optimizer.py 23.1 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 14
# 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.

15
from collections import defaultdict
Q
Qiao Longfei 已提交
16

17
import framework
Q
Qiao Longfei 已提交
18
import layers
F
fengjiayi 已提交
19
from backward import append_backward
Y
Yu Yang 已提交
20 21
from framework import program_guard
import unique_name
22 23 24
from initializer import Constant
from layer_helper import LayerHelper
from regularizer import append_regularization_ops
F
fengjiayi 已提交
25
from clip import append_gradient_clip_ops, error_clip_callback
26

27 28 29
__all__ = [
    'SGD', 'Momentum', 'Adagrad', 'Adam', 'Adamax', 'DecayedAdagrad', 'Adadelta'
]
Q
Qiao Longfei 已提交
30 31 32 33 34 35


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

    Define the common interface of an optimizer.
36 37
    User should not use this class directly,
    but need to use one of it's implementation.
Q
Qiao Longfei 已提交
38 39
    """

Y
Yu Yang 已提交
40
    def __init__(self, learning_rate, regularization=None):
41 42
        if not isinstance(learning_rate, float) and \
                not isinstance(learning_rate, framework.Variable):
Q
qiaolongfei 已提交
43
            raise TypeError("learning rate should be float or Variable")
D
dzhwinter 已提交
44
        self.regularization = regularization
45 46 47
        self._learning_rate = learning_rate
        # each program should have a independent learning rate
        # program -> Variable(learning_rate)
Q
qiaolongfei 已提交
48
        self._learning_rate_map = dict()
49 50 51
        if isinstance(self._learning_rate, framework.Variable):
            self._learning_rate_map[framework.default_main_program(
            )] = self._learning_rate
52 53 54 55 56
        # 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
57
        self.helper = None
Q
Qiao Longfei 已提交
58

Q
Qiao Longfei 已提交
59
    def _create_global_learning_rate(self):
60
        lr = self.global_learning_rate()
Q
Qiao Longfei 已提交
61

62 63 64 65
        if isinstance(lr, framework.Variable):
            return
        else:
            if not isinstance(self._learning_rate, float):
Q
qiaolongfei 已提交
66
                raise TypeError(
67 68
                    "learning rate variable is create outside optimizer,"
                    "can not create new learning rate variable for new program")
Q
Qiao Longfei 已提交
69

70 71 72 73 74 75 76 77 78 79
        # 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',
            persistable=True)

    def global_learning_rate(self, program=None):
Q
Qiao Longfei 已提交
80 81 82 83
        """
        get global decayed learning rate
        :return:
        """
84 85
        if program is None:
            program = framework.default_main_program()
Q
qiaolongfei 已提交
86
        return self._learning_rate_map.get(program, None)
Q
Qiao Longfei 已提交
87

Q
Qiao Longfei 已提交
88 89 90 91 92
    def _append_optimize_op(self, block, param_and_grad):
        """ append optimize operator to block and return all the added optimize_op
        """
        raise NotImplementedError()

93 94 95 96
    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']
Q
qiaolongfei 已提交
97 98 99 100
        if param_lr == 1.0:
            return self.global_learning_rate()
        else:
            return self.global_learning_rate() * param_lr
101 102 103 104 105 106 107

    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
108
        """
109 110
        pass

111 112 113 114 115 116 117 118 119 120 121 122 123
    def _finish_update(self, block):
        """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:
            list of finish ops or None
        """
        pass

Q
Qiao Longfei 已提交
124
    def _add_accumulator(self, name, param, dtype=None, fill_value=0.0):
125 126 127 128 129 130 131 132 133 134 135
        """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
        """
        if (name in self._accumulators and
                param.name in self._accumulators[name]):
136
            raise Exception("Accumulator {} already exists for parameter {}".
137
                            format(name, param.name))
Q
Qiao Longfei 已提交
138 139 140

        assert isinstance(self.helper, LayerHelper)
        var = self.helper.create_global_variable(
Y
Yu Yang 已提交
141
            name=unique_name.generate(name),
Q
Qiao Longfei 已提交
142
            persistable=True,
F
fengjiayi 已提交
143
            dtype=dtype or param.dtype,
Q
Qiao Longfei 已提交
144 145 146
            type=param.type,
            shape=param.shape)
        self.helper.set_variable_initializer(
147
            var, initializer=Constant(value=float(fill_value)))
Q
Qiao Longfei 已提交
148
        self._accumulators[name][param.name] = var
149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165

    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
        """
        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]

Q
Qiao Longfei 已提交
166 167 168
    def create_optimization_pass(self,
                                 parameters_and_grads,
                                 loss,
169
                                 startup_program=None):
Q
Qiao Longfei 已提交
170 171 172 173 174 175 176
        """Add optimization operators to update gradients to variables.

