optimizer.py 172.1 KB
Newer Older
1
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
D
dzhwinter 已提交
2
#
D
dzhwinter 已提交
3 4 5
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
D
dzhwinter 已提交
6
#
D
dzhwinter 已提交
7
#     http://www.apache.org/licenses/LICENSE-2.0
D
dzhwinter 已提交
8
#
D
dzhwinter 已提交
9 10 11 12 13
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
14 15

from __future__ import print_function
16

17
import numpy as np
18
from collections import defaultdict
19

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

23 24
from . import framework
from . import layers
25
from . import unique_name
26
from .backward import append_backward, _some_in_set_, _append_grad_suffix_, _get_no_grad_set_name
27
from .clip import GradientClipBase, error_clip_callback, append_gradient_clip_ops
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
from .dygraph import base as imperative_base
34
from .dygraph import no_grad
35 36 37 38
from .dygraph.learning_rate_scheduler import LearningRateDecay
from paddle.fluid import core
from paddle.fluid.layers import tensor
from functools import reduce
39
from .wrapped_decorator import signature_safe_contextmanager
M
mapingshuo 已提交
40
from .. import compat as cpt
41

42
__all__ = [
43 44 45 46
    'SGD', 'Momentum', 'Adagrad', 'Adam', 'Adamax', 'Dpsgd', 'DecayedAdagrad',
    'Ftrl', 'SGDOptimizer', 'MomentumOptimizer', 'AdagradOptimizer',
    'AdamOptimizer', 'AdamaxOptimizer', 'DpsgdOptimizer',
    'DecayedAdagradOptimizer', 'RMSPropOptimizer', 'FtrlOptimizer', 'Adadelta',
Z
Zeng Jinle 已提交
47 48 49 50
    'AdadeltaOptimizer', 'ModelAverage', 'LarsMomentum',
    'LarsMomentumOptimizer', 'DGCMomentumOptimizer', 'LambOptimizer',
    'ExponentialMovingAverage', 'PipelineOptimizer', 'LookaheadOptimizer',
    'RecomputeOptimizer'
51
]
Q
Qiao Longfei 已提交
52 53 54 55 56 57


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

    Define the common interface of an optimizer.
58 59
    User should not use this class directly,
    but need to use one of it's implementation.
Q
Qiao Longfei 已提交
60 61
    """

62
    @imperative_base.no_grad
63 64 65 66 67
    def __init__(self,
                 learning_rate,
                 parameter_list=None,
                 regularization=None,
                 name=None):
68
        self._parameter_list = parameter_list
L
lujun 已提交
69
        if framework.in_dygraph_mode():
M
minqiyang 已提交
70 71 72 73 74
            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))
75 76 77 78
            if name is not None:
                self._name = unique_name.generate(name)
            else:
                self._name = unique_name.generate(self.__class__.__name__)
79
            if self._parameter_list is None:
80 81 82
                raise AttributeError(
                    "parameter_list argument given to the Optimizer should not be None in dygraph mode."
                )
M
minqiyang 已提交
83 84 85 86 87 88
        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))
89
            self._name = name
M
minqiyang 已提交
90

D
dzhwinter 已提交
91
        self.regularization = regularization
92
        self._learning_rate = learning_rate
D
dzhwinter 已提交
93 94
        # the learning rate type should be inferenced from loss
        self._dtype = None
95
        # each program should have a independent learning rate
96
        # program -> Variable(learning_rate)
Q
qiaolongfei 已提交
97
        self._learning_rate_map = dict()
98
        if isinstance(self._learning_rate, framework.Variable):
99 100
            self._learning_rate_map[framework.default_main_program(
            )] = self._learning_rate
101 102 103 104 105
        # 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
106
        self.helper = None
107
        self._opti_name_list = []
H
hong 已提交
108
        self._accumulators_holder = {}
109
        self._param_device_map = dict()
110 111
        # if pass grad_clip into minimize, it will not be None
        self._grad_clip = None
H
hong 已提交
112 113 114 115

    @framework.dygraph_only
    def state_dict(self):
        '''
T
tianshuo78520a 已提交
116 117
        Get state dict information from optimizer. It contain all the variable used by optimizer. For Adam optimizer, contains beta1, beta2, momentum etc. If LearningRateDecay have been used, global_step will be include in state dict.
        If the optimizer never be called(minimize function), the state_dict is empty.
H
hong 已提交
118 119 120

        Args: None
        Return:
T
tianshuo78520a 已提交
121
            state_dict(dict) : dict contains all the variable used by optimizer
H
hong 已提交
122 123 124 125 126
        
        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
127 128 129 130 131 132

                with fluid.dygraph.guard():
                    emb = fluid.dygraph.Embedding([10, 10])

                    adam = fluid.optimizer.Adam(0.001, parameter_list=emb.parameters())
                    state_dict = adam.state_dict()
H
hong 已提交
133 134 135 136 137 138 139 140

        '''
        state_dict = {}
        for k, v in self._accumulators.items():
            for para_name, var_tmp in v.items():
                state_dict[var_tmp.name] = var_tmp
        # global step if use lr decay
        if isinstance(self._learning_rate, LearningRateDecay):
141
            var_tmp = None
L
Leo Chen 已提交
142
            if framework.in_dygraph_mode():
143 144
                var_temp = framework._varbase_creator(
                    None, name='global_step', dtype='int32')
L
Leo Chen 已提交
145 146
            else:
                var_temp = Variable(None, name='global_step', dtype='int32')
147

H
hong 已提交
148 149 150 151 152 153 154 155 156
            tensor.fill_constant(
                [1], "int32", self._learning_rate.step_num, out=var_temp)

            state_dict['global_step'] = var_temp
        return state_dict

    @framework.dygraph_only
    def set_dict(self, state_dict):
        '''
T
tianshuo78520a 已提交
157
        Load optimizer state dict. For Adam optimizer, contains beta1, beta2, momentum etc. If LearningRateDecay have been used, global_step will be changed.
H
hong 已提交
158 159 160 161 162 163 164 165

        Args: 
            state_dict(dict) : Dict contains all the Variable needed by optimizer
        Return:
            None
        
        Examples:
            .. code-block:: python
166

H
hong 已提交
167
                with fluid.dygraph.guard():
168
                    emb = fluid.dygraph.Embedding([10, 10])
169

H
hong 已提交
170
                    state_dict = emb.state_dict()
171
                    fluid.save_dygraph(state_dict, "paddle_dy")
172

173 174
                    adam = fluid.optimizer.Adam(learning_rate=fluid.layers.noam_decay( 100, 10000), 
                                                parameter_list=emb.parameters())
H
hong 已提交
175
                    state_dict = adam.state_dict()
176
                    fluid.save_dygraph(state_dict, "paddle_dy")
177

H
hong 已提交
178
                    para_state_dict, opti_state_dict = fluid.load_dygraph( "paddle_dy")
179

180
                    adam.set_dict(opti_state_dict)
H
hong 已提交
181 182 183 184 185 186 187 188 189

        '''

        if isinstance(self._learning_rate, LearningRateDecay):
            assert 'global_step' in state_dict, \
                    'Global step not in state dict, Dygraph use LearningRateDecay, global_step must in state_dict'
            global_step = state_dict['global_step']

            if isinstance(global_step, core.VarBase):
190
                step_np = global_step
H
hong 已提交
191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214
                step_np = np.array(step_np.value().get_tensor())
                assert step_np.shape == (1,),  \
                        "global step shape is (1,), the shape is {}".format( step_np.shape )

                self._learning_rate.step_num = int(step_np[0])
            elif isinstance(global_step, Variable):
                step_np = global_step.numpy()
                assert step_np.shape == (1,),  \
                        "global step shape is (1,), the shape is {}".format( step_np.shape )
                self._learning_rate.step_num = step_np[0]
            elif isinstance(global_step, np.ndarray):
                assert global_step.shape == (1,),  \
                        "global step shape is (1,), the shape is {}".format( global_step.shape )
                self._learning_rate.step_num = global_step[0]
            else:
                raise RuntimeError(
                    "Type not supprt, value in state dict must be [VarBase, Variable, numpy], the type is ",
                    type(global_step))

        self._accumulators_holder = state_dict
        for k, v in self._accumulators.items():
            for para_name, var_tmp in v.items():
                assert var_tmp.name in state_dict, \
                        "optimizer variable {} not found".format( var_tmp.name )
215
                var = var_tmp.value()
H
hong 已提交
216 217 218 219 220 221 222 223
                tensor = var.get_tensor()
                model_np = np.array(tensor)

                load_para = state_dict[var_tmp.name]

                if isinstance(load_para, Variable):
                    load_para_np = load_para.numpy()
                elif isinstance(load_para, core.VarBase):
224
                    load_para_np = load_para.numpy()
H
hong 已提交
225 226 227 228 229 230 231 232 233 234 235 236 237 238 239
                elif isinstance(load_para, np.ndarray):
                    load_para_np = load_para
                else:
                    raise RuntimeError("State dict type {} not supprt".format(
                        str(type(load_para))))

                assert model_np.shape == load_para_np.shape,  \
                                          "Parameter shape not match, Dygraph Parameter [ {} ] need tensor with shape {} but load tensor with shape {}".format(
                                                 item.name, model_np.shape, load_para_np.shape)

                assert model_np.dtype == load_para_np.dtype, \
                                          "Parameter dtype not match, Dygraph Parameter [ {} ] need tensor with dtype {}  but load tensor with dtype {}".format(
                                                item.name, model_np.dtype, load_para_np.dtype)

                tensor.set(load_para_np, framework._current_expected_place())
240

241 242
    def get_opti_var_name_list(self):
        return self._opti_name_list
Q
Qiao Longfei 已提交
243

Q
Qiao Longfei 已提交
244
    def _create_global_learning_rate(self):
245 246 247
        if imperative_base.enabled():
            # create learning rate Variable
            if isinstance(self._learning_rate, float):
M
minqiyang 已提交
248 249 250 251 252 253 254 255 256 257 258 259
                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)
260
            # get learning rate Variable from LearningRateDecay
M
minqiyang 已提交
261
            elif isinstance(self._learning_rate, LearningRateDecay):
262 263 264
                self._learning_rate_map[framework.default_main_program(
                )] = self._learning_rate()
            else:
Q
qiaolongfei 已提交
265
                raise TypeError(
266 267
                    "optimizer's learning rate must be float or LearningRateDecay"
                )
268
        else:
269 270 271 272
            lr = self._global_learning_rate()

            if isinstance(lr, framework.Variable):
                return
M
minqiyang 已提交
273 274 275 276 277 278
            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
279

280 281 282 283 284 285 286 287
            # 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)
288

289 290 291 292
    @framework.dygraph_only
    def current_step_lr(self):
        """
        .. note::
T
tianshuo78520a 已提交
293
          **This API is ONLY available in Dygraph mode**
294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350
        
        Get current step learning rate. The return value is all the same When LearningRateDecay is not used,
        otherwise return the step learning rate.

        Returns:
            float: The learning rate of the current step.

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
                import numpy as np

                # example1: LearningRateDecay is not used, return value is all the same
                with fluid.dygraph.guard():
                    emb = fluid.dygraph.Embedding([10, 10])
                    adam = fluid.optimizer.Adam(0.001, parameter_list = emb.parameters())
                    lr = adam.current_step_lr()
                    print(lr) # 0.001

                # example2: PiecewiseDecay is used, return the step learning rate
                with fluid.dygraph.guard():
                    inp = np.random.uniform(-0.1, 0.1, [10, 10]).astype("float32")
                    linear = fluid.dygraph.nn.Linear(10, 10)
                    inp = fluid.dygraph.to_variable(inp)
                    out = linear(inp)
                    loss = fluid.layers.reduce_mean(out)
                    
                    bd = [2, 4, 6, 8]
                    value = [0.2, 0.4, 0.6, 0.8, 1.0]
                    adam = fluid.optimizer.Adam(fluid.dygraph.PiecewiseDecay(bd, value, 0),
                                           parameter_list=linear.parameters())

                    # first step: learning rate is 0.2
                    np.allclose(adam.current_step_lr(), 0.2, rtol=1e-06, atol=0.0) # True

                    # learning rate for different steps
                    ret = [0.2, 0.2, 0.4, 0.4, 0.6, 0.6, 0.8, 0.8, 1.0, 1.0, 1.0, 1.0]
                    for i in range(12):
                        adam.minimize(loss)
                        lr = adam.current_step_lr()
                        np.allclose(lr, ret[i], rtol=1e-06, atol=0.0) # True

        """
        current_lr = self._global_learning_rate()
        if current_lr:
            return self._global_learning_rate().numpy()[0]

        if isinstance(self._learning_rate, float):
            return self._learning_rate
        else:
            step_lr = self._learning_rate.step()
            if isinstance(step_lr, (float, int)):
                return step_lr
            else:
                return step_lr.numpy()[0]

Y
yuyang18 已提交
351
    def _global_learning_rate(self, program=None):
Q
Qiao Longfei 已提交
352 353 354 355
        """
        get global decayed learning rate
        :return:
        """
356 357
        if program is None:
            program = framework.default_main_program()
Q
qiaolongfei 已提交
358
        return self._learning_rate_map.get(program, None)
Q
Qiao Longfei 已提交
359

Q
Qiao Longfei 已提交
360 361 362 363 364
    def _append_optimize_op(self, block, param_and_grad):
        """ append optimize operator to block and return all the added optimize_op
        """
        raise NotImplementedError()

365 366 367 368
    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
369 370
        if type(param_lr) == Variable:
            return param_lr
Q
qiaolongfei 已提交
371
        else:
W
Wu Yi 已提交
372
            if param_lr == 1.0:
Y
yuyang18 已提交
373
                return self._global_learning_rate()
W
Wu Yi 已提交
374
            else:
X
Xin Pan 已提交
375 376 377
                with default_main_program()._lr_schedule_guard(
                        is_with_opt=True), framework.name_scope(
                            'scale_with_param_lr'):
378
                    return self._global_learning_rate() * param_lr
379 380 381 382 383 384 385

    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
386
        """
387 388
        pass

389
    def _finish_update(self, block, parameters_and_grads):
390 391 392 393 394 395 396 397
        """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
398
            None
399 400 401
        """
        pass

402 403 404 405 406
    def _add_accumulator(self,
                         name,
                         param,
                         dtype=None,
                         fill_value=0.0,
407
                         shape=None,
408
                         type=None,
409
                         device=None):
410 411 412 413 414 415 416 417 418
        """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
419 420
        if self._name is not None:
            name = self._name + "_" + name
421 422
        if (name in self._accumulators and
                param.name in self._accumulators[name]):
L
lujun 已提交
423
            if framework.in_dygraph_mode():
X
polish  
Xin Pan 已提交
424
                return self._accumulators[name][param.name]
425
            raise Exception("Accumulator {} already exists for parameter {}".
426
                            format(name, param.name))
427 428
        if shape == None:
            shape = param.shape
Q
Qiao Longfei 已提交
429
        assert isinstance(self.helper, LayerHelper)
430 431 432 433 434

        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
435
        var = self.helper.create_global_variable(
436
            name=var_name,
Q
Qiao Longfei 已提交
437
            persistable=True,
F
fengjiayi 已提交
438
            dtype=dtype or param.dtype,
439
            type=param.type if type is None else type,
H
hong 已提交
440 441
            shape=shape,
            belong_to_optimizer=True)
442 443 444 445 446
        if device is None:
            device = self._get_device_for_param(param.name)
        with device_guard(device):
            self.helper.set_variable_initializer(
                var, initializer=Constant(value=float(fill_value)))
H
hong 已提交
447 448 449 450 451 452 453

        if framework.in_dygraph_mode():
            if len(self._accumulators_holder) > 0:
                assert var_name in self._accumulators_holder, \
                        "Optimizer set error, {} should in state dict".format( var_name )
                var.set_value(self._accumulators_holder[var_name])

Q
Qiao Longfei 已提交
454
        self._accumulators[name][param.name] = var
455
        return var
456 457 458 459 460 461 462 463 464 465 466

    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
467 468
        if self._name is not None:
            name = self._name + "_" + name
469 470 471 472 473 474
        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]

475 476 477 478 479 480 481 482 483 484 485 486
    def _update_param_device_map(self, parameters_and_grads, target_block):
        for param_and_grad in parameters_and_grads:
            if param_and_grad[0].trainable is True:
                param_name = param_and_grad[0].name
                ops = target_block.ops
                device_attr_name = core.op_proto_and_checker_maker.kOpDeviceAttrName(
                )
                for op in ops:
                    input_arg_names = op.input_arg_names
                    if param_name in input_arg_names:
                        self._param_device_map[param_name] = op.attr(
                            device_attr_name)
487
                        break
488 489 490 491 492 493 494

    def _get_device_for_param(self, param_name):
        device = None
        if param_name in self._param_device_map:
            device = self._param_device_map[param_name]
        return device

495
    def _create_optimization_pass(self, parameters_and_grads):
Q
Qiao Longfei 已提交
496 497 498
        """Add optimization operators to update gradients to variables.

