optimizer.py 177.4 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
import logging
19
from collections import defaultdict
20

Q
Qiao Longfei 已提交
21
from paddle.fluid.distribute_lookup_table import find_distributed_lookup_table
22
from paddle.fluid.framework import Program, Variable, name_scope, default_main_program, default_startup_program, device_guard
23

24 25
from . import framework
from . import layers
26
from . import unique_name
27
from .backward import append_backward, _some_in_set_, _append_grad_suffix_, _get_no_grad_set_name
28
from .clip import GradientClipBase, GradientClipByNorm, error_clip_callback, append_gradient_clip_ops
29 30 31
from .framework import program_guard
from .initializer import Constant
from .layer_helper import LayerHelper
S
sneaxiy 已提交
32
from .layers import ops
33
from .regularizer import append_regularization_ops
34
from .dygraph import base as imperative_base
35
from .dygraph import no_grad
36 37 38 39
from .dygraph.learning_rate_scheduler import LearningRateDecay
from paddle.fluid import core
from paddle.fluid.layers import tensor
from functools import reduce
40
from .wrapped_decorator import signature_safe_contextmanager
M
mapingshuo 已提交
41
from .. import compat as cpt
42

43
__all__ = [
44 45 46 47
    'SGD', 'Momentum', 'Adagrad', 'Adam', 'Adamax', 'Dpsgd', 'DecayedAdagrad',
    'Ftrl', 'SGDOptimizer', 'MomentumOptimizer', 'AdagradOptimizer',
    'AdamOptimizer', 'AdamaxOptimizer', 'DpsgdOptimizer',
    'DecayedAdagradOptimizer', 'RMSPropOptimizer', 'FtrlOptimizer', 'Adadelta',
Z
Zeng Jinle 已提交
48 49 50 51
    'AdadeltaOptimizer', 'ModelAverage', 'LarsMomentum',
    'LarsMomentumOptimizer', 'DGCMomentumOptimizer', 'LambOptimizer',
    'ExponentialMovingAverage', 'PipelineOptimizer', 'LookaheadOptimizer',
    'RecomputeOptimizer'
52
]
Q
Qiao Longfei 已提交
53 54 55 56 57 58


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

    Define the common interface of an optimizer.
59 60
    User should not use this class directly,
    but need to use one of it's implementation.
Q
Qiao Longfei 已提交
61 62
    """

63
    @imperative_base.no_grad
64 65 66 67
    def __init__(self,
                 learning_rate,
                 parameter_list=None,
                 regularization=None,
68
                 grad_clip=None,
69
                 name=None):
70
        self._parameter_list = parameter_list
71
        self._name = name
L
lujun 已提交
72
        if framework.in_dygraph_mode():
M
minqiyang 已提交
73 74 75 76 77
            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))
78
            if self._parameter_list is None:
79 80 81
                raise AttributeError(
                    "parameter_list argument given to the Optimizer should not be None in dygraph mode."
                )
82 83 84 85 86 87 88 89
            if regularization is not None:
                for param in self._parameter_list:
                    if param.regularizer is not None:
                        logging.info(
                            "If regularizer of a Parameter has been set by 'fluid.ParamAttr' or 'fluid.WeightNormParamAttr' already. "
                            "The Regularization[%s] in Optimizer will not take effect, and it will only be applied to other Parameters!"
                            % regularization.__str__())
                        break
M
minqiyang 已提交
90 91 92 93 94 95 96
        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))

97 98 99 100 101
        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"
                )
D
dzhwinter 已提交
102
        self.regularization = regularization
103
        self._grad_clip = grad_clip
104
        self._learning_rate = learning_rate
D
dzhwinter 已提交
105 106
        # the learning rate type should be inferenced from loss
        self._dtype = None
107
        # each program should have a independent learning rate
108
        # program -> Variable(learning_rate)
Q
qiaolongfei 已提交
109
        self._learning_rate_map = dict()
110
        if isinstance(self._learning_rate, framework.Variable):
111 112
            self._learning_rate_map[framework.default_main_program(
            )] = self._learning_rate
113 114 115 116 117
        # 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
118
        self.helper = None
119
        self._opti_name_list = []
H
hong 已提交
120
        self._accumulators_holder = {}
121
        self._param_device_map = dict()
H
hong 已提交
122 123 124 125

    @framework.dygraph_only
    def state_dict(self):
        '''
T
tianshuo78520a 已提交
126 127
        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
128 129 130

        Args: None
        Return:
T
tianshuo78520a 已提交
131
            state_dict(dict) : dict contains all the variable used by optimizer
H
hong 已提交
132 133 134 135 136
        
        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
137 138 139 140 141 142

                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
143 144 145 146 147 148 149 150