        Args:
          loss: the target that this optimization is for.
          parameters_and_grads: a list of (variable, gradient) pair to update.

        Returns:
177 178 179 180
          return_op_list: a list of operators that will complete one step of
          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
181
          :param startup_program:
Q
Qiao Longfei 已提交
182
        """
183 184 185 186 187
        # 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
188
        # for parameters and extend _finish_update method to add custom ops.
189 190

        # Create any accumulators
Q
Qiao Longfei 已提交
191
        program = loss.block.program
192
        with program_guard(program, startup_program):
Y
Yancey1989 已提交
193 194
            global_block = framework.default_main_program().global_block()
            start = len(global_block.ops)
195 196 197
            self.helper = LayerHelper(self.__class__.__name__)
            self._create_accumulators(loss.block,
                                      [p[0] for p in parameters_and_grads])
Q
Qiao Longfei 已提交
198
            self._create_global_learning_rate()
199 200 201 202 203 204 205 206 207 208 209

            optimize_ops = []
            for param_and_grad in parameters_and_grads:
                if param_and_grad[0].trainable is True and param_and_grad[
                        1] is not None:
                    optimize_op = self._append_optimize_op(loss.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
Y
Yancey1989 已提交
210
            self._finish_update(loss.block)
211

Y
Yancey1989 已提交
212 213
            end = len(global_block.ops)
            return global_block.slice_ops(start, end)
Q
Qiao Longfei 已提交
214

Q
Qiao Longfei 已提交
215 216
    def minimize(self,
                 loss,
217
                 startup_program=None,
Q
Qiao Longfei 已提交
218 219
                 parameter_list=None,
                 no_grad_set=None):
Q
Qiao Longfei 已提交
220 221
        """Add operations to minimize `loss` by updating `parameter_list`.

F
fengjiayi 已提交
222
        This method combines interface `append_backward()` and
Q
Qiao Longfei 已提交
223 224
        `create_optimization_pass()` into one.
        """
F
fengjiayi 已提交
225
        params_grads = append_backward(loss, parameter_list, no_grad_set,
Y
Yang Yang 已提交
226
                                       [error_clip_callback])
Y
Yu Yang 已提交
227

Y
Yu Yang 已提交
228 229
        params_grads = sorted(params_grads, key=lambda x: x[0].name)

Y
Yu Yang 已提交
230 231
        params_grads = append_gradient_clip_ops(params_grads)

F
fengjiayi 已提交
232
        # Add regularization if any
D
dzhwinter 已提交
233 234
        params_grads = append_regularization_ops(params_grads,
                                                 self.regularization)
Y
Yu Yang 已提交
235

Q
Qiao Longfei 已提交
236
        optimize_ops = self.create_optimization_pass(params_grads, loss,
237
                                                     startup_program)
T
typhoonzero 已提交
238
        return optimize_ops, params_grads
Q
Qiao Longfei 已提交
239 240 241 242 243 244


class SGDOptimizer(Optimizer):
    """ Simple SGD optimizer without any state.
    """

D
dzhwinter 已提交
245
    def __init__(self, learning_rate, **kwargs):
Q
Qiao Longfei 已提交
246
        assert learning_rate is not None
Q
Qiao Longfei 已提交
247 248
        super(SGDOptimizer, self).__init__(
            learning_rate=learning_rate, **kwargs)
Q
Qiao Longfei 已提交
249 250
        self.type = "sgd"

251 252
    def _append_optimize_op(self, block, param_and_grad):
        assert isinstance(block, framework.Block)
253

Q
Qiao Longfei 已提交
254 255 256 257 258 259
        # create the optimize op
        sgd_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
260
                "LearningRate": self._create_param_lr(param_and_grad)
Q
Qiao Longfei 已提交
261
            },
262
            outputs={"ParamOut": param_and_grad[0]})
Q
Qiao Longfei 已提交
263 264

        return sgd_op
265 266 267 268 269 270 271


class MomentumOptimizer(Optimizer):
    """Simple Momentum optimizer with velocity state
    """
    _velocity_acc_str = "velocity"

D
dzhwinter 已提交
272
    def __init__(self, learning_rate, momentum, use_nesterov=False, **kwargs):
273 274
        assert learning_rate is not None
        assert momentum is not None
Q
Qiao Longfei 已提交
275 276
        super(MomentumOptimizer, self).__init__(
            learning_rate=learning_rate, **kwargs)
277 278
        self.type = "momentum"
        self._momentum = momentum
279
        self._use_nesterov = bool(use_nesterov)
280 281 282 283 284

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

        for p in parameters:
Q
Qiao Longfei 已提交
285
            self._add_accumulator(self._velocity_acc_str, p)
286 287 288 289 290 291 292 293 294 295 296 297 298

