0.001,
                 beta1=0.9,
                 beta2=0.999,
1246
                 epsilon=1e-8,
X
Xin Pan 已提交
1247 1248
                 regularization=None,
                 name=None):
1249 1250 1251 1252
        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
1253
        super(AdamaxOptimizer, self).__init__(
X
Xin Pan 已提交
1254 1255 1256
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
1257 1258 1259 1260 1261 1262 1263 1264
        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
1265 1266
            self._add_accumulator(self._moment_acc_str, p)
            self._add_accumulator(self._inf_norm_acc_str, p)
Q
qiaolongfei 已提交
1267 1268 1269 1270 1271 1272
            self._add_accumulator(
                name=self._beta1_pow_acc_str,
                param=p,
                dtype='float32',
                fill_value=self._beta1,
                shape=[1])
1273 1274 1275 1276 1277 1278 1279

    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
1280 1281
        beta1_pow_acc = self._get_accumulator(self._beta1_pow_acc_str,
                                              param_and_grad[0])
1282 1283 1284 1285 1286 1287
        # create the adamax optimize op
        adamax_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
1288
                "LearningRate": self._create_param_lr(param_and_grad),
1289 1290
                "Moment": moment,
                "InfNorm": inf_norm,
Q
qiaolongfei 已提交
1291
                "Beta1Pow": beta1_pow_acc
1292 1293 1294 1295 1296 1297 1298 1299 1300 1301
            },
            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
1302 1303
            },
            stop_gradient=True)
1304 1305 1306

        return adamax_op

1307
    def _finish_update(self, block, parameters_and_grads):
1308 1309 1310
        """Update Beta1 Power accumulator
        """
        assert isinstance(block, framework.Block)
Q
Qiao Longfei 已提交
1311
        main_block = block.program.global_block()
1312 1313 1314
        for param, grad in parameters_and_grads:
            if grad is None:
                continue
X
Xin Pan 已提交
1315 1316
            with param.block.program._optimized_guard(
                [param, grad]), name_scope('adamx'):
1317 1318 1319 1320 1321 1322
                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
1323 1324
                    attrs={"scale": self._beta1},
                    stop_gradient=True)
1325 1326 1327


class DecayedAdagradOptimizer(Optimizer):
1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349
    """
    **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
1350 1351 1352
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
1353 1354 1355 1356 1357 1358

    Examples:
        .. code-block:: python

            optimizer = fluid.optimizer.DecayedAdagrad(learning_rate=0.2)
            optimizer.minimize(cost)
C
chengduo 已提交
1359 1360 1361

    Notes:
       Currently, DecayedAdagradOptimizer doesn't support sparse parameter optimization.
1362 1363 1364
    """
    _moment_acc_str = "moment"

X
Xin Pan 已提交
1365 1366 1367 1368 1369 1370
    def __init__(self,
                 learning_rate,
                 decay=0.95,
                 epsilon=1.0e-6,
                 regularization=None,
                 name=None):
1371 1372 1373 1374
        assert learning_rate is not None
        assert decay is not None
        assert epsilon is not None

Q
Qiao Longfei 已提交
1375
        super(DecayedAdagradOptimizer, self).__init__(
X
Xin Pan 已提交
1376 1377 1378
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405
        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
1406 1407
            attrs={"epsilon": self._epsilon},
            stop_gradient=True)
1408 1409

        return decayed_adagrad_op
1410 1411


1412
class AdadeltaOptimizer(Optimizer):
1413 1414
    """
    **Adadelta Optimizer**
Q
qiaolongfei 已提交
1415

1416
    Simple Adadelta optimizer with average squared grad state and
1417
    average squared update state.
1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429
    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
1430
        learning_rate(float): global learning rate
1431 1432
        rho(float): rho in equation
        epsilon(float): epsilon in equation
X
Xin Pan 已提交
1433 1434 1435
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
1436 1437 1438 1439 1440 1441 1442

    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
1443 1444 1445

    Notes:
       Currently, AdadeltaOptimizer doesn't support sparse parameter optimization.
1446
    """
1447

1448 1449 1450
    _avg_squared_grad_acc_str = "_avg_squared_grad"
    _avg_squared_update_acc_str = "_avg_squared_update"

X
Xin Pan 已提交
1451 1452 1453 1454 1455 1456
    def __init__(self,
                 learning_rate,
                 epsilon=1.0e-6,
                 rho=0.95,
                 regularization=None,
                 name=None):
1457 1458 1459 1460 1461 1462
        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.")
1463
        super(AdadeltaOptimizer, self).__init__(
X
Xin Pan 已提交
1464 1465 1466
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
1467 1468 1469 1470 1471
        self.type = "adadelta"
        self._epsilon = epsilon
        self._rho = rho

    def _create_accumulators(self, block, parameters):
1472 1473
        if not isinstance(block, framework.Block):
            raise TypeError("block is not instance of framework.Block.")
1474 1475 1476 1477 1478 1479