        '''
        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):
151
            var_tmp = None
L
Leo Chen 已提交
152
            if framework.in_dygraph_mode():
153 154
                var_temp = framework._varbase_creator(
                    None, name='global_step', dtype='int32')
L
Leo Chen 已提交
155 156
            else:
                var_temp = Variable(None, name='global_step', dtype='int32')
157

H
hong 已提交
158 159 160 161 162 163 164 165 166
            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
167
        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
168 169 170 171 172 173 174 175

        Args: 
            state_dict(dict) : Dict contains all the Variable needed by optimizer
        Return:
            None
        
        Examples:
            .. code-block:: python
176

H
hong 已提交
177
                with fluid.dygraph.guard():
178
                    emb = fluid.dygraph.Embedding([10, 10])
179

H
hong 已提交
180
                    state_dict = emb.state_dict()
181
                    fluid.save_dygraph(state_dict, "paddle_dy")
182

183 184
                    adam = fluid.optimizer.Adam(learning_rate=fluid.layers.noam_decay( 100, 10000), 
                                                parameter_list=emb.parameters())
H
hong 已提交
185
                    state_dict = adam.state_dict()
186
                    fluid.save_dygraph(state_dict, "paddle_dy")
187

H
hong 已提交
188
                    para_state_dict, opti_state_dict = fluid.load_dygraph( "paddle_dy")
189

190
                    adam.set_dict(opti_state_dict)
H
hong 已提交
191 192 193 194 195 196 197 198 199

        '''

        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):
200
                step_np = global_step
H
hong 已提交
201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224
                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 )
225
                var = var_tmp.value()
H
hong 已提交
226 227 228 229 230 231 232 233
                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):
234
                    load_para_np = load_para.numpy()
H
hong 已提交
235 236 237 238 239 240 241 242 243 244 245 246 247 248 249
                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())
250

251 252
    def get_opti_var_name_list(self):
        return self._opti_name_list
Q
Qiao Longfei 已提交
253

Q
Qiao Longfei 已提交
254
    def _create_global_learning_rate(self):
255 256 257
        if imperative_base.enabled():
            # create learning rate Variable
            if isinstance(self._learning_rate, float):
M
minqiyang 已提交
258 259 260 261 262 263 264 265 266 267 268 269
                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)
270
            # get learning rate Variable from LearningRateDecay
M
minqiyang 已提交
271
            elif isinstance(self._learning_rate, LearningRateDecay):
272 273 274
                self._learning_rate_map[framework.default_main_program(
                )] = self._learning_rate()
            else:
Q
qiaolongfei 已提交
275
                raise TypeError(
276 277
                    "optimizer's learning rate must be float or LearningRateDecay"
                )
278
        else:
279 280 281 282
            lr = self._global_learning_rate()

            if isinstance(lr, framework.Variable):
                return
M
minqiyang 已提交
283 284 285 286 287 288
            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
289

290 291 292 293 294 295 296 297
            # 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)
298

299 300 301 302
    @framework.dygraph_only
    def current_step_lr(self):
        """
        .. note::
T
tianshuo78520a 已提交
303
          **This API is ONLY available in Dygraph mode**
304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360
        
        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
361
    def _global_learning_rate(self, program=None):
Q
Qiao Longfei 已提交
362 363 364 365
        """
        get global decayed learning rate
        :return:
        """
366 367
        if program is None:
            program = framework.default_main_program()
Q
qiaolongfei 已提交
368
        return self._learning_rate_map.get(program, None)
Q
Qiao Longfei 已提交
369

Q
Qiao Longfei 已提交
370 371 372 373 374
    def _append_optimize_op(self, block, param_and_grad):
        """ append optimize operator to block and return all the added optimize_op
        """
        raise NotImplementedError()

375 376 377 378
    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
379 380
        if type(param_lr) == Variable:
            return param_lr
Q
qiaolongfei 已提交
381
        else:
W
Wu Yi 已提交
382
            if param_lr == 1.0:
Y
yuyang18 已提交
383
                return self._global_learning_rate()
W
Wu Yi 已提交
384
            else:
X
Xin Pan 已提交
385 386 387
                with default_main_program()._lr_schedule_guard(
                        is_with_opt=True), framework.name_scope(
                            'scale_with_param_lr'):
388
                    return self._global_learning_rate() * param_lr
389 390 391 392 393 394 395

    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
396
        """
397 398
        pass

399
    def _finish_update(self, block, parameters_and_grads):
400 401 402 403 404 405 406 407
        """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
408
            None
409 410 411
        """
        pass