    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,
299
                "LearningRate": self._create_param_lr(param_and_grad)
300 301 302 303 304
            },
            outputs={
                "ParamOut": param_and_grad[0],
                "VelocityOut": velocity_acc
            },
305
            attrs={"mu": self._momentum,
306
                   "use_nesterov": self._use_nesterov})
307 308

        return momentum_op
309 310 311 312 313 314 315


class AdagradOptimizer(Optimizer):
    """Simple Adagrad optimizer with moment state
    """
    _moment_acc_str = "moment"

D
dzhwinter 已提交
316
    def __init__(self, learning_rate, epsilon=1.0e-6, **kwargs):
317 318
        assert learning_rate is not None
        assert epsilon is not None
Q
Qiao Longfei 已提交
319 320
        super(AdagradOptimizer, self).__init__(
            learning_rate=learning_rate, **kwargs)
321 322 323 324 325 326 327
        self.type = "adagrad"
        self._epsilon = epsilon

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

        for p in parameters:
Q
Qiao Longfei 已提交
328
            self._add_accumulator(self._moment_acc_str, p)
329 330 331 332 333 334 335

    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])

336
        # Create the adagrad optimizer op
337 338 339 340 341 342
        adagrad_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
                "Moment": moment_acc,
343
                "LearningRate": self._create_param_lr(param_and_grad)
344 345 346 347 348 349
            },
            outputs={"ParamOut": param_and_grad[0],
                     "MomentOut": moment_acc},
            attrs={"epsilon": self._epsilon})

        return adagrad_op
350 351 352 353 354 355 356 357 358 359 360 361


class AdamOptimizer(Optimizer):
    """Implements the Adam Optimizer
    """
    _moment1_acc_str = "moment1"
    _moment2_acc_str = "moment2"

    def __init__(self,
                 learning_rate=0.001,
                 beta1=0.9,
                 beta2=0.999,
362
                 epsilon=1e-8,
D
dzhwinter 已提交
363
                 **kwargs):
364 365 366 367
        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
368 369
        super(AdamOptimizer, self).__init__(
            learning_rate=learning_rate, **kwargs)
370 371 372 373 374 375 376 377
        self.type = "adam"
        self._beta1 = beta1
        self._beta2 = beta2
        self._epsilon = epsilon

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

Q
Qiao Longfei 已提交
378
        main_block = block.program.global_block()
379 380
        # Create beta1 and beta2 power tensors
        beta_shape = [1]
Q
Qiao Longfei 已提交
381
        self._beta1_pow_acc = self.helper.create_global_variable(
Y
Yu Yang 已提交
382
            name=unique_name.generate('beta1_pow_acc'),
Q
Qiao Longfei 已提交
383 384 385 386 387
            dtype='float32',
            shape=beta_shape,
            lod_level=0,
            persistable=True)
        self.helper.set_variable_initializer(
388
            self._beta1_pow_acc, initializer=Constant(self._beta1))
Q
Qiao Longfei 已提交
389 390

        self._beta2_pow_acc = self.helper.create_global_variable(
Y
Yu Yang 已提交
391
            name=unique_name.generate('beta2_pow_acc'),
Q
Qiao Longfei 已提交
392 393 394 395 396 397
            dtype='float32',
            shape=beta_shape,
            lod_level=0,
            persistable=True)

        self.helper.set_variable_initializer(
398
            self._beta2_pow_acc, initializer=Constant(self._beta2))
399 400 401

        # Create accumulator tensors for first and second moments
        for p in parameters:
Q
Qiao Longfei 已提交
402 403
            self._add_accumulator(self._moment1_acc_str, p)
            self._add_accumulator(self._moment2_acc_str, p)
404 405 406 407 408 409 410 411

    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])
412
        # create the adam optimize op
413 414 415 416 417
        adam_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
418
                "LearningRate": self._create_param_lr(param_and_grad),
419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440
                "Moment1": moment1,
                "Moment2": moment2,
                "Beta1Pow": self._beta1_pow_acc,
                "Beta2Pow": self._beta2_pow_acc
            },
            outputs={
                "ParamOut": param_and_grad[0],
                "Moment1Out": moment1,
                "Moment2Out": moment2
            },
            attrs={
                "beta1": self._beta1,
                "beta2": self._beta2,
                "epsilon": self._epsilon
            })

        return adam_op

    def _finish_update(self, block):
        """Update Beta1 and Beta2 Power accumulators
        """
        assert isinstance(block, framework.Block)
Q
Qiao Longfei 已提交
441 442
        main_block = block.program.global_block()
        scale_beta1 = main_block.append_op(
443 444 445 446 447
            type="scale",
            inputs={"X": self._beta1_pow_acc},
            outputs={"Out": self._beta1_pow_acc},
            attrs={"scale": self._beta1})