        Args:
Q
qiaolongfei 已提交
499
          parameters_and_grads(list(tuple(Variable, Variable))):
500
            a list of (variable, gradient) pair to update.
Q
Qiao Longfei 已提交
501 502

        Returns:
503
          return_op_list: a list of operators that will complete one step of
504 505 506
            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
507
        """
508 509 510 511 512
        # 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
513
        # for parameters and extend _finish_update method to add custom ops.
514

515
        # Allways called under program_guard use global block as loss block
516 517 518
        # But if current block is in control flow, append optimize op in the
        # grad block of current block

519
        global_block = framework.default_main_program().global_block()
520 521 522 523 524 525 526 527 528
        target_block = global_block
        current_block = framework.default_main_program().current_block()
        if current_block.idx != global_block.idx:
            assert current_block.backward_block_idx != -1, \
                "current block is not global_block, but it doesn't have backward block."
            target_block = framework.default_main_program().blocks[
                current_block.backward_block_idx]

        start = len(target_block.ops)
529
        self.helper = LayerHelper(self.__class__.__name__)
530
        self._update_param_device_map(parameters_and_grads, target_block)
C
chengduo 已提交
531
        self._create_accumulators(
532
            target_block,
C
chengduo 已提交
533
            [p[0] for p in parameters_and_grads if p[0].trainable])
534 535
        self._create_global_learning_rate()

M
minqiyang 已提交
536
        if framework.in_dygraph_mode():
537 538 539
            for param_and_grad in parameters_and_grads:
                if param_and_grad[1] is None:
                    continue
540 541
                if param_and_grad[0].trainable is True:
                    self._append_optimize_op(target_block, param_and_grad)
542 543 544 545 546 547 548
        else:
            for param_and_grad in parameters_and_grads:
                if param_and_grad[1] is None:
                    continue
                with param_and_grad[0].block.program._optimized_guard(
                        param_and_grad), name_scope("optimizer"):
                    if param_and_grad[0].trainable is True:
549 550 551 552 553
                        device = self._get_device_for_param(param_and_grad[0]
                                                            .name)
                        with device_guard(device):
                            optimize_op = self._append_optimize_op(
                                target_block, param_and_grad)
554 555 556

        # Get custom finish ops for subclasses
        # FIXME: Need to fix this once we figure out how to handle dependencies
557
        self._finish_update(target_block, parameters_and_grads)
558

559 560
        end = len(target_block.ops)
        return target_block._slice_ops(start, end)
561 562

    def _process_distribute_lookuptable(self, param_grads):
Q
Qiao Longfei 已提交
563 564 565 566 567 568 569 570 571
        """
        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
        """
572 573
        program = framework.default_main_program()
        global_block = framework.default_main_program().global_block()
Q
Qiao Longfei 已提交
574 575 576 577 578 579 580 581 582 583 584 585 586 587 588
        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:
589 590 591 592 593 594 595 596 597 598 599 600 601
            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
602 603
        return new_param_grads, (table_param, table_grad), sgd_op

604 605 606
    def _append_dgc_ops(self, param_and_grad):
        pass

607 608 609 610 611 612 613
    def backward(self,
                 loss,
                 startup_program=None,
                 parameter_list=None,
                 no_grad_set=None,
                 callbacks=None):
        """
614
        The first part of ``minimize``, do auto-diff to append backward operations for
615 616 617
        the current program.

        Args:
618 619 620 621
            loss (Variable): ``loss`` variable to run optimizations.
            startup_program (Program, optional): :ref:`api_fluid_Program` for
                initializing parameters in ``parameter_list``. The default value
                is None, at this time :ref:`api_fluid_default_startup_program` will be used.
622
            parameter_list (list, optional): List of ``Variable`` or ``Variable.name`` to update
623 624
                to minimize ``loss``. The default value is None, at this time all parameters
                will be updated.
625
            no_grad_set (set, optional): Set of ``Variable``  or ``Variable.name`` that don't need
626 627 628
                to be updated. The default value is None.
            callbacks (list, optional): list of callable objects to run when appending backward
                operator for one parameter. The default value is None.
M
minqiyang 已提交
629

630
        Return:
631 632
            list: list of (param, grad) variable pairs, param is ``Parameter``,
                grad is the gradient value corresponding to the parameter.
M
minqiyang 已提交
633

634
        Examples:
635
            See examples in ``apply_gradients``.
636
        """
637
        act_no_grad_set = None
L
Leo Chen 已提交
638
        if framework.in_dygraph_mode():
639
            pass
L
Leo Chen 已提交
640 641
        else:
            act_no_grad_set = self._get_no_grad_set(loss, no_grad_set)
G
gongweibao 已提交
642

C
chengduo 已提交
643
        self._dtype = loss.dtype
L
lujun 已提交
644
        if framework.in_dygraph_mode():
C
chengduo 已提交
645
            params_grads = []
646
            for param in self._parameter_list:
C
chengduo 已提交
647 648
                if not param.trainable:
                    continue
649
                if param._grad_ivar() is not None:
C
chengduo 已提交
650
                    # create gradient variable
651
                    grad_var = param._grad_ivar()
C
chengduo 已提交
652
                    params_grads.append((param, grad_var))
653
        else:
C
chengduo 已提交
654 655 656 657 658
            if callbacks is None:
                callbacks = [error_clip_callback]
            else:
                assert (isinstance(callbacks, list))
            program = loss.block.program
C
chengduo 已提交
659 660 661 662
            assert len(loss.shape) == 1 and loss.shape[0] == 1, \
                "The loss.shape should be (1L,), but the current loss.shape is {}. " \
                "Maybe that you should call fluid.layers.mean to process the current loss.".format(
                    loss.shape)
663 664
            parameter_list = parameter_list if parameter_list \
                else self._parameter_list
C
chengduo 已提交
665 666
            with program_guard(program, startup_program):
                params_grads = append_backward(loss, parameter_list,
667
                                               act_no_grad_set, callbacks)
C
chengduo 已提交
668 669 670 671
                # Note: since we can't use all_reduce_op now,
                #  dgc_op should be the last op of one grad.
                self._append_dgc_ops(params_grads)
        return params_grads
672 673 674 675 676 677 678 679

    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
680

681 682
        Returns:
            list: A list of operators appended to the current program.
M
minqiyang 已提交
683

684 685 686
        Examples:
            .. code-block:: python

687
                import paddle.fluid as fluid
688 689 690 691 692 693 694
                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)
        """
695

696 697 698 699 700
        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)

701 702 703 704 705
        # 'minimize(grad_clip)' or 'set_gradient_clip'
        if self._grad_clip is not None:
            params_grads = self._grad_clip(params_grads)
        else:
            params_grads = append_gradient_clip_ops(params_grads)
706 707 708 709 710 711 712 713 714 715 716 717

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

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

        return optimize_ops

C
chengduo 已提交
718 719 720 721 722 723 724 725 726 727 728 729
    def apply_optimize(self, loss, startup_program, params_grads):
        """
        Second part of `minimize`, appending optimization operators for
        given `params_grads` pairs.
        Args:
            loss (Variable): loss variable to run optimizations.
            startup_program (Program): startup_program for initializing parameters
                in `parameter_list`.
            params_grads (list): list of (param, grad) pair to do optimization.
        Returns:
            list: A list of operators appended to the current program.
        """
L
lujun 已提交
730
        if framework.in_dygraph_mode():
C
chengduo 已提交
731 732
            with program_guard(framework.default_main_program(),
                               framework.default_startup_program()):
733 734
                if self._grad_clip is not None:
                    params_grads = self._grad_clip(params_grads)
735 736
                params_grads = append_regularization_ops(params_grads,
                                                         self.regularization)
C
chengduo 已提交
737 738 739 740 741 742 743
                optimize_ops = self._create_optimization_pass(params_grads)
        else:
            program = loss.block.program
            with program_guard(program, startup_program):
                optimize_ops = self.apply_gradients(params_grads)
        return optimize_ops

G
gongweibao 已提交
744
    def _get_no_grad_set(self, loss, no_grad_set=None):
745
        no_grad_set = _get_no_grad_set_name(no_grad_set)
G
gongweibao 已提交
746 747 748 749 750 751 752 753
        parameters = loss.block.program.global_block().all_parameters()
        param_no_trainable = set(
            [param.name for param in parameters if param.trainable is False])
        # If the parameter is no trainable, it should not have a gradient.
        no_grad_set.update(param_no_trainable)

        return no_grad_set

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
    @framework.dygraph_only
    def clear_gradients(self):
        """
        Clear the gradients of all optimized parameters for model.
        
        Returns:
            None
        
        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
                import numpy as np

                with fluid.dygraph.guard():
                    value = np.arange(26).reshape(2, 13).astype("float32")
                    a = fluid.dygraph.to_variable(value)
                    linear = fluid.Linear(13, 5, dtype="float32")
                    # This can be any optimizer supported by dygraph.
                    adam = fluid.optimizer.Adam(learning_rate = 0.01, 
                                                parameter_list = linear.parameters())
                    out = linear(a)
                    out.backward()
                    adam.minimize(out)
                    adam.clear_gradients()

        """
        for p in self._parameter_list:
            if p.trainable:
                p.clear_gradient()

785
    @imperative_base.no_grad
Q
Qiao Longfei 已提交
786 787
    def minimize(self,
                 loss,
788
                 startup_program=None,
Q
Qiao Longfei 已提交
789
                 parameter_list=None,
790 791
                 no_grad_set=None,
                 grad_clip=None):
792
        """
793
        Add operations to minimize ``loss`` by updating ``parameter_list``.
M
minqiyang 已提交
794

795
        Args:
796 797 798 799
            loss (Variable): A ``Variable`` containing the value to minimize.
            startup_program (Program, optional): :ref:`api_fluid_Program` for
                initializing parameters in ``parameter_list``. The default value
                is None, at this time :ref:`api_fluid_default_startup_program` will be used.
800
            parameter_list (list, optional): List of ``Variable`` or ``Variable.name`` to update
801 802
                to minimize ``loss``. The default value is None, at this time all parameters
                will be updated.
803
            no_grad_set (set, optional): Set of ``Variable``  or ``Variable.name`` that don't need
804 805 806 807 808 809
                to be updated. The default value is None.   
            grad_clip (GradientClipBase, optional): Gradient cliping strategy, it's an instance of 
                some derived class of ``GradientClipBase`` . There are three cliping strategies 
                ( :ref:`api_fluid_clip_GradientClipByGlobalNorm` , :ref:`api_fluid_clip_GradientClipByNorm` , 
                :ref:`api_fluid_clip_GradientClipByValue` ). Default value: None, and there is no 
                gradient clipping.
Q
Qiao Longfei 已提交
810

811
        Returns:
812 813 814
            tuple: tuple (optimize_ops, params_grads), A list of operators appended
            by minimize and a list of (param, grad) variable pairs, param is
            ``Parameter``, grad is the gradient value corresponding to the parameter.
815 816 817
            The returned tuple can be passed to ``fetch_list`` in ``Executor.run()`` to 
            indicate program pruning. If so, the program will be pruned by ``feed`` and 
            ``fetch_list`` before run, see details in ``Executor``.
818 819 820

        Examples:
            Please refer to the example of current Optimizer.
Q
Qiao Longfei 已提交
821
        """
C
chengduo 已提交
822
        assert isinstance(loss, Variable), "The loss should be an Variable."
823 824 825 826 827 828
        if grad_clip is not None:
            if not isinstance(grad_clip, GradientClipBase):
                raise TypeError(
                    "'grad_clip' should be an instance of GradientClipBase's derived class"
                )
            self._grad_clip = grad_clip
829 830
        parameter_list = parameter_list if parameter_list \
            else self._parameter_list
C
chengduo 已提交
831 832 833 834 835
        params_grads = self.backward(
            loss,
            startup_program=startup_program,
            parameter_list=parameter_list,
            no_grad_set=no_grad_set)
836

C
chengduo 已提交
837 838
        optimize_ops = self.apply_optimize(
            loss, startup_program=startup_program, params_grads=params_grads)
M
minqiyang 已提交
839

Q
Qiao Longfei 已提交
840
        return optimize_ops, params_grads
Q
Qiao Longfei 已提交
841 842 843


class SGDOptimizer(Optimizer):
Q
qiaolongfei 已提交
844 845 846 847 848 849 850
    """
    Optimizer of the stochastic gradient descent algorithm.

    .. math::

        param\_out = param - learning\_rate * grad

851 852 853
    Parameters:
        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.
854 855 856
        parameter_list (list, optional):  List of ``Variable`` names to update to minimize ``loss``. \
            This parameter is required in dygraph mode. \
            The default value is None in static mode, at this time all parameters will be updated.
857 858 859 860 861
        regularization (WeightDecayRegularizer, optional): The strategy of regularization. There are two method: \
             :ref:`api_fluid_regularizer_L1Decay` , :ref:`api_fluid_regularizer_L2Decay` . If a parameter has set \
            regularizer using :ref:`api_fluid_ParamAttr` already, the regularization setting here in optimizer will be \
            ignored for this parameter. Otherwise, the regularization setting here in optimizer will take effect.  \
            Default None, meaning there is no regularization.
862 863
        name (str, optional): This parameter is used by developers to print debugging information. \
            For details, please refer to :ref:`api_guide_Name`. Default is None.
Q
qiaolongfei 已提交
864 865 866 867

    Examples:
        .. code-block:: python

868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892
            import paddle
            import paddle.fluid as fluid
            import numpy as np

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

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

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

Q
Qiao Longfei 已提交
893 894
    """

895 896 897 898 899
    def __init__(self,
                 learning_rate,
                 parameter_list=None,
                 regularization=None,
                 name=None):
Q
Qiao Longfei 已提交
900
        assert learning_rate is not None
Q
Qiao Longfei 已提交
901
        super(SGDOptimizer, self).__init__(
X
Xin Pan 已提交
902
            learning_rate=learning_rate,
903
            parameter_list=parameter_list,
X
Xin Pan 已提交
904 905
            regularization=regularization,
            name=name)
Q
Qiao Longfei 已提交
906 907
        self.type = "sgd"

908
    @no_grad
909
    def _append_optimize_op(self, block, param_and_grad):
910
        lr = self._create_param_lr(param_and_grad)
911
        if framework.in_dygraph_mode():
912 913 914
            core.ops.sgd(param_and_grad[0], lr, param_and_grad[1],
                         param_and_grad[0])
            return None
915

916
        assert isinstance(block, framework.Block)
Q
Qiao Longfei 已提交
917 918 919 920 921 922
        # create the optimize op
        sgd_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
923
                "LearningRate": lr
Q
Qiao Longfei 已提交
924
            },
M
minqiyang 已提交
925 926
            outputs={"ParamOut": param_and_grad[0]},
            stop_gradient=True)
Q
Qiao Longfei 已提交
927 928

        return sgd_op
929 930 931


class MomentumOptimizer(Optimizer):
Q
qiaolongfei 已提交
932 933 934 935 936 937 938 939 940 941 942 943 944 945
    """

    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):

946
        &\quad   param = param - (gradient + mu * velocity) * learning\_rate
Q
qiaolongfei 已提交
947 948 949

        & else:

Q
qiaolongfei 已提交
950
        &\quad   param = param - learning\_rate * velocity
Q
qiaolongfei 已提交
951

952 953 954 955
    Parameters:
        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
956 957 958
        parameter_list (list, optional):  List of ``Variable`` names to update to minimize ``loss``. \
            This parameter is required in dygraph mode. \
            The default value is None in static mode, at this time all parameters will be updated.
959
        use_nesterov (bool, optional): Enables Nesterov momentum, default is false.
960 961 962 963 964
        regularization (WeightDecayRegularizer, optional): The strategy of regularization. There are two method: \
             :ref:`api_fluid_regularizer_L1Decay` , :ref:`api_fluid_regularizer_L2Decay` . If a parameter has set \
            regularizer using :ref:`api_fluid_ParamAttr` already, the regularization setting here in optimizer will be \
            ignored for this parameter. Otherwise, the regularization setting here in optimizer will take effect.  \
            Default None, meaning there is no regularization.
965 966
        name (str, optional): This parameter is used by developers to print debugging information. \
            For details, please refer to :ref:`api_guide_Name`. Default is None.
Q
qiaolongfei 已提交
967 968 969 970

    Examples:
        .. code-block:: python

971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995
            import paddle
            import paddle.fluid as fluid
            import numpy as np

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

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

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

996 997 998
    """
    _velocity_acc_str = "velocity"

X
Xin Pan 已提交
999 1000 1001
    def __init__(self,
                 learning_rate,
                 momentum,
1002
                 parameter_list=None,
X
Xin Pan 已提交
1003 1004 1005
                 use_nesterov=False,
                 regularization=None,
                 name=None):
1006 1007
        assert learning_rate is not None
        assert momentum is not None
Q
Qiao Longfei 已提交
1008
        super(MomentumOptimizer, self).__init__(
X
Xin Pan 已提交
1009
            learning_rate=learning_rate,
1010
            parameter_list=parameter_list,
X
Xin Pan 已提交
1011 1012
            regularization=regularization,
            name=name)
1013 1014
        self.type = "momentum"
        self._momentum = momentum
1015
        self._use_nesterov = bool(use_nesterov)
1016 1017 1018 1019 1020

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

        for p in parameters:
Q
Qiao Longfei 已提交
1021
            self._add_accumulator(self._velocity_acc_str, p)
1022 1023 1024 1025 1026 1027

    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])
1028 1029 1030 1031 1032 1033 1034 1035
        lr = self._create_param_lr(param_and_grad)

        if framework.in_dygraph_mode():
            _, _ = core.ops.momentum(param_and_grad[0], param_and_grad[1],
                                     velocity_acc, lr, param_and_grad[0],
                                     velocity_acc, 'mu', self._momentum,
                                     'use_nesterov', self._use_nesterov)
            return None
1036

1037
        attrs = {"mu": self._momentum, "use_nesterov": self._use_nesterov}
1038 1039 1040 1041
        inputs = {
            "Param": [param_and_grad[0]],
            "Grad": [param_and_grad[1]],
            "Velocity": [velocity_acc],
1042
            "LearningRate": [lr]
1043 1044 1045 1046 1047 1048
        }

        outputs = {
            "ParamOut": [param_and_grad[0]],
            "VelocityOut": [velocity_acc]
        }
1049 1050 1051
        # create the momentum optimize op
        momentum_op = block.append_op(
            type=self.type,
1052 1053 1054
            inputs=inputs,
            outputs=outputs,
            attrs=attrs,
M
minqiyang 已提交
1055
            stop_gradient=True)
1056 1057

        return momentum_op
1058 1059


1060
class DGCMomentumOptimizer(Optimizer):
1061
    """
1062
    DGC (Deep Gradient Compression) Momentum Optimizer. Original paper is https://arxiv.org/abs/1712.01887
1063

G
gongweibao 已提交
1064
    DGC reduces the communication bandwidth by sending only the important gradients (sparse update):\
1065 1066
        only gradients larger than a threshold are transmitted.

G
gongweibao 已提交
1067
    To avoid losing information, DGC accumulates the rest of the gradients locally.
1068 1069 1070

    Eventually, these gradients become large enough to be transmitted.