412 413 414 415 416
    def _add_accumulator(self,
                         name,
                         param,
                         dtype=None,
                         fill_value=0.0,
417
                         shape=None,
418
                         type=None,
419
                         device=None):
420 421 422 423 424 425 426 427 428
        """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
429 430
        if self._name is not None:
            name = self._name + "_" + name
431 432
        if (name in self._accumulators and
                param.name in self._accumulators[name]):
L
lujun 已提交
433
            if framework.in_dygraph_mode():
X
polish  
Xin Pan 已提交
434
                return self._accumulators[name][param.name]
435
            raise Exception("Accumulator {} already exists for parameter {}".
436
                            format(name, param.name))
437 438
        if shape == None:
            shape = param.shape
Q
Qiao Longfei 已提交
439
        assert isinstance(self.helper, LayerHelper)
440 441 442 443 444

        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
445
        var = self.helper.create_global_variable(
446
            name=var_name,
Q
Qiao Longfei 已提交
447
            persistable=True,
F
fengjiayi 已提交
448
            dtype=dtype or param.dtype,
449
            type=param.type if type is None else type,
H
hong 已提交
450 451
            shape=shape,
            belong_to_optimizer=True)
452 453 454 455 456
        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
457 458 459 460 461 462 463

        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
464
        self._accumulators[name][param.name] = var
465
        return var
466 467 468 469 470 471 472 473 474 475 476

    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
477 478
        if self._name is not None:
            name = self._name + "_" + name
479 480 481 482 483 484
        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]

485 486 487 488 489 490 491 492 493 494 495 496
    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)
497
                        break
498 499 500 501 502 503 504

    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

505
    def _create_optimization_pass(self, parameters_and_grads):
Q
Qiao Longfei 已提交
506 507 508
        """Add optimization operators to update gradients to variables.

        Args:
Q
qiaolongfei 已提交
509
          parameters_and_grads(list(tuple(Variable, Variable))):
510
            a list of (variable, gradient) pair to update.
Q
Qiao Longfei 已提交
511 512

        Returns:
513
          return_op_list: a list of operators that will complete one step of
514 515 516
            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
517
        """
518 519 520 521 522
        # 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
523
        # for parameters and extend _finish_update method to add custom ops.
524

525
        # Allways called under program_guard use global block as loss block
526 527 528
        # But if current block is in control flow, append optimize op in the
        # grad block of current block

529
        global_block = framework.default_main_program().global_block()
530 531 532 533 534 535 536 537 538
        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)
539
        self.helper = LayerHelper(self.__class__.__name__)
540
        self._update_param_device_map(parameters_and_grads, target_block)
C
chengduo 已提交
541
        self._create_accumulators(
542
            target_block,
C
chengduo 已提交
543
            [p[0] for p in parameters_and_grads if p[0].trainable])
544 545
        self._create_global_learning_rate()

M
minqiyang 已提交
546
        if framework.in_dygraph_mode():
547 548 549
            for param_and_grad in parameters_and_grads:
                if param_and_grad[1] is None:
                    continue
550 551
                if param_and_grad[0].trainable is True:
                    self._append_optimize_op(target_block, param_and_grad)
552 553 554 555 556 557 558
        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:
559 560 561 562 563
                        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)
564 565 566

        # Get custom finish ops for subclasses
        # FIXME: Need to fix this once we figure out how to handle dependencies
567
        self._finish_update(target_block, parameters_and_grads)
568

569 570
        end = len(target_block.ops)
        return target_block._slice_ops(start, end)
571 572

    def _process_distribute_lookuptable(self, param_grads):
Q
Qiao Longfei 已提交
573 574 575 576 577 578 579 580 581
        """
        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
        """
582 583
        program = framework.default_main_program()
        global_block = framework.default_main_program().global_block()
Q
Qiao Longfei 已提交
584 585 586 587 588 589 590 591 592 593 594 595 596 597 598
        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:
599 600 601 602 603 604 605 606 607 608 609 610 611
            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
612 613
        return new_param_grads, (table_param, table_grad), sgd_op

614 615 616
    def _append_dgc_ops(self, param_and_grad):
        pass

617 618 619 620 621 622 623
    def backward(self,
                 loss,
                 startup_program=None,
                 parameter_list=None,
                 no_grad_set=None,
                 callbacks=None):
        """
624
        The first part of ``minimize``, do auto-diff to append backward operations for
625 626 627
        the current program.