Q
Qiao Longfei 已提交
448
        scale_beta2 = main_block.append_op(
449 450 451 452 453 454
            type="scale",
            inputs={"X": self._beta2_pow_acc},
            outputs={"Out": self._beta2_pow_acc},
            attrs={"scale": self._beta2})

        return [scale_beta1, scale_beta2]
455 456 457 458 459 460 461 462 463 464 465 466


class AdamaxOptimizer(Optimizer):
    """Implements the Adamax Optimizer
    """
    _moment_acc_str = "moment"
    _inf_norm_acc_str = "inf_norm"

    def __init__(self,
                 learning_rate=0.001,
                 beta1=0.9,
                 beta2=0.999,
467
                 epsilon=1e-8,
D
dzhwinter 已提交
468
                 **kwargs):
469 470 471 472
        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
473 474
        super(AdamaxOptimizer, self).__init__(
            learning_rate=learning_rate, **kwargs)
475 476 477 478 479 480 481 482
        self.type = "adamax"
        self._beta1 = beta1
        self._beta2 = beta2
        self._epsilon = epsilon

    def _create_accumulators(self, block, parameters):
        # Create beta1 power accumulator tensor
        beta_shape = [1]
Q
Qiao Longfei 已提交
483
        self._beta1_pow_acc = self.helper.create_global_variable(
Y
Yu Yang 已提交
484
            name=unique_name.generate('beta1_pow_acc'),
Q
Qiao Longfei 已提交
485 486 487 488 489
            dtype='float32',
            shape=beta_shape,
            lod_level=0,
            persistable=True)
        self.helper.set_variable_initializer(
490
            self._beta1_pow_acc, initializer=Constant(self._beta1))
491 492 493

        # Create accumulator tensors for first moment and infinity norm
        for p in parameters:
Q
Qiao Longfei 已提交
494 495
            self._add_accumulator(self._moment_acc_str, p)
            self._add_accumulator(self._inf_norm_acc_str, p)
496 497 498 499 500 501 502 503 504 505 506 507 508

    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])
        # create the adamax optimize op
        adamax_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
509
                "LearningRate": self._create_param_lr(param_and_grad),
510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530
                "Moment": moment,
                "InfNorm": inf_norm,
                "Beta1Pow": self._beta1_pow_acc
            },
            outputs={
                "ParamOut": param_and_grad[0],
                "MomentOut": moment,
                "InfNormOut": inf_norm
            },
            attrs={
                "beta1": self._beta1,
                "beta2": self._beta2,
                "epsilon": self._epsilon
            })

        return adamax_op

    def _finish_update(self, block):
        """Update Beta1 Power accumulator
        """
        assert isinstance(block, framework.Block)
Q
Qiao Longfei 已提交
531 532
        main_block = block.program.global_block()
        scale_beta1 = main_block.append_op(
533 534 535 536 537 538
            type="scale",
            inputs={"X": self._beta1_pow_acc},
            outputs={"Out": self._beta1_pow_acc},
            attrs={"scale": self._beta1})

        return [scale_beta1]
539 540 541 542 543 544 545


class DecayedAdagradOptimizer(Optimizer):
    """Simple Decayed Adagrad optimizer with moment state
    """
    _moment_acc_str = "moment"

D
dzhwinter 已提交
546
    def __init__(self, learning_rate, decay=0.95, epsilon=1.0e-6, **kwargs):
547 548 549 550
        assert learning_rate is not None
        assert decay is not None
        assert epsilon is not None

Q
Qiao Longfei 已提交
551 552
        super(DecayedAdagradOptimizer, self).__init__(
            learning_rate=learning_rate, **kwargs)
553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582
        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},
            attrs={"epsilon": self._epsilon})

        return decayed_adagrad_op
583 584


585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636
class AdadeltaOptimizer(Optimizer):
    """Simple Adadelta optimizer with average squared grad state and
    average squared update state.
    """
    _avg_squared_grad_acc_str = "_avg_squared_grad"
    _avg_squared_update_acc_str = "_avg_squared_update"

    def __init__(self, learning_rate, epsilon=1.0e-6, rho=0.95, **kwargs):
        assert learning_rate is not None
        assert epsilon is not None
        assert rho is not None
        super(AdadeltaOptimizer, self).__init__(
            learning_rate=learning_rate, **kwargs)
        self.type = "adadelta"
        self._epsilon = epsilon
        self._rho = rho

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

        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):
        assert isinstance(block, framework.Block)

        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,
                   "rho": self._rho})

        return adadelta_op


637 638 639 640 641 642 643 644 645 646 647 648 649 650
# 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
651
Adadelta = AdadeltaOptimizer