1071
    Thus, DGC sends the large gradients immediately but eventually sends all of the gradients over time.
1072

G
gongweibao 已提交
1073
    To ensure no loss of accuracy, DGC employs momentum correction and local gradient clipping on top of the gradient sparsification to maintain model performance.
1074 1075 1076 1077

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

    This optimizer will do two things:
1078

1079 1080
        1. Compress the gradient by get TopK import value from tensor \
            and use it for allreduce to reduce network bandwidth.
1081

1082
        2. Call momentum to optimize the cost.
1083 1084

    Args:
1085 1086
        learning_rate (float|Variable): The learning rate used to update parameters. \
            It can be a float value or a Variable with one float value as a data element.
1087
        momentum (float): Momentum factor.
G
gongweibao 已提交
1088
        rampup_begin_step (int): The beginning step from which gradient compression is implemented.
1089 1090 1091 1092 1093 1094 1095
        rampup_step (int): Time steps used in sparsity warm-up 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 100, \
                it will use 0.75 at 0~19 steps, and 0.9375 at 20~39 steps, 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). \
            Default is [0.999]. For example, if the sparsity is [0.99, 0.999], \
                the top [1%, 0.1%] important element will be transmitted.
1096 1097 1098
        parameter_list (list, optional):  List of ``Variable`` names to update to minimize ``loss``. \
            This parameter is required in dygraph mode. \
            The default value is None in static mode, at this time all parameters will be updated.
1099 1100 1101
        use_nesterov (bool): Enables Nesterov momentum. True means use Nesterov. Default is False.
        local_grad_clip_norm (float, optional): Local gradient clip norm value. Optional, default is None, represent no need clip.
        num_trainers (int, optional): The number of training nodes. Optional, default is None.
1102 1103 1104 1105 1106
        regularization (WeightDecayRegularizer, optional): The strategy of regularization. There are two method: \
             :ref:`api_fluid_regularizer_L1Decay` , :ref:`api_fluid_regularizer_L2Decay` . If a parameter has set \
            regularizer using :ref:`api_fluid_ParamAttr` already, the regularization setting here in optimizer will be \
            ignored for this parameter. Otherwise, the regularization setting here in optimizer will take effect.  \
            Default None, meaning there is no regularization.
1107 1108
        name (str, optional): This parameter is used by developers to print debugging information. \
            For details, please refer to :ref:`api_guide_Name`. Default is None.
1109 1110 1111 1112

    Examples:
        .. code-block:: python

1113
            import paddle.fluid as fluid
1114
            optimizer = fluid.optimizer.DGCMomentumOptimizer(
G
gongweibao 已提交
1115 1116 1117 1118 1119
                        learning_rate=0.0001,
                        momentum=0.9,
                        rampup_step=1000,
                        rampup_begin_step=1252,
                        sparsity=[0.999, 0.999])
1120 1121

    """
1122 1123
    _u_velocity_acc_str = "_dgc_u_"
    _v_velocity_acc_str = "_dgc_v_"
1124 1125 1126 1127 1128 1129 1130

    def __init__(self,
                 learning_rate,
                 momentum,
                 rampup_begin_step,
                 rampup_step=1,
                 sparsity=[0.999],
1131
                 parameter_list=None,
1132 1133 1134 1135 1136
                 use_nesterov=False,
                 local_grad_clip_norm=None,
                 num_trainers=None,
                 regularization=None,
                 name=None):
Z
zhongpu 已提交
1137 1138
        if framework.in_dygraph_mode():
            raise Exception("In dygraph, don't support DGCMomentumOptimizer.")
1139 1140 1141 1142

        assert core.is_compiled_with_cuda(), \
            "Paddle is not compiled with CUDA. DGC is only support GPU for now."

1143 1144 1145 1146
        assert learning_rate is not None
        assert momentum is not None
        super(DGCMomentumOptimizer, self).__init__(
            learning_rate=learning_rate,
1147
            parameter_list=parameter_list,
1148 1149 1150 1151 1152
            regularization=regularization,
            name=name)
        self.type = "dgc_momentum"
        self._momentum = momentum
        self._use_nesterov = bool(use_nesterov)
1153

1154
        assert rampup_begin_step >= 0, "rampup_begin_step must >= 0"
1155
        self._rampup_begin_step = rampup_begin_step
1156 1157
        self._rampup_step = rampup_step
        self._sparsity = sparsity
1158

1159
        self._rampup_begin_step_var = None
1160
        self._global_step_var = None
1161

1162 1163 1164 1165 1166 1167 1168 1169 1170
        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
1171
            self._clip_norm = local_grad_clip_norm * (num_trainers**-0.5)
1172

1173 1174
        self.regular_type, self.regular_coeff = self._get_regularization_param(
            self.regularization)
1175
        self._grad_clip = None
1176

1177 1178 1179
    def _get_regularization_param(self, regularization):
        regular_type = 0
        regular_coeff = 0.0
1180

1181 1182
        if regularization is not None:
            regular_coeff = regularization._regularization_coeff
1183
            from .regularizer import L1Decay, L2Decay
1184 1185 1186 1187
            if isinstance(regularization, L1Decay):
                regular_type = 1
            elif isinstance(regularization, L2Decay):
                regular_type = 2
1188 1189
            else:
                assert False, 'regularization must be None|L1Decay|L2Deacy'
1190
        return regular_type, regular_coeff
1191

1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202
    def _is_use_dgc(self, param_var, grad_var):
        var_numel = abs(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 :
            return False
        return True

    def _append_optimize_op(self, block, param_and_grad):
        assert isinstance(block, framework.Block)
1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217
        velocity_acc = self._get_accumulator(self._u_velocity_acc_str,
                                             param_and_grad[0])
        assert velocity_acc is not None

        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, "use_nesterov": self._use_nesterov}
1218 1219

        if not self._is_use_dgc(param_and_grad[0], param_and_grad[1]):
1220 1221 1222
            type = "momentum"
        else:
            type = "dgc_momentum"
1223 1224 1225 1226 1227
            inputs.update({
                "current_step": self._global_step_var,
                "nranks": self._nranks_var
            })
            outputs.update({'Grad_out': param_and_grad[1]})
1228
            attrs.update({"rampup_begin_step": float(self._rampup_begin_step)})
1229 1230 1231

        # create the dgc momentum optimize op
        dgc_momentum_op = block.append_op(
1232 1233 1234 1235
            type=type,
            inputs=inputs,
            outputs=outputs,
            attrs=attrs,
1236 1237 1238
            stop_gradient=True)
        return dgc_momentum_op

1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257
    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

1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270
    def _add_nranks_var(self, name, value=-1):
        helper = LayerHelper('global_step_counter')
        counter, is_new_var = helper.create_or_get_global_variable(
            name=name, dtype='float32', shape=[1], persistable=True)
        if is_new_var:
            helper.set_variable_initializer(
                counter,
                initializer=Constant(
                    value=float(value), force_cpu=True))
            counter.stop_gradient = True

        return counter

1271 1272 1273 1274 1275 1276
    def _append_dgc_ops(self, param_and_grads):
        main_program = default_main_program()
        main_program._enable_dgc = True

        # step counter
        self._global_step_var = self._add_auto_increment_var(
G
gongweibao 已提交
1277
            counter_name=core.dgc.kDGCCounterName(), begin=0)
1278

1279 1280 1281
        self._nranks_var = self._add_nranks_var(
            name=core.dgc.kDGCNRanksName(), value=-1)

1282 1283 1284 1285 1286
        # rampup begin step var for all_reduce_op_handle
        self._rampup_begin_step_var = tensor.create_global_var(
            shape=[1],
            dtype=core.VarDesc.VarType.FP32,
            persistable=True,
G
gongweibao 已提交
1287
            name=core.dgc.kDGCRampUpBeginStepName(),
1288 1289 1290
            value=self._rampup_begin_step * 1.0,
            force_cpu=True)

1291 1292
        self.helper = LayerHelper(self.__class__.__name__)

1293
        for param_var, grad_var in param_and_grads:
1294 1295 1296
            # reuse velocity in dgc_op and dgc_momentum_op
            u_var = self._add_accumulator(self._u_velocity_acc_str, param_var)

1297
            if not self._is_use_dgc(param_var, grad_var):
1298 1299
                continue

1300
            v_var = self._add_accumulator(self._v_velocity_acc_str, param_var)
1301 1302 1303 1304 1305

            k_var = tensor.create_global_var(
                shape=[1],
                dtype=param_var.dtype,
                persistable=True,
G
gongweibao 已提交
1306
                name=param_var.name + core.dgc.kDGCKName(),
1307 1308 1309 1310 1311 1312 1313
                value=0.0,
                force_cpu=True)

            encoded_var = tensor.create_global_var(
                shape=[1],
                dtype=param_var.dtype,
                persistable=True,
G
gongweibao 已提交
1314
                name=param_var.name + core.dgc.kDGCEncodedName(),
1315 1316 1317
                value=0.0,
                force_cpu=False)

1318 1319 1320 1321 1322 1323 1324 1325
            gather_var = tensor.create_global_var(
                shape=[1],
                dtype=param_var.dtype,
                persistable=True,
                name=param_var.name + core.dgc.kDGCGatherName(),
                value=0.0,
                force_cpu=False)

1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347
            # 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,
1348
                         encoded_var, gather_var)
1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363

    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:
1364 1365
            name = unique_name.generate_with_ignorable_key(".".join(
                [helper.name, 'tmp']))
1366 1367 1368 1369 1370

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

        helper.append_op(
G
gongweibao 已提交
1371
            type="dgc_clip_by_norm",
1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383
            inputs={"X": x,
                    "current_step": self._global_step_var},
            attrs={
                "max_norm": max_norm,
                "rampup_begin_step": float(self._rampup_begin_step)
            },
            outputs={"Out": out})
        return out

    def _append_clip_norm(self, grad_var, clip_norm):
        with grad_var.block.program._backward_role_guard():
            return self._clip_by_norm(
G
gongweibao 已提交
1384
                x=grad_var, max_norm=clip_norm, name=grad_var.name)
1385 1386

    def _dgc_op(self, param_var, clip_var, grad_var, u_var, v_var, k_var,
1387
                encoded_var, gather_var):
1388 1389
        block = framework.default_main_program().global_block()
        op_maker = core.op_proto_and_checker_maker
1390

1391 1392 1393 1394 1395 1396 1397
        regular_type = self.regular_type
        regular_coeff = self.regular_coeff
        # The regularizer of the Parameters have higher priority
        if param_var.regularizer is not None:
            regular_type, regular_coeff = self._get_regularization_param(
                param_var.regularizer)

1398 1399 1400 1401 1402 1403
        dgc_op = block.append_op(
            type="dgc",
            inputs={
                "U": u_var,
                "V": v_var,
                "Grad": clip_var,
1404
                "Param": param_var,
1405 1406
                "current_step": self._global_step_var,
                "nranks": self._nranks_var,
1407 1408 1409 1410 1411 1412
            },
            outputs={
                "U_out": u_var,
                "V_out": v_var,
                "EncodeGrad": encoded_var,
                "k": k_var,
1413 1414
                "Grad_out": grad_var,
                "GatherBuff": gather_var,
1415 1416 1417 1418 1419 1420
            },
            attrs={
                "m": self._momentum,
                "sparsity": self._sparsity,
                "use_nesterov": self._use_nesterov,
                "rampup_begin_step": float(self._rampup_begin_step),
1421
                "rampup_step": float(self._rampup_step),
1422 1423
                "regular_coeff": float(regular_coeff),
                "regular_type": int(regular_type),
1424 1425 1426 1427 1428 1429 1430 1431
            },
            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])

1432
    @imperative_base.no_grad
1433 1434 1435 1436 1437 1438 1439
    def apply_gradients(self, 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)

        not_dgc_params_grads = []
        dgc_params_grads = []
1440
        # DGC clip and regularization in optimizer.backward
1441 1442 1443 1444 1445 1446
        for param, grad in params_grads:
            if not self._is_use_dgc(param, grad):
                not_dgc_params_grads.append((param, grad))
            else:
                dgc_params_grads.append((param, grad))

1447 1448 1449 1450 1451 1452
        # 'minimize(grad_clip)' or 'set_gradient_clip'
        if self._grad_clip is not None:
            not_dgc_params_grads = self._grad_clip(not_dgc_params_grads)
        else:
            not_dgc_params_grads = append_gradient_clip_ops(
                not_dgc_params_grads)
1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466

        not_dgc_params_grads = append_regularization_ops(not_dgc_params_grads,
                                                         self.regularization)

        params_grads = not_dgc_params_grads + dgc_params_grads
        params_grads = sorted(params_grads, key=lambda x: x[0].name)

        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

1467

1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482
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

1483 1484 1485 1486 1487 1488
    Parameters:
        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.
1489 1490 1491
        parameter_list (list, optional):  List of ``Variable`` names to update to minimize ``loss``. \
            This parameter is required in dygraph mode. \
            The default value is None in static mode, at this time all parameters will be updated.
1492 1493 1494 1495 1496
        regularization (WeightDecayRegularizer, optional): The strategy of regularization. There are two method: \
             :ref:`api_fluid_regularizer_L1Decay` , :ref:`api_fluid_regularizer_L2Decay` . If a parameter has set \
            regularizer using :ref:`api_fluid_ParamAttr` already, the regularization setting here in optimizer will be \
            ignored for this parameter. Otherwise, the regularization setting here in optimizer will take effect.  \
            Default None, meaning there is no regularization.
1497 1498
        name (str, optional): This parameter is used by developers to print debugging information. \
            For details, please refer to :ref:`api_guide_Name`. Default is None.
1499 1500 1501 1502

    Examples:
        .. code-block:: python

1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518
            import paddle.fluid as fluid
            import numpy as np

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

            exe = fluid.Executor(fluid.CPUPlace())
            exe.run(fluid.default_startup_program())
            exe.run(
                feed={"inp": np_inp},
                fetch_list=[out.name])
1519 1520 1521 1522 1523 1524 1525 1526
    """
    _velocity_acc_str = "velocity"

    def __init__(self,
                 learning_rate,
                 momentum,
                 lars_coeff=0.001,
                 lars_weight_decay=0.0005,
1527
                 parameter_list=None,
1528 1529 1530 1531 1532 1533
                 regularization=None,
                 name=None):
        assert learning_rate is not None
        assert momentum is not None
        super(LarsMomentumOptimizer, self).__init__(
            learning_rate=learning_rate,
1534
            parameter_list=parameter_list,
1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569
            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
1570 1571
            },
            stop_gradient=True)
1572 1573 1574 1575

        return momentum_op


1576
class AdagradOptimizer(Optimizer):
Q
qiaolongfei 已提交
1577
    """
1578 1579
    The Adaptive Gradient optimizer (Adagrad for short) can adaptively assign
    different learning rates to individual parameters.
Q
qiaolongfei 已提交
1580

1581
    The parameter ``param_out`` update rule with gradient ``grad``:
Q
qiaolongfei 已提交
1582 1583 1584 1585 1586 1587 1588

    .. math::

        moment\_out &= moment + grad * grad

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

1589 1590 1591 1592 1593 1594
    Related paper: `Adaptive Subgradient Methods for Online Learning and
    Stochastic Optimization <http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf>`_.

    The original paper does not have the ``epsilon`` attribute. It is added here
    in our implementation as also proposed `Per-parameter adaptive learning rate
    methods <http://cs231n.github.io/neural-networks-3/#ada>`_
Q
qiaolongfei 已提交
1595 1596 1597
    for numerical stability to avoid the division by zero error.

    Args:
1598 1599 1600 1601
        learning_rate (float|Variable): The learning rate used to update ``Parameter``.
            It can be a float value or a ``Variable`` with a float type.
        epsilon (float, optional): A small float value for numerical stability.
            The default value is 1e-06.
1602 1603 1604
        parameter_list (list, optional):  List of ``Variable`` names to update to minimize ``loss``. \
            This parameter is required in dygraph mode. \
            The default value is None in static mode, at this time all parameters will be updated.
1605 1606 1607 1608 1609
        regularization (WeightDecayRegularizer, optional): The strategy of regularization. There are two method: \
             :ref:`api_fluid_regularizer_L1Decay` , :ref:`api_fluid_regularizer_L2Decay` . If a parameter has set \
            regularizer using :ref:`api_fluid_ParamAttr` already, the regularization setting here in optimizer will be \
            ignored for this parameter. Otherwise, the regularization setting here in optimizer will take effect.  \
            Default None, meaning there is no regularization.
1610 1611 1612 1613 1614
        name (str, optional): Normally there is no need for user to set this property.
            For more information, please refer to :ref:`api_guide_Name`.
            The default value is None.
        initial_accumulator_value (float, optional): Initial value for moment accumulator.
            The default value is 0.0.
Q
qiaolongfei 已提交
1615 1616 1617 1618

    Examples:
        .. code-block:: python

1619
            import numpy as np
1620
            import paddle.fluid as fluid
1621 1622

            np_inp = np.array([[1.0, 2.0], [3.0, 4.0]], dtype=np.float32)
1623
            inp = fluid.data(name="inp", shape=[2, 2])
1624 1625
            out = fluid.layers.fc(inp, size=3)
            out = fluid.layers.reduce_sum(out)
1626
            optimizer = fluid.optimizer.AdagradOptimizer(learning_rate=0.2)
1627 1628 1629 1630 1631 1632 1633
            optimizer.minimize(out)

            exe = fluid.Executor(fluid.CPUPlace())
            exe.run(fluid.default_startup_program())
            exe.run(
                feed={"inp": np_inp},
                fetch_list=[out.name])
1634 1635 1636
    """
    _moment_acc_str = "moment"

X
Xin Pan 已提交
1637 1638 1639
    def __init__(self,
                 learning_rate,
                 epsilon=1.0e-6,
1640
                 parameter_list=None,
X
Xin Pan 已提交
1641
                 regularization=None,
1642
                 name=None,
X
xuezhong 已提交
1643
                 initial_accumulator_value=0.0):
1644 1645
        assert learning_rate is not None
        assert epsilon is not None
Q
Qiao Longfei 已提交
1646
        super(AdagradOptimizer, self).__init__(
X
Xin Pan 已提交
1647
            learning_rate=learning_rate,
1648
            parameter_list=parameter_list,
X
Xin Pan 已提交
1649 1650
            regularization=regularization,
            name=name)
1651 1652
        self.type = "adagrad"
        self._epsilon = epsilon
1653
        self.initial_accumulator_value = initial_accumulator_value
1654 1655 1656 1657 1658

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

        for p in parameters:
Z
zhongpu 已提交
1659 1660 1661 1662
            self._add_accumulator(
                self._moment_acc_str,
                p,
                fill_value=self.initial_accumulator_value)
1663 1664 1665 1666 1667 1668

    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])
1669
        # Create the adagrad optimizer op
1670 1671 1672 1673 1674 1675
        adagrad_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
                "Moment": moment_acc,
1676
                "LearningRate": self._create_param_lr(param_and_grad)
1677 1678 1679
            },
            outputs={"ParamOut": param_and_grad[0],
                     "MomentOut": moment_acc},
M
minqiyang 已提交
1680 1681
            attrs={"epsilon": self._epsilon},
            stop_gradient=True)
1682 1683

        return adagrad_op
1684 1685 1686


class AdamOptimizer(Optimizer):
Q
qiaolongfei 已提交
1687
    """
T
tianshuo78520a 已提交
1688
    The Adam optimizer uses an optimization described at the end
1689 1690 1691 1692 1693
    of section 2 of `Adam paper <https://arxiv.org/abs/1412.6980>`_ ,
    it can dynamically adjusts the learning rate of each parameter using
    the 1st moment estimates and the 2nd moment estimates of the gradient.
    