        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):
1480 1481
        if not isinstance(block, framework.Block):
            raise TypeError("block is not instance of framework.Block.")
1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502

        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
1503 1504
                   "rho": self._rho},
            stop_gradient=True)
1505 1506 1507 1508

        return adadelta_op


Q
qingqing01 已提交
1509 1510 1511 1512 1513 1514 1515 1516 1517 1518
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
1519
        r(w, t) & = \\rho r(w, t-1) + (1 - \\rho)(\\nabla Q_{i}(w))^2
Q
qingqing01 已提交
1520 1521 1522 1523

        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
1524
    each weight. Then dividing the gradient by :math:`sqrt{v(w,t)}`.
Q
qingqing01 已提交
1525 1526 1527 1528 1529 1530

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

    ..  math::

Q
qiaolongfei 已提交
1531
        r(w, t) & = \\rho r(w, t-1) + (1 - \\rho)(\\nabla Q_{i}(w))^2
Q
qingqing01 已提交
1532

1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546
        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
1547 1548 1549 1550
            \\epsilon}} \\nabla Q_{i}(w)

        w & = w - v(w, t)

Q
qiaolongfei 已提交
1551
    where, :math:`\\rho` is a hyperparameter and typical values are 0.9, 0.95
Q
qingqing01 已提交
1552 1553 1554 1555 1556 1557
    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
1558
        learning_rate(float): global learning rate.
Q
qingqing01 已提交
1559 1560 1561
        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
1562
        momentum(float): :math:`\\beta` in equation is the momentum term,
Q
qingqing01 已提交
1563
            set 0.0 by default.
1564 1565 1566 1567
        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
1568 1569 1570
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
Q
qingqing01 已提交
1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583

    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"
1584
    _mean_grad_acc_str = "mean_grad"
Q
qingqing01 已提交
1585 1586 1587 1588 1589 1590

    def __init__(self,
                 learning_rate,
                 rho=0.95,
                 epsilon=1.0e-6,
                 momentum=0.0,
1591
                 centered=False,
X
Xin Pan 已提交
1592 1593
                 regularization=None,
                 name=None):
Q
qingqing01 已提交
1594
        super(RMSPropOptimizer, self).__init__(
X
Xin Pan 已提交
1595 1596 1597
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
Q
qingqing01 已提交
1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610
        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
1611
        self._centered = centered
Q
qingqing01 已提交
1612 1613 1614 1615 1616 1617 1618 1619

    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)
1620
            self._add_accumulator(self._mean_grad_acc_str, p)
Q
qingqing01 已提交
1621 1622 1623 1624 1625 1626 1627 1628 1629

    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])
1630 1631
        mean_grad_acc = self._get_accumulator(self._mean_grad_acc_str,
                                              param_and_grad[0])
Q
qingqing01 已提交
1632 1633 1634 1635 1636 1637 1638
        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,
1639
                "MeanGrad": mean_grad_acc,
Q
qingqing01 已提交
1640 1641 1642 1643 1644
                "LearningRate": self._create_param_lr(param_and_grad),
            },
            outputs={
                "ParamOut": param_and_grad[0],
                "MomentOut": momentum_acc,
1645 1646
                "MeanSquareOut": mean_square_acc,
                "MeanGradOut": mean_grad_acc
Q
qingqing01 已提交
1647 1648 1649 1650
            },
            attrs={
                "epsilon": self._epsilon,
                "decay": self._rho,
1651 1652
                "momentum": self._momentum,
                "centered": self._centered
M
minqiyang 已提交
1653 1654
            },
            stop_gradient=True)
Q
qingqing01 已提交
1655 1656 1657 1658

        return rmsprop_op


Q
qiaolongfei 已提交
1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700
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
1701 1702 1703
        l1 (float): L1 regularization strength.
        l2 (float): L2 regularization strength.
        lr_power (float): Learning Rate Power.
X
Xin Pan 已提交
1704 1705 1706
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
Q
qiaolongfei 已提交
1707 1708 1709 1710 1711 1712 1713 1714 1715

    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
1716 1717 1718

    Notes:
       Currently, FtrlOptimizer doesn't support sparse parameter optimization.
Q
qiaolongfei 已提交
1719 1720 1721 1722 1723
    """