    The parameter ``param_out`` update rule with gradient ``grad``:
Q
qiaolongfei 已提交
1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707

    .. 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}

1708 1709
    Related paper: `Adam: A Method for Stochastic Optimization <https://arxiv.org/abs/1412.6980>`_

Q
qiaolongfei 已提交
1710
    Args:
1711 1712
        learning_rate (float|Variable, optional): The learning rate used to update ``Parameter``.
            It can be a float value or a ``Variable`` with a float type. The default value is 0.001.
1713 1714
        beta1 (float|Variable, optional): The exponential decay rate for the 1st moment estimates.
            It should be a float number or a Variable with shape [1] and data type as float32.
1715
            The default value is 0.9.
1716 1717
        beta2 (float|Variable, optional): The exponential decay rate for the 2nd moment estimates.
            It should be a float number or a Variable with shape [1] and data type as float32.
1718 1719 1720
            The default value is 0.999.
        epsilon (float, optional): A small float value for numerical stability.
            The default value is 1e-08.
1721 1722 1723
        parameter_list (list, optional):  List of ``Variable`` names to update to minimize ``loss``. \
            This parameter is required in dygraph mode. \
            The default value is None in static mode, at this time all parameters will be updated.
1724 1725 1726 1727 1728
        regularization (WeightDecayRegularizer, optional): The strategy of regularization. There are two method: \
             :ref:`api_fluid_regularizer_L1Decay` , :ref:`api_fluid_regularizer_L2Decay` . If a parameter has set \
            regularizer using :ref:`api_fluid_ParamAttr` already, the regularization setting here in optimizer will be \
            ignored for this parameter. Otherwise, the regularization setting here in optimizer will take effect.  \
            Default None, meaning there is no regularization.
1729 1730 1731 1732 1733 1734 1735 1736 1737 1738
        name (str, optional): Normally there is no need for user to set this property.
            For more information, please refer to :ref:`api_guide_Name`.
            The default value is None.
        lazy_mode (bool, optional): 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 in 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.
            The default value is False.
Q
qiaolongfei 已提交
1739 1740 1741 1742

    Examples:
        .. code-block:: python

1743 1744 1745 1746 1747 1748
            import paddle
            import paddle.fluid as fluid

            place = fluid.CPUPlace()
            main = fluid.Program()
            with fluid.program_guard(main):
1749 1750
                x = fluid.data(name='x', shape=[None, 13], dtype='float32')
                y = fluid.data(name='y', shape=[None, 1], dtype='float32')
1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765
                y_predict = fluid.layers.fc(input=x, size=1, act=None)
                cost = fluid.layers.square_error_cost(input=y_predict, label=y)
                avg_cost = fluid.layers.mean(cost)

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

                fetch_list = [avg_cost]
                train_reader = paddle.batch(
                    paddle.dataset.uci_housing.train(), batch_size=1)
                feeder = fluid.DataFeeder(place=place, feed_list=[x, y])
                exe = fluid.Executor(place)
                exe.run(fluid.default_startup_program())
                for data in train_reader():
                    exe.run(main, feed=feeder.feed(data), fetch_list=fetch_list)
Q
qiaolongfei 已提交
1766

1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783
        .. code-block:: python

            # Adam with beta1/beta2 as Variable
            import paddle
            import paddle.fluid as fluid
            import paddle.fluid.layers.learning_rate_scheduler as lr_scheduler

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

                # define beta decay variable
1784
                def get_decayed_betas(beta1_init, beta2_init, decay_steps, decay_rate):
1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812
                    global_step = lr_scheduler._decay_step_counter()

                    beta1 = fluid.layers.create_global_var(
                        shape=[1],
                        value=float(beta1_init),
                        dtype='float32',
                        # set persistable for save checkpoints and resume
                        persistable=True,
                        name="beta1")
                    beta2 = fluid.layers.create_global_var(
                        shape=[1],
                        value=float(beta2_init),
                        dtype='float32',
                        # set persistable for save checkpoints and resume
                        persistable=True,
                        name="beta2")

                    div_res = global_step / decay_steps
                    decayed_beta1 = beta1_init * (decay_rate**div_res)
                    decayed_beta2 = beta2_init * (decay_rate**div_res)
                    fluid.layers.assign(decayed_beta1, beta1)
                    fluid.layers.assign(decayed_beta2, beta2)

                    return beta1, beta2

                beta1, beta2 = get_decayed_betas(0.9, 0.99, 1e5, 0.9)
                adam_optimizer = fluid.optimizer.AdamOptimizer(
                                                    learning_rate=0.01,
1813
                                                    beta1=beta1,
1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824
                                                    beta2=beta2)
                adam_optimizer.minimize(avg_cost)

                fetch_list = [avg_cost]
                train_reader = paddle.batch(
                    paddle.dataset.uci_housing.train(), batch_size=1)
                feeder = fluid.DataFeeder(place=place, feed_list=[x, y])
                exe = fluid.Executor(place)
                exe.run(fluid.default_startup_program())
                for data in train_reader():
                    exe.run(main, feed=feeder.feed(data), fetch_list=fetch_list)
1825 1826 1827
    """
    _moment1_acc_str = "moment1"
    _moment2_acc_str = "moment2"
Q
qiaolongfei 已提交
1828 1829
    _beta1_pow_acc_str = "beta1_pow_acc"
    _beta2_pow_acc_str = "beta2_pow_acc"
1830 1831 1832 1833 1834

    def __init__(self,
                 learning_rate=0.001,
                 beta1=0.9,
                 beta2=0.999,
1835
                 epsilon=1e-8,
1836
                 parameter_list=None,
X
Xin Pan 已提交
1837
                 regularization=None,
Q
Qiao Longfei 已提交
1838
                 name=None,
Q
Qiao Longfei 已提交
1839
                 lazy_mode=False):
1840 1841 1842 1843
        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
1844
        super(AdamOptimizer, self).__init__(
X
Xin Pan 已提交
1845
            learning_rate=learning_rate,
1846
            parameter_list=parameter_list,
X
Xin Pan 已提交
1847 1848
            regularization=regularization,
            name=name)
1849 1850 1851 1852
        self.type = "adam"
        self._beta1 = beta1
        self._beta2 = beta2
        self._epsilon = epsilon
Q
Qiao Longfei 已提交
1853
        self._lazy_mode = lazy_mode
1854 1855 1856 1857 1858 1859

    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
1860 1861
            self._add_accumulator(self._moment1_acc_str, p)
            self._add_accumulator(self._moment2_acc_str, p)
Q
qiaolongfei 已提交
1862 1863 1864
            self._add_accumulator(
                name=self._beta1_pow_acc_str,
                param=p,
1865 1866
                fill_value=0.9 if isinstance(self._beta1, Variable) \
                        else self._beta1,
1867
                shape=[1],
1868
                type=core.VarDesc.VarType.LOD_TENSOR, device='cpu')
Q
qiaolongfei 已提交
1869 1870 1871
            self._add_accumulator(
                name=self._beta2_pow_acc_str,
                param=p,
1872 1873
                fill_value=0.999 if isinstance(self._beta2, Variable) \
                        else self._beta2,
1874
                shape=[1],
1875
                type=core.VarDesc.VarType.LOD_TENSOR, device='cpu')
1876 1877 1878 1879 1880 1881 1882 1883

    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
1884 1885 1886 1887
        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])
1888
        lr = self._create_param_lr(param_and_grad)
1889
        # create the adam optimize op
1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904

        if framework.in_dygraph_mode():
            _beta1 = self._beta1 if not isinstance(
                self._beta1, Variable) else self._beta1.numpy().item(0)
            _beta2 = self._beta2 if not isinstance(
                self._beta2, Variable) else self._beta2.numpy().item(0)
            _, _, _, _, _ = core.ops.adam(
                param_and_grad[0], param_and_grad[1], lr, moment1, moment2,
                beta1_pow_acc, beta2_pow_acc, param_and_grad[0], moment1,
                moment2, beta1_pow_acc, beta2_pow_acc, 'epsilon', self._epsilon,
                'lazy_mode', self._lazy_mode, 'min_row_size_to_use_multithread',
                1000, 'beta1', _beta1, 'beta2', _beta2)

            return None

1905
        inputs = {
1906 1907
            "Param": [param_and_grad[0]],
            "Grad": [param_and_grad[1]],
1908
            "LearningRate": [lr],
1909 1910 1911 1912
            "Moment1": [moment1],
            "Moment2": [moment2],
            "Beta1Pow": [beta1_pow_acc],
            "Beta2Pow": [beta2_pow_acc]
1913 1914
        }
        outputs = {
1915 1916 1917 1918 1919
            "ParamOut": [param_and_grad[0]],
            "Moment1Out": [moment1],
            "Moment2Out": [moment2],
            "Beta1PowOut": [beta1_pow_acc],
            "Beta2PowOut": [beta2_pow_acc],
1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935
        }
        attrs = {
            "epsilon": self._epsilon,
            "lazy_mode": self._lazy_mode,
            "min_row_size_to_use_multithread": 1000
        }

        if isinstance(self._beta1, Variable):
            inputs['Beta1Tensor'] = self._beta1
        else:
            attrs['beta1'] = self._beta1
        if isinstance(self._beta2, Variable):
            inputs['Beta2Tensor'] = self._beta2
        else:
            attrs['beta2'] = self._beta2

1936 1937
        adam_op = block.append_op(
            type=self.type,
1938 1939 1940
            inputs=inputs,
            outputs=outputs,
            attrs=attrs,
M
minqiyang 已提交
1941
            stop_gradient=True)
1942 1943 1944

        return adam_op

1945 1946

class AdamaxOptimizer(Optimizer):
Q
qiaolongfei 已提交
1947
    """
1948 1949 1950 1951
    The Adamax optimizer is implemented based on the Adamax Optimization 
    in Section 7 of `Adam paper <https://arxiv.org/abs/1412.6980>`_.
    The Adamax algorithm is a variant of the Adam algorithm based on the infinite norm,
    which makes the learning rate update algorithm more stable and simple.
Q
qiaolongfei 已提交
1952

1953
    The parameter ``param_out`` update rule with gradient ``grad``:
Q
qiaolongfei 已提交
1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966

    .. 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}

1967
    Related paper: `Adam: A Method for Stochastic Optimization <https://arxiv.org/abs/1412.6980>`_
Q
qiaolongfei 已提交
1968

1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980
    The original paper does not have an ``epsilon`` attribute,
    it is added here for numerical stability to prevent the division by 0 error.

    Args:
        learning_rate (float|Variable, optional): The learning rate used to update ``Parameter``.
            It can be a float value or a ``Variable`` with a float type. The default value is 0.001.
        beta1 (float, optional): The exponential decay rate for the 1st moment estimates.
            The default value is 0.9.
        beta2 (float, optional): The exponential decay rate for the 2nd moment estimates.
            The default value is 0.999.
        epsilon (float, optional): A small float value for numerical stability.
            The default value is 1e-08.
1981 1982 1983
        parameter_list (list, optional):  List of ``Variable`` names to update to minimize ``loss``. \
            This parameter is required in dygraph mode. \
            The default value is None in static mode, at this time all parameters will be updated.
1984 1985 1986 1987 1988
        regularization (WeightDecayRegularizer, optional): The strategy of regularization. There are two method: \
             :ref:`api_fluid_regularizer_L1Decay` , :ref:`api_fluid_regularizer_L2Decay` . If a parameter has set \
            regularizer using :ref:`api_fluid_ParamAttr` already, the regularization setting here in optimizer will be \
            ignored for this parameter. Otherwise, the regularization setting here in optimizer will take effect.  \
            Default None, meaning there is no regularization.
1989 1990 1991 1992 1993 1994
        name (str, optional): Normally there is no need for user to set this property.
            For more information, please refer to :ref:`api_guide_Name`.
            The default value is None.

    **Notes**:
        **Currently, AdamaxOptimizer doesn't support sparse parameter optimization.**
Q
qiaolongfei 已提交
1995

1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008
    Examples:
        .. code-block:: python

          import paddle.fluid as fluid
          import numpy

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

          train_program = fluid.Program()
          startup_program = fluid.Program()
          with fluid.program_guard(train_program, startup_program):
2009
              data = fluid.data(name='X', shape=[None, 1], dtype='float32')
2010 2011
              hidden = fluid.layers.fc(input=data, size=10)
              loss = fluid.layers.mean(hidden)
2012
              adam = fluid.optimizer.AdamaxOptimizer(learning_rate=0.2)
2013 2014 2015 2016 2017 2018 2019 2020 2021
              adam.minimize(loss)

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

          x = numpy.random.random(size=(10, 1)).astype('float32')
          outs = exe.run(program=train_program,
                        feed={'X': x},
                         fetch_list=[loss.name])
2022 2023 2024
    """
    _moment_acc_str = "moment"
    _inf_norm_acc_str = "inf_norm"
Q
qiaolongfei 已提交
2025
    _beta1_pow_acc_str = "beta1_pow_acc"
2026 2027 2028 2029 2030

    def __init__(self,
                 learning_rate=0.001,
                 beta1=0.9,
                 beta2=0.999,
2031
                 epsilon=1e-8,
2032
                 parameter_list=None,
X
Xin Pan 已提交
2033 2034
                 regularization=None,
                 name=None):
2035 2036 2037 2038
        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
2039
        super(AdamaxOptimizer, self).__init__(
X
Xin Pan 已提交
2040
            learning_rate=learning_rate,
2041
            parameter_list=parameter_list,
X
Xin Pan 已提交
2042 2043
            regularization=regularization,
            name=name)
2044 2045 2046 2047 2048 2049 2050 2051
        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
2052 2053
            self._add_accumulator(self._moment_acc_str, p)
            self._add_accumulator(self._inf_norm_acc_str, p)
Q
qiaolongfei 已提交
2054 2055 2056 2057 2058
            self._add_accumulator(
                name=self._beta1_pow_acc_str,
                param=p,
                fill_value=self._beta1,
                shape=[1])
2059 2060 2061 2062 2063 2064 2065

    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
2066 2067
        beta1_pow_acc = self._get_accumulator(self._beta1_pow_acc_str,
                                              param_and_grad[0])
2068 2069 2070 2071 2072 2073
        # create the adamax optimize op
        adamax_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
2074
                "LearningRate": self._create_param_lr(param_and_grad),
2075 2076
                "Moment": moment,
                "InfNorm": inf_norm,
Q
qiaolongfei 已提交
2077
                "Beta1Pow": beta1_pow_acc
2078 2079 2080 2081 2082 2083 2084 2085 2086 2087
            },
            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
2088 2089
            },
            stop_gradient=True)
2090 2091 2092

        return adamax_op

2093
    def _finish_update(self, block, parameters_and_grads):
2094 2095 2096
        """Update Beta1 Power accumulator
        """
        assert isinstance(block, framework.Block)
2097
        for param, grad in parameters_and_grads:
C
chengduo 已提交
2098
            if grad is None or param.trainable is False:
2099
                continue
X
Xin Pan 已提交
2100 2101
            with param.block.program._optimized_guard(
                [param, grad]), name_scope('adamx'):
2102 2103
                beta1_pow_acc = self._get_accumulator(self._beta1_pow_acc_str,
                                                      param)
2104
                block.append_op(
2105 2106 2107
                    type="scale",
                    inputs={"X": beta1_pow_acc},
                    outputs={"Out": beta1_pow_acc},
M
minqiyang 已提交
2108 2109
                    attrs={"scale": self._beta1},
                    stop_gradient=True)
2110 2111


2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149
class DpsgdOptimizer(Optimizer):
    """
    We implement the Dpsgd optimizer according to CCS16 paper -
    Deep Learning with Differential Privacy.

    Examples:
        .. code-block:: python

          import paddle.fluid as fluid
          import numpy

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

          train_program = fluid.Program()
          startup_program = fluid.Program()
          with fluid.program_guard(train_program, startup_program):
              data = fluid.layers.data(name='X', shape=[1], dtype='float32')
              hidden = fluid.layers.fc(input=data, size=10)
              loss = fluid.layers.mean(hidden)
              optimizer = fluid.optimizer.Dpsgd(learning_rate=0.01, clip=10.0, batch_size=16.0, sigma=1.0)
              optimizer.minimize(loss)

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

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

    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.
        clip (float): clipping threshold
        batch_size (float): batch size.
        sigma (float): for gaussian noise.
2150 2151 2152
        parameter_list (list, optional):  List of ``Variable`` names to update to minimize ``loss``. \
            This parameter is required in dygraph mode. \
            The default value is None in static mode, at this time all parameters will be updated.
2153 2154 2155 2156 2157 2158 2159 2160
    Notes:
       Currently, DpsgdOptimizer doesn't support sparse parameter optimization.
    """

    def __init__(self,
                 learning_rate=0.001,
                 clip=0.9,
                 batch_size=0.999,
2161 2162
                 sigma=1e-8,
                 parameter_list=None):
2163 2164 2165 2166
        assert learning_rate is not None
        assert clip is not None
        assert batch_size is not None
        assert sigma is not None
2167 2168
        super(DpsgdOptimizer, self).__init__(
            learning_rate=learning_rate, parameter_list=parameter_list)
2169 2170 2171 2172
        self.type = "dpsgd"
        self._clip = clip
        self._batch_size = batch_size
        self._sigma = sigma
Z
zhongpu 已提交
2173 2174 2175 2176 2177 2178 2179
        '''
        Note(wangzhongpu):
        This property is only used for debugging, do not need to set it!
        Dpsgd operator use time(NULL) as random seed to generate random number.
        However, during debugging, we need determinated result, so we will set self._seed to a fixed number.
        '''
        self._seed = None
2180 2181 2182 2183 2184

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

        # create the dpsgd optimize op
Z
zhongpu 已提交
2185 2186 2187
        if self._seed == None:
            self._seed = 0

2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198
        dpsgd_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
                "LearningRate": self._create_param_lr(param_and_grad)
            },
            outputs={"ParamOut": param_and_grad[0]},
            attrs={
                "clip": self._clip,
                "batch_size": self._batch_size,
Z
zhongpu 已提交
2199 2200
                "sigma": self._sigma,
                "seed": self._seed
2201 2202 2203 2204 2205 2206
            },
            stop_gradient=True)

        return dpsgd_op


2207
class DecayedAdagradOptimizer(Optimizer):
2208
    """
2209 2210 2211
    The Decayed Adagrad optimizer can be seen as an Adagrad algorithm that introduces
    the decay rate to solve the problem of a sharp drop in the learning rate
    during model training when using the AdagradOptimizer.
2212

2213
    The parameter ``param_out`` update rule with gradient ``grad``:
2214 2215 2216 2217 2218 2219 2220

    .. math::

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

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

2221 2222 2223 2224
    Related paper: `Adaptive Subgradient Methods for Online Learning and Stochastic
    Optimization <http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf>`_.

    The original paper does not have an ``epsilon`` attribute. It is added here for numerical
2225 2226 2227
    stability to avoid the division by zero error.

    Args:
2228 2229 2230 2231 2232
        learning_rate (float|Variable): The learning rate used to update ``Parameter``.
            It can be a float value or a ``Variable`` with a float type.
        decay (float, optional): The decay rate. The default value is 0.95.
        epsilon (float, optional): A small float value for numerical stability.
            The default value is 1e-06.
2233 2234 2235
        parameter_list (list, optional):  List of ``Variable`` names to update to minimize ``loss``. \
            This parameter is required in dygraph mode. \
            The default value is None in static mode, at this time all parameters will be updated.
2236 2237 2238 2239 2240
        regularization (WeightDecayRegularizer, optional): The strategy of regularization. There are two method: \
             :ref:`api_fluid_regularizer_L1Decay` , :ref:`api_fluid_regularizer_L2Decay` . If a parameter has set \
            regularizer using :ref:`api_fluid_ParamAttr` already, the regularization setting here in optimizer will be \
            ignored for this parameter. Otherwise, the regularization setting here in optimizer will take effect.  \
            Default None, meaning there is no regularization.
2241 2242 2243 2244 2245 2246
        name (str, optional): Normally there is no need for user to set this property.
            For more information, please refer to :ref:`api_guide_Name`.
            The default value is None.

    **Notes**:
        **Currently, DecayedAdagradOptimizer doesn't support sparse parameter optimization.**
2247 2248 2249 2250

    Examples:
        .. code-block:: python

2251 2252
            import paddle.fluid as fluid

2253 2254 2255 2256
            x = fluid.data( name='x', shape=[None, 10], dtype='float32' )
            trans = fluid.layers.fc( x, 100 )
            cost = fluid.layers.reduce_mean( trans )
            optimizer = fluid.optimizer.DecayedAdagradOptimizer(learning_rate=0.2)
2257
            optimizer.minimize(cost)
2258 2259 2260
    """
    _moment_acc_str = "moment"

X
Xin Pan 已提交
2261 2262 2263 2264
    def __init__(self,
                 learning_rate,
                 decay=0.95,
                 epsilon=1.0e-6,
2265
                 parameter_list=None,
X
Xin Pan 已提交
2266 2267
                 regularization=None,
                 name=None):
2268 2269 2270 2271
        assert learning_rate is not None
        assert decay is not None
        assert epsilon is not None

Q
Qiao Longfei 已提交
2272
        super(DecayedAdagradOptimizer, self).__init__(
X
Xin Pan 已提交
2273
            learning_rate=learning_rate,
2274
            parameter_list=parameter_list,
X
Xin Pan 已提交
2275 2276
            regularization=regularization,
            name=name)
2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303
        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},
2304 2305
            attrs={"epsilon": self._epsilon,
                   "decay": self._decay},
M
minqiyang 已提交
2306
            stop_gradient=True)
2307 2308

        return decayed_adagrad_op
2309 2310


2311
class AdadeltaOptimizer(Optimizer):
2312
    """
Z
Zeng Jinle 已提交
2313
    **Notes: This API does not support sparse parameter optimization.**
Q
qiaolongfei 已提交
2314

Z
Zeng Jinle 已提交
2315
    Adadelta Optimizer. Please refer to this for details:
Z
Zeng Jinle 已提交
2316 2317 2318
    `ADADELTA: AN ADAPTIVE LEARNING RATE METHOD <https://arxiv.org/abs/1212.5701>`_.

    The update is done as follows:
2319

Z
Zeng Jinle 已提交
2320 2321
    .. math::

Z
Zeng Jinle 已提交
2322
        E(g_t^2) &= \\rho * E(g_{t-1}^2) + (1-\\rho) * g^2
2323

Z
Zeng Jinle 已提交
2324
        learning\_rate &= \sqrt{ ( E(dx_{t-1}^2) + \\epsilon ) / ( E(g_t^2) + \\epsilon ) }
Z
Zeng Jinle 已提交
2325

Z
Zeng Jinle 已提交
2326
        E(dx_t^2) &= \\rho * E(dx_{t-1}^2) + (1-\\rho) * (-g*learning\_rate)^2
2327 2328

    Args:
Z
Zeng Jinle 已提交
2329 2330 2331
        learning_rate (float|Variable): global learning rate.
        epsilon (float): a small float number for numeric stability. Default 1.0e-6.
        rho (float): a floating point value indicating the decay rate. Default 0.95.
2332 2333 2334
        parameter_list (list, optional):  List of ``Variable`` names to update to minimize ``loss``. \
            This parameter is required in dygraph mode. \
            The default value is None in static mode, at this time all parameters will be updated.
2335 2336 2337 2338 2339
        regularization (WeightDecayRegularizer, optional): The strategy of regularization. There are two method: \
             :ref:`api_fluid_regularizer_L1Decay` , :ref:`api_fluid_regularizer_L2Decay` . If a parameter has set \
            regularizer using :ref:`api_fluid_ParamAttr` already, the regularization setting here in optimizer will be \
            ignored for this parameter. Otherwise, the regularization setting here in optimizer will take effect.  \
            Default None, meaning there is no regularization.
2340 2341 2342
        name (str, optional): The default value is None. Normally there is no need for user
                to set this property. For more information, please refer to
                :ref:`api_guide_Name` .
2343 2344 2345 2346

    Examples:
        .. code-block:: python

2347
            import paddle.fluid as fluid
Z
Zeng Jinle 已提交
2348

2349
            image = fluid.data(name='image', shape=[None, 28], dtype='float32')
Z
Zeng Jinle 已提交
2350 2351
            fc = fluid.layers.fc(image, size=10)
            cost = fluid.layers.reduce_mean(fc)
2352 2353
            optimizer = fluid.optimizer.Adadelta(
                learning_rate=0.0003, epsilon=1.0e-6, rho=0.95)
C
chengduo 已提交
2354

Z
Zeng Jinle 已提交
2355 2356 2357 2358
            # optimizer_ops is a list of optimizer operators to update parameters
            # params_grads is a list of (param, param_grad), where param is each
            # parameter and param_grad is the gradient variable of param.
            optimizer_ops, params_grads = optimizer.minimize(cost)
2359
    """
2360

2361 2362 2363
    _avg_squared_grad_acc_str = "_avg_squared_grad"
    _avg_squared_update_acc_str = "_avg_squared_update"

X
Xin Pan 已提交
2364 2365 2366 2367
    def __init__(self,
                 learning_rate,
                 epsilon=1.0e-6,
                 rho=0.95,
2368
                 parameter_list=None,
X
Xin Pan 已提交
2369 2370
                 regularization=None,
                 name=None):
2371 2372 2373 2374 2375 2376
        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.")
2377
        super(AdadeltaOptimizer, self).__init__(
X
Xin Pan 已提交
2378
            learning_rate=learning_rate,
2379
            parameter_list=parameter_list,
X
Xin Pan 已提交
2380 2381
            regularization=regularization,
            name=name)
2382 2383 2384 2385 2386
        self.type = "adadelta"
        self._epsilon = epsilon
        self._rho = rho

    def _create_accumulators(self, block, parameters):
2387 2388
        if not isinstance(block, framework.Block):
            raise TypeError("block is not instance of framework.Block.")
2389 2390 2391 2392 2393 2394

        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):
2395 2396
        if not isinstance(block, framework.Block):
            raise TypeError("block is not instance of framework.Block.")
2397 2398 2399 2400 2401 2402 2403 2404 2405 2406 2407 2408 2409 2410 2411 2412 2413 2414 2415 2416 2417

        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
2418 2419
                   "rho": self._rho},
            stop_gradient=True)
2420 2421 2422 2423

        return adadelta_op


Q
qingqing01 已提交
2424 2425 2426 2427 2428 2429 2430 2431 2432 2433
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
2434
        r(w, t) & = \\rho r(w, t-1) + (1 - \\rho)(\\nabla Q_{i}(w))^2
Q
qingqing01 已提交
2435 2436 2437 2438

        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
2439
    each weight. Then dividing the gradient by :math:`sqrt{v(w,t)}`.
Q
qingqing01 已提交
2440 2441 2442 2443 2444 2445

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

    ..  math::

Q
qiaolongfei 已提交
2446
        r(w, t) & = \\rho r(w, t-1) + (1 - \\rho)(\\nabla Q_{i}(w))^2
Q
qingqing01 已提交
2447

2448 2449 2450 2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461
        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
2462 2463 2464 2465
            \\epsilon}} \\nabla Q_{i}(w)

        w & = w - v(w, t)

Q
qiaolongfei 已提交
2466
    where, :math:`\\rho` is a hyperparameter and typical values are 0.9, 0.95
Q
qingqing01 已提交
2467 2468 2469 2470 2471
    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.


2472 2473 2474
    Parameters:
        learning_rate(float): Global learning rate.
        rho(float): rho is :math: `\\rho` in equation, default is 0.95.
Q
qingqing01 已提交
2475
        epsilon(float): :math: `\\epsilon` in equation is smoothing term to
2476
            avoid division by zero, default is 1e-6.
Q
qiaolongfei 已提交
2477
        momentum(float): :math:`\\beta` in equation is the momentum term,
2478
            default is 0.0.
2479 2480 2481 2482
        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.
2483 2484 2485
        parameter_list (list, optional):  List of ``Variable`` names to update to minimize ``loss``. \
            This parameter is required in dygraph mode. \
            The default value is None in static mode, at this time all parameters will be updated.
2486 2487 2488 2489 2490
        regularization (WeightDecayRegularizer, optional): The strategy of regularization. There are two method: \
             :ref:`api_fluid_regularizer_L1Decay` , :ref:`api_fluid_regularizer_L2Decay` . If a parameter has set \
            regularizer using :ref:`api_fluid_ParamAttr` already, the regularization setting here in optimizer will be \
            ignored for this parameter. Otherwise, the regularization setting here in optimizer will take effect.  \
            Default None, meaning there is no regularization.
2491 2492
        name (str, optional): This parameter is used by developers to print debugging information. \
            For details, please refer to :ref:`api_guide_Name`. Default is None.
Q
qingqing01 已提交
2493 2494 2495 2496 2497 2498 2499

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

    Examples:
          .. code-block:: python

2500 2501 2502 2503 2504 2505 2506 2507 2508 2509 2510 2511 2512 2513 2514 2515 2516 2517 2518 2519 2520 2521 2522 2523 2524
            import paddle
            import paddle.fluid as fluid
            import numpy as np

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

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

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

Q
qingqing01 已提交
2525 2526 2527 2528
    """

    _momentum_acc_str = "momentum"
    _mean_square_acc_str = "mean_square"
2529
    _mean_grad_acc_str = "mean_grad"
Q
qingqing01 已提交
2530 2531 2532 2533 2534 2535

    def __init__(self,
                 learning_rate,
                 rho=0.95,
                 epsilon=1.0e-6,
                 momentum=0.0,
2536
                 centered=False,
2537
                 parameter_list=None,
X
Xin Pan 已提交
2538 2539
                 regularization=None,
                 name=None):
Q
qingqing01 已提交
2540
        super(RMSPropOptimizer, self).__init__(
X
Xin Pan 已提交
2541
            learning_rate=learning_rate,
2542
            parameter_list=parameter_list,
X
Xin Pan 已提交
2543 2544
            regularization=regularization,
            name=name)
Q
qingqing01 已提交
2545 2546 2547 2548 2549 2550 2551 2552 2553 2554 2555 2556 2557
        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
2558
        self._centered = centered
Q
qingqing01 已提交
2559 2560 2561 2562 2563 2564 2565 2566

    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)
2567
            self._add_accumulator(self._mean_grad_acc_str, p)
Q
qingqing01 已提交
2568 2569 2570 2571 2572 2573 2574 2575 2576

    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])
2577 2578
        mean_grad_acc = self._get_accumulator(self._mean_grad_acc_str,
                                              param_and_grad[0])
Q
qingqing01 已提交
2579 2580 2581 2582 2583 2584 2585
        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,
2586
                "MeanGrad": mean_grad_acc,
Q
qingqing01 已提交
2587 2588 2589 2590 2591
                "LearningRate": self._create_param_lr(param_and_grad),
            },
            outputs={
                "ParamOut": param_and_grad[0],
                "MomentOut": momentum_acc,
2592 2593
                "MeanSquareOut": mean_square_acc,
                "MeanGradOut": mean_grad_acc
Q
qingqing01 已提交
2594 2595 2596 2597
            },
            attrs={
                "epsilon": self._epsilon,
                "decay": self._rho,
2598 2599
                "momentum": self._momentum,
                "centered": self._centered
M
minqiyang 已提交
2600 2601
            },
            stop_gradient=True)
Q
qingqing01 已提交
2602 2603 2604 2605

        return rmsprop_op


Q
qiaolongfei 已提交
2606 2607 2608 2609 2610 2611 2612 2613 2614 2615 2616 2617 2618 2619 2620 2621 2622 2623 2624 2625 2626 2627 2628 2629 2630 2631 2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645
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

2646 2647 2648 2649 2650
    Parameters:
        learning_rate (float|Variable): Global learning rate.
        l1 (float): L1 regularization strength, default is 0.0.
        l2 (float): L2 regularization strength, default is 0.0.
        lr_power (float): Learning Rate Power, default is -0.5.
2651 2652 2653
        parameter_list (list, optional):  List of ``Variable`` names to update to minimize ``loss``. \
            This parameter is required in dygraph mode. \
            The default value is None in static mode, at this time all parameters will be updated.
2654 2655 2656 2657 2658
        regularization (WeightDecayRegularizer, optional): The strategy of regularization. There are two method: \
             :ref:`api_fluid_regularizer_L1Decay` , :ref:`api_fluid_regularizer_L2Decay` . If a parameter has set \
            regularizer using :ref:`api_fluid_ParamAttr` already, the regularization setting here in optimizer will be \
            ignored for this parameter. Otherwise, the regularization setting here in optimizer will take effect.  \
            Default None, meaning there is no regularization.
2659 2660
        name (str, optional): This parameter is used by developers to print debugging information. \
            For details, please refer to :ref:`api_guide_Name`. Default is None.
Q
qiaolongfei 已提交
2661 2662 2663 2664 2665 2666 2667

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

    Examples:
          .. code-block:: python

2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691
            import paddle
            import paddle.fluid as fluid
            import numpy as np

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

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

                fetch_list = [avg_cost]
                train_reader = paddle.batch(
                    paddle.dataset.uci_housing.train(), batch_size=1)
                feeder = fluid.DataFeeder(place=place, feed_list=[x, y])
                exe = fluid.Executor(place)
                exe.run(fluid.default_startup_program())
                for data in train_reader():
                    exe.run(main, feed=feeder.feed(data), fetch_list=fetch_list)
C
chengduo 已提交
2692

2693
    NOTE:
C
chengduo 已提交
2694
       Currently, FtrlOptimizer doesn't support sparse parameter optimization.
Q
qiaolongfei 已提交
2695 2696 2697 2698 2699
    """

    _squared_acc_str = "squared"
    _linear_acc_str = "linear"

X
Xin Pan 已提交
2700 2701 2702 2703 2704
    def __init__(self,
                 learning_rate,
                 l1=0.0,
                 l2=0.0,
                 lr_power=-0.5,
2705
                 parameter_list=None,
X
Xin Pan 已提交
2706 2707
                 regularization=None,
                 name=None):
Q
qiaolongfei 已提交
2708
        super(FtrlOptimizer, self).__init__(
X
Xin Pan 已提交
2709
            learning_rate=learning_rate,
2710
            parameter_list=parameter_list,
X
Xin Pan 已提交
2711 2712
            regularization=regularization,
            name=name)
Q
qiaolongfei 已提交
2713 2714 2715 2716 2717 2718 2719 2720 2721 2722 2723 2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748 2749 2750 2751 2752
        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
2753 2754
                   "lr_power": self._lr_power},
            stop_gradient=True)
Q
qiaolongfei 已提交
2755 2756 2757 2758

        return ftrl_op


Y
Yibing Liu 已提交
2759 2760 2761 2762 2763 2764
class LambOptimizer(AdamOptimizer):
    """
    LAMB (Layer-wise Adaptive Moments optimizer for Batching training) Optimizer.