    _squared_acc_str = "squared"
    _linear_acc_str = "linear"

X
Xin Pan 已提交
1724 1725 1726 1727 1728 1729 1730
    def __init__(self,
                 learning_rate,
                 l1=0.0,
                 l2=0.0,
                 lr_power=-0.5,
                 regularization=None,
                 name=None):
Q
qiaolongfei 已提交
1731
        super(FtrlOptimizer, self).__init__(
X
Xin Pan 已提交
1732 1733 1734
            learning_rate=learning_rate,
            regularization=regularization,
            name=name)
Q
qiaolongfei 已提交
1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774
        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
1775 1776
                   "lr_power": self._lr_power},
            stop_gradient=True)
Q
qiaolongfei 已提交
1777 1778 1779 1780

        return ftrl_op


1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794
# 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
1795
Adadelta = AdadeltaOptimizer
Q
qingqing01 已提交
1796
RMSProp = RMSPropOptimizer
Q
qiaolongfei 已提交
1797
Ftrl = FtrlOptimizer
1798
LarsMomentum = LarsMomentumOptimizer
1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813


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
1814 1815 1816
        regularization: A Regularizer, such as
                        fluid.regularizer.L2DecayRegularizer.
        name: A optional name prefix.
1817
    Examples:
Q
qiaolongfei 已提交
1818 1819 1820

      .. code-block:: python

1821
        optimizer = fluid.optimizer.Momentum()
1822 1823
        optimizer.minimize(cost)
        model_average = fluid.optimizer.ModelAverage(0.15,
1824 1825 1826 1827 1828
                                                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()...)
1829 1830 1831 1832

            with model_average.apply(exe):
                for data in test_reader():
                    exe.run(inference_program...)
1833 1834 1835
    """

    def __init__(self,
W
wanghaoshuang 已提交
1836
                 average_window_rate,
1837 1838
                 min_average_window=10000,
                 max_average_window=10000,
X
Xin Pan 已提交
1839 1840 1841 1842
                 regularization=None,
                 name=None):
        super(ModelAverage, self).__init__(
            0.0, regularization=regularization, name=name)
1843 1844 1845
        self.average_window = average_window_rate
        self.min_average_window = min_average_window
        self.max_average_window = max_average_window
1846

1847
        self.params_grads = []
1848 1849
        for param in framework.default_main_program().global_block(
        ).all_parameters():
1850
            if param.do_model_average != False:
1851 1852 1853 1854
                grad = param.block.create_var(
                    name=unique_name.generate(".".join([param.name, 'tmp'])),
                    dtype=param.dtype,
                    persistable=False,
W
wanghaoshuang 已提交
1855
                    stop_gradient=True)
1856
                self.params_grads.append((param, grad))
1857

1858
        for param, grad in self.params_grads:
1859 1860
            if grad is None:
                continue
X
Xin Pan 已提交
1861 1862
            with param.block.program._optimized_guard(
                [param, grad]), name_scope('move_average'):
1863
                self._append_average_accumulate_op(param)
1864

1865 1866 1867 1868
        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:
1869
                self._add_average_apply_op(block, param_grad)
1870 1871 1872 1873 1874

        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:
1875
                self._add_average_restore_op(block, param_grad)
1876

1877
    def _add_average_apply_op(self, block, param_grad):
L
Luo Tao 已提交
1878 1879 1880 1881 1882 1883
        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(
1884
            self._get_accumulator('num_accumulates', param))
L
Luo Tao 已提交
1885
        old_num_accumulates = block._clone_variable(
1886
            self._get_accumulator('old_num_accumulates', param))
L
Luo Tao 已提交
1887
        num_updates = block._clone_variable(
1888 1889 1890 1891 1892 1893
            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
1894 1895 1896 1897
        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
1898
        ops._elementwise_div(x=sum, y=tmp, out=param)
1899 1900

    def _add_average_restore_op(self, block, param_grad):
L
Luo Tao 已提交
1901 1902
        param = block._clone_variable(param_grad[0])
        grad = block._clone_variable(param_grad[1])
1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939
        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
1940 1941
            },
            stop_gradient=True)
1942

S
rename  
sneaxiy 已提交
1943
    @signature_safe_contextmanager
1944
    def apply(self, executor, need_restore=True):
1945 1946
        """Apply average values to parameters of current model.
        """
1947 1948 1949 1950 1951 1952
        executor.run(self.apply_program)
        try:
            yield
        finally:
            if need_restore:
                self.restore(executor)
1953 1954 1955 1956

    def restore(self, executor):
        """Restore parameter values of current model.
        """
1957
        executor.run(self.restore_program)