    LAMB Optimizer is designed to scale up the batch size of training without losing 
    accuracy, which supports adaptive element-wise updating and accurate layer-wise 
Y
Yibing Liu 已提交
2765 2766
    correction. For more information, please refer to `Large Batch Optimization for 
    Deep Learning: Training BERT in 76 minutes <https://arxiv.org/abs/1904.00962>`_ .
Y
Yibing Liu 已提交
2767 2768 2769 2770 2771

    The updating of parameters follows:

    ..  math::

Y
Yibing Liu 已提交
2772
        m_t &= \\beta_1 m_{t - 1}+ (1 - \\beta_1)g_t 
Y
Yibing Liu 已提交
2773

Y
Yibing Liu 已提交
2774
        v_t &= \\beta_2 v_{t - 1}  + (1 - \\beta_2)g_t^2
Y
Yibing Liu 已提交
2775

Y
Yibing Liu 已提交
2776
        r_t &= \\frac{m_t}{\\sqrt{v_t}+\\epsilon}
Y
Yibing Liu 已提交
2777

Y
Yibing Liu 已提交
2778
        w_t &= w_{t-1} -\\eta_t \\frac{\\left \| w_{t-1}\\right \|}{\\left \| r_t + \\lambda w_{t-1}\\right \|} (r_t + \\lambda w_{t-1})
Y
Yibing Liu 已提交
2779 2780 2781 2782 2783 2784


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

    Args:
Y
Yibing Liu 已提交
2785 2786 2787 2788 2789 2790 2791 2792
        learning_rate (float|Variable, optional): the learning rate used to update parameters. \
            Can be a float value or a Variable with data type float32. Default 0.001.
        lamb_weight_decay (float, optional): The LAMB weight decay rate. Default 0.01.
        beta1 (float, optional): The exponential decay rate for the 1st moment estimates.
            Default 0.9.
        beta2 (float, optional): The exponential decay rate for the 2nd moment estimates.
            Default 0.999.
        epsilon (float, optional): A small float value for numerical stability. Default 1e-6.
2793 2794 2795
        parameter_list (list, optional):  List of ``Variable`` names to update to minimize ``loss``. \
            This parameter is required in dygraph mode. \
            The default value is None in static mode, at this time all parameters will be updated.
2796 2797 2798 2799 2800
        regularization (WeightDecayRegularizer, optional): The strategy of regularization. There are two method: \
             :ref:`api_fluid_regularizer_L1Decay` , :ref:`api_fluid_regularizer_L2Decay` . If a parameter has set \
            regularizer using :ref:`api_fluid_ParamAttr` already, the regularization setting here in optimizer will be \
            ignored for this parameter. Otherwise, the regularization setting here in optimizer will take effect.  \
            Default None, meaning there is no regularization.
Y
Yibing Liu 已提交
2801 2802 2803 2804 2805
        exclude_from_weight_decay_fn (function|None): Exclude a parameter from weight 
            decay when **exclude_from_weight_decay_fn(parameter)** returns true. 
            Default None.
        name(str|None): For detailed information, please refer to 
            :ref:`api_guide_Name` . Usually name is no need to set and None by default.
Y
Yibing Liu 已提交
2806 2807 2808 2809 2810 2811

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

Y
Yibing Liu 已提交
2812
            data = fluid.data(name='x', shape=[-1, 5], dtype='float32')
Y
Yibing Liu 已提交
2813 2814 2815
            hidden = fluid.layers.fc(input=data, size=10)
            cost = fluid.layers.mean(hidden)

Y
Yibing Liu 已提交
2816 2817 2818 2819 2820
            def exclude_fn(param):
                return param.name.endswith('.b_0')

            optimizer = fluid.optimizer.Lamb(learning_rate=0.002,
                                             exclude_from_weight_decay_fn=exclude_fn)
Y
Yibing Liu 已提交
2821 2822 2823 2824
            optimizer.minimize(cost)
    """
    _moment1_acc_str = "moment1"
    _moment2_acc_str = "moment2"
Y
Yibing Liu 已提交
2825
    # these two not used in op temporarily
Y
Yibing Liu 已提交
2826 2827 2828 2829 2830 2831 2832 2833 2834
    _beta1_pow_acc_str = "beta1_pow_acc"
    _beta2_pow_acc_str = "beta2_pow_acc"

    def __init__(self,
                 learning_rate=0.001,
                 lamb_weight_decay=0.01,
                 beta1=0.9,
                 beta2=0.999,
                 epsilon=1e-6,
2835
                 parameter_list=None,
Y
Yibing Liu 已提交
2836
                 regularization=None,
Y
Yibing Liu 已提交
2837
                 exclude_from_weight_decay_fn=None,
Y
Yibing Liu 已提交
2838 2839 2840 2841 2842 2843 2844 2845
                 name=None):
        assert learning_rate is not None
        assert lamb_weight_decay is not None
        assert beta1 is not None
        assert beta2 is not None
        assert epsilon is not None
        super(LambOptimizer, self).__init__(
            learning_rate=learning_rate,
2846
            parameter_list=parameter_list,
Y
Yibing Liu 已提交
2847 2848 2849 2850 2851 2852 2853
            regularization=regularization,
            beta1=beta1,
            beta2=beta2,
            epsilon=epsilon,
            name=name)
        self.type = "lamb"
        self._weight_decay = lamb_weight_decay
Y
Yibing Liu 已提交
2854
        self._exclude_from_weight_decay_fn = exclude_from_weight_decay_fn
Y
Yibing Liu 已提交
2855 2856 2857

    def _append_optimize_op(self, block, param_and_grad):
        assert isinstance(block, framework.Block)
2858
        block.program._use_lamb = True
Y
Yibing Liu 已提交
2859 2860 2861 2862 2863 2864 2865 2866 2867 2868

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

Y
Yibing Liu 已提交
2869 2870 2871 2872 2873 2874
        if self._exclude_from_weight_decay_fn is not None \
            and self._exclude_from_weight_decay_fn(param_and_grad[0]):
            weight_decay = 0.0
        else:
            weight_decay = self._weight_decay

Y
Yibing Liu 已提交
2875 2876 2877 2878 2879 2880 2881 2882 2883 2884 2885 2886 2887 2888 2889 2890 2891 2892 2893 2894 2895
        # create the lamb optimize op
        lamb_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
                "LearningRate": self._create_param_lr(param_and_grad),
                "Moment1": moment1,
                "Moment2": moment2,
                "Beta1Pow": beta1_pow_acc,
                "Beta2Pow": beta2_pow_acc
            },
            outputs={
                "ParamOut": param_and_grad[0],
                "Moment1Out": moment1,
                "Moment2Out": moment2
            },
            attrs={
                "beta1": self._beta1,
                "beta2": self._beta2,
                "epsilon": self._epsilon,
Y
Yibing Liu 已提交
2896
                "weight_decay": weight_decay
Y
Yibing Liu 已提交
2897 2898 2899 2900 2901 2902
            },
            stop_gradient=True)

        return lamb_op


2903 2904 2905 2906 2907 2908 2909 2910 2911 2912 2913 2914 2915
# 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
2916
Dpsgd = DpsgdOptimizer
2917
DecayedAdagrad = DecayedAdagradOptimizer
2918
Adadelta = AdadeltaOptimizer
Q
qingqing01 已提交
2919
RMSProp = RMSPropOptimizer
Q
qiaolongfei 已提交
2920
Ftrl = FtrlOptimizer
2921
LarsMomentum = LarsMomentumOptimizer
Y
Yibing Liu 已提交
2922
Lamb = LambOptimizer
2923 2924 2925


class ModelAverage(Optimizer):
2926 2927 2928 2929 2930 2931 2932 2933 2934 2935 2936 2937 2938 2939 2940 2941 2942 2943 2944
    """
    The ModelAverage optimizer accumulates specific continuous historical parameters
    during training. The accumulated historical range can be controlled by the passed
    ``average_window_rate`` argument. The averaged ``Parameter`` are used in the prediction,
    which usually can improve the accuracy of the prediction.

    Accumulate the average of the ``Parameter`` in the sliding window, the result will be saved
    in a temporary variable, can be applied to the current model's ``Parameter`` by calling
    the ``apply()`` method, and the current model ``Parameter`` can be restored by calling
    the ``restore()`` method.

    The window size for calculating the average is determined by ``average_window_rate``,
    ``min_average_window``, ``max_average_window`` and the current ``Parameter`` update times (num_updates).

    When the cumulative times (num_accumulates) is greater than the specific window
    threshold (average_window), the accumulated ``Parameter`` temporary variable is set to 0.0.
    The following example will help to understand the role of these arguments:

    ::
2945

2946 2947 2948 2949 2950 2951 2952 2953 2954
        if num_accumulates >= min_average_window and num_accumulates >= min(max_average_window, num_updates * average_window_rate):
            num_accumulates = 0

    In the above conditional judgment statement, ``num_accumulates`` indicates the current
    accumulated number, which can be abstractly understood as the length of the cumulative window.
    The length of the window must be at least the length set by the ``min_average_window`` argument,
    and cannot exceed the length specified by the ``max_average_window`` argument or
    ``num_updates * average_window_rate``, where ``num_updates`` indicates the current ``Parameter``
    update times, ``average_window_rate`` is a coefficient that calculates the length of the window.
2955 2956

    Args:
2957 2958 2959
        average_window_rate (float): The calculate ratio of the window length relative to ``Parameter`` update times.
        min_average_window (int, optional): the minimum size of average window length. The default value is 10000.
        max_average_window (int, optional): The maximum size of average window length. The default value is 10000.
2960 2961 2962 2963 2964
        regularization (WeightDecayRegularizer, optional): The strategy of regularization. There are two method: \
             :ref:`api_fluid_regularizer_L1Decay` , :ref:`api_fluid_regularizer_L2Decay` . If a parameter has set \
            regularizer using :ref:`api_fluid_ParamAttr` already, the regularization setting here in optimizer will be \
            ignored for this parameter. Otherwise, the regularization setting here in optimizer will take effect.  \
            Default None, meaning there is no regularization.
2965 2966 2967
        name (str, optional): Normally there is no need for user to set this property.
            For more information, please refer to :ref:`api_guide_Name`.
            The default value is None.
2968

2969
    Examples:
Q
qiaolongfei 已提交
2970 2971 2972

      .. code-block:: python

2973 2974 2975 2976 2977 2978
        import paddle.fluid as fluid
        import numpy

        # First create the Executor.
        place = fluid.CPUPlace()  # fluid.CUDAPlace(0)
        exe = fluid.Executor(place)
2979

2980 2981 2982 2983
        train_program = fluid.Program()
        startup_program = fluid.Program()
        with fluid.program_guard(train_program, startup_program):
            # build net
2984
            data = fluid.data(name='X', shape=[None, 1], dtype='float32')
2985 2986 2987 2988 2989 2990 2991 2992
            hidden = fluid.layers.fc(input=data, size=10)
            loss = fluid.layers.mean(hidden)
            optimizer = fluid.optimizer.Momentum(learning_rate=0.2, momentum=0.1)
            optimizer.minimize(loss)

            # build ModelAverage optimizer
            model_average = fluid.optimizer.ModelAverage(0.15,
                                                         min_average_window=10000,
2993
                                                         max_average_window=12500)
2994 2995

            exe.run(startup_program)
2996 2997 2998 2999 3000
            for i in range(12500):
                x = numpy.random.random(size=(10, 1)).astype('float32')
                outs = exe.run(program=train_program,
                               feed={'X': x},
                               fetch_list=[loss.name])
3001 3002

            # apply ModelAverage
3003
            with model_average.apply(exe):
3004 3005 3006 3007
                x = numpy.random.random(size=(10, 1)).astype('float32')
                exe.run(program=train_program,
                        feed={'X': x},
                        fetch_list=[loss.name])
3008 3009 3010
    """

    def __init__(self,
W
wanghaoshuang 已提交
3011
                 average_window_rate,
3012 3013
                 min_average_window=10000,
                 max_average_window=10000,
X
Xin Pan 已提交
3014 3015
                 regularization=None,
                 name=None):
Z
zhongpu 已提交
3016 3017
        if framework.in_dygraph_mode():
            raise Exception("In dygraph, don't support ModelAverage.")
X
Xin Pan 已提交
3018 3019
        super(ModelAverage, self).__init__(
            0.0, regularization=regularization, name=name)
3020 3021 3022
        self.average_window = average_window_rate
        self.min_average_window = min_average_window
        self.max_average_window = max_average_window
3023

3024
        self.params_grads = []
3025 3026
        for param in framework.default_main_program().global_block(
        ).all_parameters():
3027
            if param.do_model_average != False:
3028
                grad = param.block.create_var(
3029 3030
                    name=unique_name.generate_with_ignorable_key(".".join(
                        [param.name, 'tmp'])),
3031 3032
                    dtype=param.dtype,
                    persistable=False,
W
wanghaoshuang 已提交
3033
                    stop_gradient=True)
3034
                self.params_grads.append((param, grad))
3035

3036
        for param, grad in self.params_grads:
3037 3038
            if grad is None:
                continue
X
Xin Pan 已提交
3039 3040
            with param.block.program._optimized_guard(
                [param, grad]), name_scope('move_average'):
3041
                self._append_average_accumulate_op(param)
3042

3043 3044 3045 3046
        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:
3047
                self._add_average_apply_op(block, param_grad)
3048 3049 3050 3051 3052

        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:
3053
                self._add_average_restore_op(block, param_grad)
3054

3055
    def _add_average_apply_op(self, block, param_grad):
L
Luo Tao 已提交
3056 3057 3058 3059 3060 3061
        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(
3062
            self._get_accumulator('num_accumulates', param))
L
Luo Tao 已提交
3063
        old_num_accumulates = block._clone_variable(
3064
            self._get_accumulator('old_num_accumulates', param))
L
Luo Tao 已提交
3065
        num_updates = block._clone_variable(
3066 3067 3068 3069 3070 3071
            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
3072 3073 3074 3075
        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
3076
        ops._elementwise_div(x=sum, y=tmp, out=param)
3077 3078

    def _add_average_restore_op(self, block, param_grad):
L
Luo Tao 已提交
3079 3080
        param = block._clone_variable(param_grad[0])
        grad = block._clone_variable(param_grad[1])
3081 3082 3083 3084 3085 3086 3087 3088 3089 3090 3091 3092 3093 3094 3095 3096 3097 3098 3099 3100 3101 3102 3103 3104 3105 3106 3107 3108 3109 3110 3111 3112 3113 3114 3115 3116 3117
        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
3118 3119
            },
            stop_gradient=True)
3120

S
rename  
sneaxiy 已提交
3121
    @signature_safe_contextmanager
3122
    def apply(self, executor, need_restore=True):
3123 3124
        """
        Apply the average of the cumulative ``Parameter`` to the parameters of the current model.
3125 3126

        Args:
3127 3128 3129 3130 3131 3132 3133 3134 3135 3136 3137 3138 3139 3140 3141 3142 3143 3144 3145 3146 3147 3148 3149 3150 3151 3152 3153 3154 3155 3156 3157 3158 3159 3160 3161 3162 3163 3164 3165 3166 3167 3168 3169 3170
            executor(fluid.Executor): The current network executor.
            need_restore(bool): Restore flag variable, if set to True, the network will restore
                the parameters of the network to the default value, if set to False,
                it will not be restored. The default value is True.

        Examples:

          .. code-block:: python

            import paddle.fluid as fluid
            import numpy

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

            train_program = fluid.Program()
            startup_program = fluid.Program()
            with fluid.program_guard(train_program, startup_program):
                # build net
                data = fluid.data(name='X', shape=[None, 1], dtype='float32')
                hidden = fluid.layers.fc(input=data, size=10)
                loss = fluid.layers.mean(hidden)
                optimizer = fluid.optimizer.Momentum(learning_rate=0.2, momentum=0.1)
                optimizer.minimize(loss)

                # build ModelAverage optimizer
                model_average = fluid.optimizer.ModelAverage(0.15,
                                                            min_average_window=10000,
                                                            max_average_window=12500)

                exe.run(startup_program)
                for i in range(12500):
                    x = numpy.random.random(size=(10, 1)).astype('float32')
                    outs = exe.run(program=train_program,
                                feed={'X': x},
                                fetch_list=[loss.name])

                # apply ModelAverage
                with model_average.apply(exe):
                    x = numpy.random.random(size=(10, 1)).astype('float32')
                    exe.run(program=train_program,
                            feed={'X': x},
                            fetch_list=[loss.name])
3171
        """
3172 3173 3174 3175 3176 3177
        executor.run(self.apply_program)
        try:
            yield
        finally:
            if need_restore:
                self.restore(executor)
3178 3179

    def restore(self, executor):
3180 3181
        """
        Restore ``Parameter`` values of current model.
3182 3183
        
        Args:
3184 3185 3186 3187 3188 3189 3190 3191 3192 3193 3194 3195 3196 3197 3198 3199 3200 3201 3202 3203 3204 3205 3206 3207 3208 3209 3210 3211 3212 3213 3214 3215 3216 3217 3218 3219 3220 3221 3222 3223 3224 3225 3226 3227
            executor(fluid.Executor): The current network executor.

        Examples:

          .. code-block:: python

            import paddle.fluid as fluid
            import numpy

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

            train_program = fluid.Program()
            startup_program = fluid.Program()
            with fluid.program_guard(train_program, startup_program):
                # build net
                data = fluid.data(name='X', shape=[None, 1], dtype='float32')
                hidden = fluid.layers.fc(input=data, size=10)
                loss = fluid.layers.mean(hidden)
                optimizer = fluid.optimizer.Momentum(learning_rate=0.2, momentum=0.1)
                optimizer.minimize(loss)

                # build ModelAverage optimizer
                model_average = fluid.optimizer.ModelAverage(0.15,
                                                            min_average_window=10000,
                                                            max_average_window=12500)

                exe.run(startup_program)
                for i in range(12500):
                    x = numpy.random.random(size=(10, 1)).astype('float32')
                    outs = exe.run(program=train_program,
                                feed={'X': x},
                                fetch_list=[loss.name])

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

                # restore Parameters
                model_average.restore(exe)
3228
        """
3229
        executor.run(self.restore_program)
3230 3231 3232 3233 3234 3235 3236 3237 3238 3239


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

    ..  math::

3240
        \\text{EMA}_0 & = 0
3241

3242 3243
	\\text{EMA}_t & = \\text{decay} * \\text{EMA}_{t-1} + (1 - \\text{decay}) * \\theta_t

Y
Yibing Liu 已提交
3244 3245 3246 3247
    The average results calculated by **update()** method will be saved in 
    temporary variables which are created and maintained by the object, and can 
    be applied to parameters of current model by calling **apply()** method. And 
    the **restore()** method is used to restore the parameters.
3248 3249 3250 3251 3252 3253 3254 3255 3256 3257 3258 3259 3260 3261 3262 3263 3264 3265 3266 3267 3268

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

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

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

    Usually **thres_steps** can be the global training steps.
3269 3270 3271


    Args:
Y
Yibing Liu 已提交
3272 3273 3274 3275 3276 3277 3278
	decay (float, optional): The exponential decay rate, usually close to 1, such as 
            0.999, 0.9999, ... . Default 0.999.
        thres_steps (Variable|None): If not `None`, schedule the decay rate. 
            Default None.
        name (str|None): For detailed information, please refer to 
            :ref:`api_guide_Name`. Usually name is no need to set and None by 
            default.
3279 3280 3281 3282 3283


    Examples:

	.. code-block:: python
3284 3285 3286 3287 3288

	    import numpy
	    import paddle
	    import paddle.fluid as fluid

Y
Yibing Liu 已提交
3289
	    data = fluid.data(name='x', shape=[-1, 5], dtype='float32')
3290 3291 3292 3293 3294 3295 3296 3297
	    hidden = fluid.layers.fc(input=data, size=10)
	    cost = fluid.layers.mean(hidden)

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

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

3298
	    global_steps = fluid.layers.autoincreased_step_counter()
3299 3300 3301 3302 3303 3304 3305 3306 3307 3308 3309 3310 3311 3312 3313 3314 3315 3316 3317 3318 3319 3320 3321 3322 3323 3324 3325 3326 3327
	    ema = fluid.optimizer.ExponentialMovingAverage(0.999, thres_steps=global_steps)
	    ema.update()

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

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

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

		 # usage 2
		with ema.apply(exe, need_restore=False):
		    data = numpy.random.random(size=(10, 5)).astype('float32')
		    exe.run(program=test_program,
			    feed={'x': data}, 
			    fetch_list=[hidden.name])
		ema.restore(exe)
3328 3329
    """

3330
    def __init__(self, decay=0.999, thres_steps=None, name=None):
Z
zhongpu 已提交
3331 3332 3333
        if framework.in_dygraph_mode():
            raise Exception(
                "In dygraph, don't support ExponentialMovingAverage.")
3334
        self._decay = decay
3335
        self._thres_steps = thres_steps
3336
        self._name = name if name is not None else ''
3337 3338
        self._decay_var = self._get_ema_decay()

3339
        self._step_counter_name = "@EMA_STEP_COUNTER@"
Y
Yibing Liu 已提交
3340
        self._params_tmps = []
3341
        for param in default_main_program().global_block().all_parameters():
3342 3343 3344 3345 3346 3347 3348
            if param.do_model_average != False:
                tmp = param.block.create_var(
                    name=unique_name.generate(".".join(
                        [self._name + param.name, 'ema_tmp'])),
                    dtype=param.dtype,
                    persistable=False,
                    stop_gradient=True)
Y
Yibing Liu 已提交
3349
                self._params_tmps.append((param, tmp))
3350

Y
Yibing Liu 已提交
3351 3352
        self._ema_vars = {}
        for param, tmp in self._params_tmps:
3353 3354
            with param.block.program._optimized_guard(
                [param, tmp]), name_scope('moving_average'):
Y
Yibing Liu 已提交
3355
                self._ema_vars[param.name] = self._create_ema_vars(param)
3356 3357 3358 3359

        self.apply_program = Program()
        block = self.apply_program.global_block()
        with program_guard(main_program=self.apply_program):
3360
            decay_pow, global_step = self._get_decay_pow(block)
Y
Yibing Liu 已提交
3361
            for param, tmp in self._params_tmps:
3362 3363
                param = block._clone_variable(param)
                tmp = block._clone_variable(tmp)
Y
Yibing Liu 已提交
3364
                ema = block._clone_variable(self._ema_vars[param.name])
3365
                layers.assign(input=param, output=tmp)
3366
                # bias correction
3367 3368 3369
                with layers.control_flow.Switch() as switch:
                    with switch.case(global_step > 0):
                        layers.assign(output=ema, input=ema / (1.0 - decay_pow))
3370 3371 3372 3373 3374
                layers.assign(input=ema, output=param)

        self.restore_program = Program()
        block = self.restore_program.global_block()
        with program_guard(main_program=self.restore_program):
Y
Yibing Liu 已提交
3375
            for param, tmp in self._params_tmps:
3376 3377 3378 3379
                tmp = block._clone_variable(tmp)
                param = block._clone_variable(param)
                layers.assign(input=tmp, output=param)

3380 3381 3382 3383 3384 3385 3386 3387 3388 3389 3390 3391 3392 3393 3394 3395 3396 3397 3398 3399 3400 3401
    def _get_ema_decay(self):
        with default_main_program()._lr_schedule_guard():
            decay_var = layers.tensor.create_global_var(
                shape=[1],
                value=self._decay,
                dtype='float32',
                persistable=True,
                name="scheduled_ema_decay_rate")

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

    def _get_decay_pow(self, block):
3402 3403 3404 3405 3406 3407 3408
        global_step = layers.create_global_var(
            name=self._step_counter_name,
            shape=[1],
            value=0,
            dtype='int64',
            persistable=True)
        global_step = layers.cast(global_step, "float32")
3409
        decay_var = block._clone_variable(self._decay_var)
3410 3411
        decay_pow_acc = layers.elementwise_pow(decay_var, global_step)
        return decay_pow_acc, global_step
3412

Y
Yibing Liu 已提交
3413
    def _create_ema_vars(self, param):
3414 3415 3416 3417 3418 3419 3420 3421 3422
        param_ema = layers.create_global_var(
            name=unique_name.generate(self._name + param.name + '_ema'),
            shape=param.shape,
            value=0.0,
            dtype=param.dtype,
            persistable=True)

        return param_ema

Y
Yibing Liu 已提交
3423 3424 3425 3426 3427
    def update(self):
        """ 
        Update Exponential Moving Average. Should only call this method in 
        train program.
        """
3428 3429
        global_step = layers.autoincreased_step_counter(
            counter_name=self._step_counter_name)
3430
        param_master_emas = []
Y
Yibing Liu 已提交
3431 3432 3433 3434
        for param, tmp in self._params_tmps:
            with param.block.program._optimized_guard(
                [param, tmp]), name_scope('moving_average'):
                param_ema = self._ema_vars[param.name]
3435
                if param.name + '.master' in self._ema_vars:
3436 3437 3438 3439 3440 3441 3442 3443 3444 3445 3446 3447 3448 3449 3450 3451 3452
                    master_ema = self._ema_vars[param.name + '.master']
                    param_master_emas.append([param_ema, master_ema])
                else:
                    ema_t = param_ema * self._decay_var + param * (
                        1 - self._decay_var)
                    layers.assign(input=ema_t, output=param_ema)

        # for fp16 params
        for param_ema, master_ema in param_master_emas:
            default_main_program().global_block().append_op(
                type="cast",
                inputs={"X": master_ema},
                outputs={"Out": param_ema},
                attrs={
                    "in_dtype": master_ema.dtype,
                    "out_dtype": param_ema.dtype
                })
Y
Yibing Liu 已提交
3453

3454 3455 3456 3457 3458 3459 3460
    @signature_safe_contextmanager
    def apply(self, executor, need_restore=True):
        """
        Apply moving average to parameters for evaluation.
        
        Args:
            executor (Executor): The Executor to execute applying.
Y
Yibing Liu 已提交
3461 3462
            need_restore (bool, optional): Whether to restore parameters after 
                applying. Default True.
3463 3464 3465 3466 3467 3468 3469 3470 3471 3472 3473 3474 3475 3476 3477
        """
        executor.run(self.apply_program)
        try:
            yield
        finally:
            if need_restore:
                self.restore(executor)

    def restore(self, executor):
        """Restore parameters.
        
        Args:
            executor (Executor): The Executor to execute restoring.
        """
        executor.run(self.restore_program)
H
hutuxian 已提交
3478 3479 3480


class PipelineOptimizer(object):
3481 3482
    """
    Pipeline Optimizer
H
hutuxian 已提交
3483

T
tianshuo78520a 已提交
3484
    Train with pipeline mode. The program will be split by cut_list. 
H
hutuxian 已提交
3485 3486

    If the len of cut_list is k, then the whole program (including \
T
tianshuo78520a 已提交
3487
    backward part) will be split to 2*k-1 sections. 
H
hutuxian 已提交
3488 3489 3490 3491
    
    So the length of place_list and concurrency_list must be also 2*k-1.

    Note: Though the asynchronous mode is applied in pipeline training to speed up, \
3492
    the final performance depends on the training progress of each pipeline heavily.
H
hutuxian 已提交
3493 3494 3495

    And we will try the synchronous mode in the future.

3496
    Args:
H
hutuxian 已提交
3497 3498 3499 3500
        optimizer (Optimizer): The based optimizer, such as SGD.
        cut_list (list of Variable list): The cut variable of the main_program.
        place_list (list of Place): The place where the section will run on.
        concurrency_list (list of int): The concurrency degree.
3501 3502
        queue_size (int): Each section will consume scopes from its in-scope queue 
                        and produce scopes to out-scope queue. And this parameter 
H
hutuxian 已提交
3503 3504 3505 3506
                        specify the scope queue size. [Optional. Default: 30].
        sync_steps (int): The synchronization steps between different cards. [Optional. Default: 1].
        start_cpu_core_id (int): specify the first cpu core id. [Optional. Default:0].

3507 3508
    Examples:
        .. code-block:: python
H
hutuxian 已提交
3509

3510
            import paddle.fluid as fluid
H
hutuxian 已提交
3511 3512 3513 3514 3515 3516 3517 3518 3519 3520 3521 3522 3523 3524 3525 3526 3527 3528 3529 3530 3531 3532 3533 3534 3535 3536 3537 3538 3539 3540 3541 3542 3543 3544
            import paddle.fluid.layers as layers

            x = fluid.layers.data(name='x', shape=[1], dtype='int64', lod_level=0)
            y = fluid.layers.data(name='y', shape=[1], dtype='int64', lod_level=0)
            emb_x = layers.embedding(input=x, param_attr=fluid.ParamAttr(name="embx"), size=[10,2], is_sparse=False)
            emb_y = layers.embedding(input=y, param_attr=fluid.ParamAttr(name="emby",learning_rate=0.9), size=[10,2], is_sparse=False)
            concat = layers.concat([emb_x, emb_y], axis=1)
            fc = layers.fc(input=concat, name="fc", size=1, num_flatten_dims=1, bias_attr=False)
            loss = layers.reduce_mean(fc)
            optimizer = fluid.optimizer.SGD(learning_rate=0.5)
            optimizer = fluid.optimizer.PipelineOptimizer(optimizer,
                    cut_list=[[emb_x, emb_y], [loss]],
                    place_list=[fluid.CPUPlace(), fluid.CUDAPlace(0), fluid.CPUPlace()],
                    concurrency_list=[1, 1, 4],
                    queue_size=2,
                    sync_steps=1,
                    )
            optimizer.minimize(loss)
            place = fluid.CPUPlace()
            exe = fluid.Executor(place)
            exe.run(fluid.default_startup_program())
            filelist = [] # you should set your own filelist, e.g. filelist = ["dataA.txt"]
            dataset = fluid.DatasetFactory().create_dataset("FileInstantDataset")
            dataset.set_use_var([x,y])
            dataset.set_batch_size(batch_size)
            dataset.set_filelist(filelist)
            exe.train_from_dataset(
                        fluid.default_main_program(),
                        dataset,
                        thread=2,
                        debug=False,
                        fetch_list=[],
                        fetch_info=[],
                        print_period=1)
3545 3546
    """

H
hutuxian 已提交
3547 3548 3549 3550 3551 3552 3553 3554
    def __init__(self,
                 optimizer,
                 cut_list=None,
                 place_list=None,
                 concurrency_list=None,
                 queue_size=30,
                 sync_steps=1,
                 start_cpu_core_id=0):
Z
zhongpu 已提交
3555 3556
        if framework.in_dygraph_mode():
            raise Exception("In dygraph, don't support PipelineOptimizer.")
H
hutuxian 已提交
3557 3558 3559 3560 3561 3562 3563 3564 3565
        # TODO: check properties
        self._optimizer = optimizer
        self._cut_list = cut_list
        self._place_list = place_list
        self._concurrency_list = concurrency_list
        self._queue_size = queue_size
        self._sync_steps = sync_steps
        self._start_cpu_core_id = start_cpu_core_id

H
hutuxian 已提交
3566
    def _create_vars(self, block, main_program):
H
hutuxian 已提交
3567 3568 3569 3570 3571 3572 3573 3574 3575 3576 3577
        used_var_set = set()
        for op_idx in range(block.desc.op_size()):
            op_desc = block.desc.op(op_idx)
            vars = op_desc.input_arg_names() + op_desc.output_arg_names()
            for var in vars:
                if var in used_var_set:
                    continue
                used_var_set.add(var)
                source_var = main_program.block(0).var(str(var))
                block._clone_variable(source_var, False)

H
hutuxian 已提交
3578
    def _extract_section_opt_ops(self, ops, cut_point_name):
H
hutuxian 已提交
3579 3580 3581 3582 3583 3584 3585 3586 3587 3588 3589 3590 3591 3592 3593
        """
        Extract opt ops in the given section
        """
        output_names = set(cut_point_name)
        relevant_op_flags = [True] * len(ops)
        for i, op in reversed(list(enumerate(ops))):
            if _some_in_set_(op.desc.output_arg_names(), output_names):
                for name in op.desc.input_arg_names():
                    output_names.add(name)
            else:
                relevant_op_flags[i] = False

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

H
hutuxian 已提交
3594
    def _find_input_output(self, ops, name, is_forward=True):
H
hutuxian 已提交
3595 3596 3597 3598 3599 3600 3601 3602 3603 3604 3605 3606 3607 3608
        """
        Find the inputs or outputs of a section
        """
        all_set = set()
        part_set = set()
        for op in ops:
            if is_forward:
                part_set.update(op.desc.output_arg_names())
            else:
                part_set.update(op.desc.input_arg_names())
            all_set.update(op.desc.output_arg_names())
            all_set.update(op.desc.input_arg_names())
        return all_set - part_set

H
hutuxian 已提交
3609
    def _find_persistable_vars(self, ops, whole_parameters):
H
hutuxian 已提交
3610 3611 3612 3613 3614 3615 3616 3617 3618 3619 3620 3621 3622 3623 3624 3625 3626 3627 3628 3629 3630 3631 3632 3633 3634 3635 3636
        """
        find the persistable input vars in current section
        """
        res = set()
        for op in ops:
            vars = op.desc.input_arg_names()
            for var in vars:
                if var in whole_parameters:
                    res.add(var)
        return res

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

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

H
hutuxian 已提交
3637
    def _extract_section_ops(self, ops, cut_point_name):
H
hutuxian 已提交
3638 3639 3640 3641 3642 3643 3644 3645 3646 3647 3648 3649 3650 3651 3652 3653 3654 3655 3656
        """
        Extract ops in the given section 
        """
        output_names = set(cut_point_name)
        relevant_op_flags = [True] * len(ops)
        for i, op in reversed(list(enumerate(ops))):
            if not self._is_opt_role_op(op) and _some_in_set_(
                    op.desc.output_arg_names(), output_names):
                for name in op.desc.input_arg_names():
                    output_names.add(name)
            elif op.desc.type() == "print" and op.desc.input_arg_names()[
                    0] in output_names:
                continue
            else:
                relevant_op_flags[i] = False

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

H
hutuxian 已提交
3657 3658
    def _find_section_opt(self, ops, params):
        res = self._extract_section_opt_ops(ops, params)
H
hutuxian 已提交
3659 3660
        return res

H
hutuxian 已提交
3661
    def _split_program(self, main_program, cut_list):
H
hutuxian 已提交
3662 3663 3664 3665 3666 3667 3668 3669 3670 3671 3672 3673 3674 3675 3676 3677 3678 3679 3680 3681
        programs = []
        block = main_program.block(0)
        whole_parameters = [e.name for e in block.all_parameters()]
        cut_var_names = []
        cut_len = len(cut_list)
        sec_params = []
        for i, cut_vars in enumerate(cut_list[:-1]):
            cut_var_names.append([cut_var.name for cut_var in cut_vars])
        for i, cut_vars in reversed(list(enumerate(cut_list[:-1]))):
            cut_var_names.append(
                [_append_grad_suffix_(cut_var.name) for cut_var in cut_vars])
            if i == 0:
                cut_var_names[-1] += [var.name for var in cut_list[-1]]
        ops = block.ops[:]
        for i, cut_vars in enumerate(cut_var_names):
            program = {
                "program": Program(),
                "input_set": set(),
                "output_set": set()
            }
H
hutuxian 已提交
3682
            cur_ops = self._extract_section_ops(ops, cut_vars)
H
hutuxian 已提交
3683 3684 3685 3686 3687 3688
            if i == 0:
                for op in ops:
                    if self._is_lr_role_op(op):
                        cur_ops.append(op)
            #prevent inplace in/out
            program["input_set"].update(
H
hutuxian 已提交
3689
                self._find_input_output(
H
hutuxian 已提交
3690 3691 3692 3693 3694 3695
                    cur_ops, [], is_forward=True))
            for e in cur_ops:
                ops.remove(e)

            if i < cut_len:
                sec_params.append(
H
hutuxian 已提交
3696
                    self._find_persistable_vars(cur_ops, whole_parameters))
H
hutuxian 已提交
3697
            if i >= cut_len - 1:
H
hutuxian 已提交
3698 3699
                opt_ops = self._find_section_opt(
                    ops, sec_params[2 * cut_len - 2 - i])
H
hutuxian 已提交
3700 3701 3702 3703 3704 3705 3706 3707 3708 3709

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

            op_descs = [op.desc for op in cur_ops]
            for op_desc in op_descs:
                ap_op = program["program"].block(0).desc.append_op()
                ap_op.copy_from(op_desc)
            program["input_set"].update(
H
hutuxian 已提交
3710
                self._find_input_output(
H
hutuxian 已提交
3711 3712 3713
                    cur_ops, cut_vars, is_forward=True))
            program["input_set"].update(sec_params[min(i, 2 * cut_len - 2 - i)])
            program["output_set"].update(
H
hutuxian 已提交
3714
                self._find_input_output(
H
hutuxian 已提交
3715 3716 3717 3718 3719 3720 3721 3722 3723 3724 3725 3726 3727 3728
                    cur_ops, cut_vars, is_forward=False))
            programs.append(program)
        program = {
            "program": Program(),
            "input_set": set(),
            "output_set": set()
        }
        op_descs = [op.desc for op in ops]
        for op_desc in op_descs:
            ap_op = program["program"].block(0).desc.append_op()
            ap_op.copy_from(op_desc)
        program["input_set"].update(
            [cut_var.name + "@GRAD" for cut_var in cut_list[0]])
        program["input_set"].update(
H
hutuxian 已提交
3729
            self._find_input_output(
H
hutuxian 已提交
3730 3731 3732 3733 3734 3735 3736 3737 3738 3739 3740 3741 3742 3743 3744 3745 3746 3747 3748 3749
                ops, [], is_forward=True))
        program["input_set"].update(sec_params[0])
        programs.append(program)
        inputs = set()
        for program in reversed(list(programs)):
            output_list = list(program["output_set"])
            for output in output_list:
                if output not in inputs:
                    program["output_set"].remove(output)
            inputs.update(program["input_set"])
        return programs

    def minimize(self,
                 loss,
                 startup_program=None,
                 parameter_list=None,
                 no_grad_set=None):
        self._optimizer.minimize(loss, startup_program, parameter_list,
                                 no_grad_set)
        program = loss.block.program
H
hutuxian 已提交
3750 3751 3752 3753 3754 3755 3756 3757
        if len(self._cut_list) == 0:
            program_list = []
            ptmp = {"program": program, "input_set": set(), "output_set": set()}
            program_list.append(ptmp)
        else:
            program_list = self._split_program(program, self._cut_list)
            for p in program_list:
                self._create_vars(p["program"].block(0), program)
H
hutuxian 已提交
3758 3759 3760 3761 3762 3763 3764 3765 3766 3767 3768 3769 3770 3771 3772 3773 3774 3775 3776 3777
        whole_parameters = [e.name for e in program.block(0).all_parameters()]
        param_need_sync = []
        for i, section_p in enumerate(program_list):
            if not isinstance(self._place_list[i], core.CUDAPlace):
                continue
            section_var = [e for e in section_p["program"].block(0).vars]
            for p in section_var:
                if p in whole_parameters:
                    param_need_sync.append(p)
        program._pipeline_opt = {
            "trainer": "PipelineTrainer",
            "device_worker": "Section",
            "section_program_list": program_list,
            "place_list": self._place_list,
            "concurrency_list": self._concurrency_list,
            "queue_size": self._queue_size,
            "start_cpu_core_id": self._start_cpu_core_id,
            "sync_steps": self._sync_steps,
            "param_need_sync": param_need_sync
        }
M
mapingshuo 已提交
3778 3779


M
mapingshuo 已提交
3780 3781 3782 3783 3784 3785 3786 3787 3788 3789 3790 3791 3792 3793 3794 3795 3796 3797 3798 3799 3800 3801 3802 3803 3804 3805 3806 3807 3808 3809 3810 3811 3812 3813 3814 3815 3816 3817 3818 3819 3820 3821 3822 3823 3824 3825 3826 3827 3828 3829 3830 3831 3832 3833 3834 3835 3836 3837 3838 3839 3840 3841
class RecomputeOptimizer(Optimizer):
    """
    Recompute Optimizer Wrapper

    Normally, a training step contains three sub-steps: first, run forward
    Operators to calculate the loss; second, run backward Operators to 
    calculate gradient of the parameters; third, apply optimization method
    to update the value of the parameters.

    In the forward computation process, all variables that are needed by 
    backward computation process will be kept in memory, which occupy a great
    amount of memory when the network becomes very deep.

    Recompute split the network to k segments. In each segment, It will 
    recompute the forward Operators, before running backward operators. It is
    very helpful for saving memory.
 
    The Variables that separate a network to segments are called as checkpoints,
    and users should set it manually. The usage is very simple:

    Args:
        optimizer (Optimizer): The optimizer that is applied to parameters.

    Examples:
        .. code-block:: python

            import paddle.fluid as fluid
            import numpy as np
            def gen_data():
                return {"x": np.random.random(size=(32, 32)).astype('float32'),
                "y": np.random.randint(2, size=(32, 1)).astype('int64')}
            def mlp(input_x, input_y, hid_dim=128, label_dim=2):
                print(input_x)
                fc_1 = fluid.layers.fc(input=input_x, size=hid_dim)
                prediction = fluid.layers.fc(input=[fc_1], size=label_dim, act='softmax')
                cost = fluid.layers.cross_entropy(input=prediction, label=input_y)
                sum_cost = fluid.layers.reduce_mean(cost)
                return sum_cost, fc_1, prediction
            input_x = fluid.layers.data(name="x", shape=[32], dtype='float32')
            input_y = fluid.layers.data(name="y", shape=[1], dtype='int64')
            cost, fc_1, pred = mlp(input_x, input_y)

            sgd = fluid.optimizer.Adam(learning_rate=0.01)
            sgd = fluid.optimizer.RecomputeOptimizer(sgd)
            sgd._set_checkpoints([fc_1, pred])
            sgd.minimize(cost)

            print("Finished optimize")
            place = fluid.CPUPlace()
            exe = fluid.Executor(place)
            exe.run(fluid.default_startup_program())
            step = 10

            for i in range(step):
                cost_val = exe.run(feed=gen_data(),
                       program=fluid.default_main_program(),
                       fetch_list=[cost.name])
                print("step=%d cost=%f" % (i, cost_val[0]))

    """

    def __init__(self, optimizer):
Z
zhongpu 已提交
3842 3843
        if framework.in_dygraph_mode():
            raise Exception("In dygraph, don't support RecomputeOptimizer.")
M
mapingshuo 已提交
3844 3845 3846 3847 3848 3849 3850 3851 3852 3853 3854 3855 3856 3857 3858 3859 3860 3861 3862 3863 3864 3865 3866 3867 3868 3869 3870 3871 3872 3873 3874 3875 3876 3877 3878 3879 3880 3881 3882 3883 3884 3885 3886 3887 3888 3889 3890 3891 3892 3893 3894 3895 3896 3897 3898 3899 3900 3901 3902 3903 3904 3905 3906 3907 3908 3909 3910 3911 3912 3913 3914 3915 3916 3917 3918
        self._optimizer = optimizer
        self._checkpoints = None

    def _set_checkpoints(self, checkpoints):
        self._checkpoints = checkpoints

    def load(self, stat_dict):
        """
        load function is not supported by Recompute Optimizer for now.
        :return: None

        Args:
            stat_dict: the dict load by load_persistable method

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
                import paddle.compat as cpt
                
                def mlp(input_x, input_y, hid_dim=128, label_dim=2):
                    fc_1 = fluid.layers.fc(input=input_x, size=hid_dim)
                    prediction = fluid.layers.fc(input=[fc_1], size=label_dim, act='softmax')
                    cost = fluid.layers.cross_entropy(input=prediction, label=input_y)
                    sum_cost = fluid.layers.reduce_mean(cost)
                    return sum_cost, fc_1, prediction
                
                input_x = fluid.layers.data(name="x", shape=[32], dtype='float32')
                input_y = fluid.layers.data(name="y", shape=[1], dtype='int64')
                cost, fc_1, pred = mlp(input_x, input_y)
                print("Finished FF")
                
                sgd = fluid.optimizer.Adam(learning_rate=0.01)
                sgd = fluid.optimizer.RecomputeOptimizer(sgd)
                sgd._set_checkpoints([fc_1, pred])
                try:
                    stat_dict = {}
                    sgd.load(stat_dict)
                except NotImplementedError as e:
                    print(cpt.get_exception_message(e))
        """
        raise NotImplementedError(
            "load function is not supported by Recompute Optimizer for now")

    def apply_gradients(self, params_grads):
        """
        call apply_gradients function of self._optimizer.

        Args:
            params_grads (list): list of (param, grad) pair to do optimization.

        Returns:
            list: A list of operators appended to the current program.

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
                import paddle.fluid.framework as framework

                def mlp(input_x, input_y, hid_dim=128, label_dim=2):
                    fc_1 = fluid.layers.fc(input=input_x, size=hid_dim)
                    prediction = fluid.layers.fc(input=[fc_1], size=label_dim, act='softmax')
                    cost = fluid.layers.cross_entropy(input=prediction, label=input_y)
                    sum_cost = fluid.layers.reduce_mean(cost)
                    return sum_cost, fc_1, prediction


                input_x = fluid.layers.data(name="x", shape=[32], dtype='float32')
                input_y = fluid.layers.data(name="y", shape=[1], dtype='int64')
                cost, fc_1, pred = mlp(input_x, input_y)
                print("Finished FF")

                sgd = fluid.optimizer.Adam(learning_rate=0.01)
                sgd = fluid.optimizer.RecomputeOptimizer(sgd)
3919
                sgd._set_checkpoints([fc_1, pred])
M
mapingshuo 已提交
3920 3921 3922 3923
                params_grads = sgd.backward(
                    cost,
                    startup_program=None,
                    parameter_list=None,
3924
                    no_grad_set=None)
M
mapingshuo 已提交
3925 3926 3927 3928 3929 3930 3931 3932 3933 3934 3935 3936 3937 3938 3939

                program = cost.block.program
                with framework.program_guard(program, None):
                    optimize_ops = sgd.apply_gradients(params_grads)

                print("Finished apply gradients")
        """

        return self._optimizer.apply_gradients(params_grads=params_grads)

    def backward(self,
                 loss,
                 startup_program=None,
                 parameter_list=None,
                 no_grad_set=None,
3940
                 callbacks=None):
M
mapingshuo 已提交
3941 3942 3943 3944 3945 3946 3947
        """
        call append_backward with checkpoints.

        Args:
            loss (Variable): loss variable to run optimizations.
            startup_program (Program): startup_program for initializing parameters
                in `parameter_list`.
3948 3949
            parameter_list (list): list of Variables or Variable.names to update.
            no_grad_set (set|None): set of Variables or Variables.names should be ignored.
M
mapingshuo 已提交
3950 3951 3952 3953 3954 3955 3956 3957 3958 3959 3960 3961 3962 3963 3964 3965 3966 3967 3968 3969 3970 3971 3972 3973
            callbacks (list|None): list of callables to run when appending backward
                operator for one parameter.
            checkpoints (list): list of Variables as checkpoints

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
    
                def mlp(input_x, input_y, hid_dim=128, label_dim=2):
                    fc_1 = fluid.layers.fc(input=input_x, size=hid_dim)
                    prediction = fluid.layers.fc(input=[fc_1], size=label_dim, act='softmax')
                    cost = fluid.layers.cross_entropy(input=prediction, label=input_y)
                    sum_cost = fluid.layers.reduce_mean(cost)
                    return sum_cost, fc_1, prediction
    
    
                input_x = fluid.layers.data(name="x", shape=[32], dtype='float32')
                input_y = fluid.layers.data(name="y", shape=[1], dtype='int64')
                cost, fc_1, pred = mlp(input_x, input_y)
                print("Finished FF")
    
                sgd = fluid.optimizer.Adam(learning_rate=0.01)
                sgd = fluid.optimizer.RecomputeOptimizer(sgd)
3974
                sgd._set_checkpoints([fc_1, pred])
M
mapingshuo 已提交
3975 3976 3977 3978
                params_grads = sgd.backward(
                    cost,
                    startup_program=None,
                    parameter_list=None,
3979
                    no_grad_set=None)
M
mapingshuo 已提交
3980 3981 3982 3983 3984 3985 3986 3987 3988 3989 3990 3991 3992 3993 3994
                print("Finished backward")
        """

        if framework.in_dygraph_mode():
            raise NotImplementedError(
                "DyGraph current does not support recompute")

        self._dtype = loss.dtype
        program = loss.block.program
        with program_guard(program, startup_program):
            params_grads = append_backward(
                loss,
                parameter_list,
                no_grad_set,
                checkpoints=self._checkpoints)
3995 3996 3997
            # Note: since we can't use all_reduce_op now,
            #  dgc_op should be the last op of one grad.
            self._optimizer._append_dgc_ops(params_grads)
M
mapingshuo 已提交
3998 3999 4000 4001 4002 4003 4004 4005 4006 4007 4008 4009 4010 4011 4012 4013 4014 4015 4016
        return params_grads

    def apply_optimize(self, loss, startup_program, params_grads):
        """
        call the apply_optimize function of self._optimizer
        Args:
            loss (Variable): loss variable to run optimizations.
            startup_program (Program): startup_program for initializing parameters
                in `parameter_list`.
            params_grads (list): list of (param, grad) pair to do optimization.
        Examples:
            .. code-block:: python
                import paddle.fluid as fluid
                
                def mlp(input_x, input_y, hid_dim=128, label_dim=2):
                    fc_1 = fluid.layers.fc(input=input_x, size=hid_dim)
                    prediction = fluid.layers.fc(input=[fc_1], size=label_dim, act='softmax')
                    cost = fluid.layers.cross_entropy(input=prediction, label=input_y)
                    sum_cost = fluid.layers.reduce_mean(cost)
M
mapingshuo 已提交
4017
                    return sum_cost, fc_1, prediction                
M
mapingshuo 已提交
4018 4019 4020 4021 4022 4023 4024 4025
                
                input_x = fluid.layers.data(name="x", shape=[32], dtype='float32')
                input_y = fluid.layers.data(name="y", shape=[1], dtype='int64')
                cost, fc_1, pred = mlp(input_x, input_y)
                print("Finished FF")
                
                sgd = fluid.optimizer.Adam(learning_rate=0.01)
                sgd = fluid.optimizer.RecomputeOptimizer(sgd)
4026
                sgd._set_checkpoints([fc_1, pred])
M
mapingshuo 已提交
4027 4028 4029 4030
                params_grads = sgd.backward(
                    cost,
                    startup_program=None,
                    parameter_list=None,
4031
                    no_grad_set=None)
M
mapingshuo 已提交
4032 4033 4034 4035 4036 4037 4038 4039 4040 4041 4042 4043 4044 4045 4046 4047
                
                optimize_ops = sgd.apply_optimize(
                    cost, startup_program=None, params_grads=params_grads)
                
                print("Finished apply_optimize")
        """

        return self._optimizer.apply_optimize(
            loss, startup_program=startup_program, params_grads=params_grads)

    def minimize(self,
                 loss,
                 startup_program=None,
                 parameter_list=None,
                 no_grad_set=None,
                 grad_clip=None):
4048
        assert isinstance(loss, Variable), "The loss should be an Variable."
M
mapingshuo 已提交
4049 4050 4051 4052 4053
        assert (self._checkpoints is not None
                ), "You should call _set_checkpoints first"
        if framework.in_dygraph_mode():
            raise NotImplementedError(
                "DyGraph current does not support recompute")
4054 4055 4056 4057 4058 4059
        if grad_clip is not None:
            if not isinstance(grad_clip, GradientClipBase):
                raise TypeError(
                    "'grad_clip' should be an instance of GradientClipBase's derived class"
                )
            self._optimizer._grad_clip = grad_clip
M
mapingshuo 已提交
4060 4061 4062 4063
        params_grads = self.backward(
            loss,
            startup_program=startup_program,
            parameter_list=parameter_list,
4064
            no_grad_set=no_grad_set)
M
mapingshuo 已提交
4065 4066 4067 4068 4069 4070 4071

        optimize_ops = self.apply_optimize(
            loss, startup_program=startup_program, params_grads=params_grads)

        return optimize_ops, params_grads


M
mapingshuo 已提交
4072 4073 4074 4075 4076 4077 4078 4079 4080 4081 4082 4083 4084 4085 4086 4087 4088 4089 4090 4091 4092 4093 4094 4095 4096 4097 4098 4099 4100 4101 4102 4103 4104 4105 4106 4107 4108 4109 4110 4111 4112 4113 4114 4115 4116 4117 4118 4119 4120 4121 4122 4123 4124 4125 4126
class LookaheadOptimizer(object):
    """
    This implements the Lookahead optimizer of the
    paper : https://arxiv.org/abs/1907.08610.

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

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

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

    Examples:
        .. code-block:: python

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

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

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

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

    """

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

Z
zhongpu 已提交
4127 4128
        if framework.in_dygraph_mode():
            raise Exception("In dygraph, don't support LookaheadOptimizer.")
M
mapingshuo 已提交
4129 4130 4131 4132 4133 4134 4135 4136 4137 4138 4139 4140 4141 4142 4143 4144 4145 4146 4147 4148 4149 4150 4151 4152 4153 4154 4155 4156 4157 4158 4159 4160 4161 4162 4163 4164 4165 4166 4167 4168 4169 4170 4171 4172 4173 4174 4175 4176 4177 4178 4179 4180 4181 4182 4183 4184 4185 4186 4187 4188 4189 4190 4191 4192 4193 4194 4195 4196 4197 4198 4199 4200 4201 4202 4203 4204 4205 4206 4207 4208 4209 4210 4211 4212 4213 4214 4215 4216 4217 4218 4219 4220 4221 4222 4223 4224
        assert (inner_optimizer is not None), "inner optimizer can not be None"
        assert (
            0.0 <= alpha <= 1.0
        ), "alpha should be larger or equal to 0.0, and less or equal than 1.0"
        assert (isinstance(k, int) and k > 0), "k should be a positive integer"

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

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

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

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

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

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

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

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

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

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

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

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

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