        Args:
628 629 630 631
            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.
632
            parameter_list (list, optional): List of ``Variable`` or ``Variable.name`` to update
633 634
                to minimize ``loss``. The default value is None, at this time all parameters
                will be updated.
635
            no_grad_set (set, optional): Set of ``Variable``  or ``Variable.name`` that don't need
636 637 638
                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
639

640
        Return:
641 642
            list: list of (param, grad) variable pairs, param is ``Parameter``,
                grad is the gradient value corresponding to the parameter.
M
minqiyang 已提交
643

644
        Examples:
645
            See examples in ``apply_gradients``.
646
        """
647
        act_no_grad_set = None
L
Leo Chen 已提交
648
        if framework.in_dygraph_mode():
649
            pass
L
Leo Chen 已提交
650 651
        else:
            act_no_grad_set = self._get_no_grad_set(loss, no_grad_set)
G
gongweibao 已提交
652

C
chengduo 已提交
653
        self._dtype = loss.dtype
L
lujun 已提交
654
        if framework.in_dygraph_mode():
C
chengduo 已提交
655
            params_grads = []
656
            for param in self._parameter_list:
C
chengduo 已提交
657 658
                if not param.trainable:
                    continue
659
                if param._grad_ivar() is not None:
C
chengduo 已提交
660
                    # create gradient variable
661
                    grad_var = param._grad_ivar()
C
chengduo 已提交
662
                    params_grads.append((param, grad_var))
663
        else:
C
chengduo 已提交
664 665 666 667 668
            if callbacks is None:
                callbacks = [error_clip_callback]
            else:
                assert (isinstance(callbacks, list))
            program = loss.block.program
C
chengduo 已提交
669 670 671 672
            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)
673 674
            parameter_list = parameter_list if parameter_list \
                else self._parameter_list
C
chengduo 已提交
675 676
            with program_guard(program, startup_program):
                params_grads = append_backward(loss, parameter_list,
677
                                               act_no_grad_set, callbacks)
C
chengduo 已提交
678 679 680 681
                # 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
682 683 684 685 686 687 688 689

    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
690

691 692
        Returns:
            list: A list of operators appended to the current program.
M
minqiyang 已提交
693

694 695 696
        Examples:
            .. code-block:: python

697
                import paddle.fluid as fluid
698 699 700 701 702 703 704
                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)
        """
705

706 707 708 709 710
        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)

711 712 713 714 715
        # '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)
716 717 718 719 720 721 722 723 724 725 726 727

        # 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
728 729 730 731 732 733 734 735 736 737 738 739
    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
740
        if framework.in_dygraph_mode():
C
chengduo 已提交
741 742
            with program_guard(framework.default_main_program(),
                               framework.default_startup_program()):
743 744
                if self._grad_clip is not None:
                    params_grads = self._grad_clip(params_grads)
745 746
                params_grads = append_regularization_ops(params_grads,
                                                         self.regularization)
C
chengduo 已提交
747 748 749 750 751 752 753
                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
754
    def _get_no_grad_set(self, loss, no_grad_set=None):
755
        no_grad_set = _get_no_grad_set_name(no_grad_set)
G
gongweibao 已提交
756 757 758 759 760 761 762 763
        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

764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794
    @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()

795
    @imperative_base.no_grad
Q
Qiao Longfei 已提交
796 797
    def minimize(self,
                 loss,
798
                 startup_program=None,
Q
Qiao Longfei 已提交
799
                 parameter_list=None,
800
                 no_grad_set=None):
801
        """
802
        Add operations to minimize ``loss`` by updating ``parameter_list``.
M
minqiyang 已提交
803

804
        Args:
805 806 807 808
            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.
809
            parameter_list (list, optional): List of ``Variable`` or ``Variable.name`` to update
810 811
                to minimize ``loss``. The default value is None, at this time all parameters
                will be updated.
812
            no_grad_set (set, optional): Set of ``Variable``  or ``Variable.name`` that don't need
813
                to be updated. The default value is None.
Q
Qiao Longfei 已提交
814

815
        Returns:
816 817 818
            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.
819 820 821
            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``.
822 823 824

        Examples:
            Please refer to the example of current Optimizer.
Q
Qiao Longfei 已提交
825
        """
C
chengduo 已提交
826
        assert isinstance(loss, Variable), "The loss should be an Variable."
827

828 829
        parameter_list = parameter_list if parameter_list \
            else self._parameter_list
C
chengduo 已提交
830 831 832 833 834
        params_grads = self.backward(
            loss,
            startup_program=startup_program,
            parameter_list=parameter_list,
            no_grad_set=no_grad_set)
835

C
chengduo 已提交
836 837
        optimize_ops = self.apply_optimize(
            loss, startup_program=startup_program, params_grads=params_grads)
M
minqiyang 已提交
838

Q
Qiao Longfei 已提交
839
        return optimize_ops, params_grads
Q
Qiao Longfei 已提交
840 841 842


class SGDOptimizer(Optimizer):
Q
qiaolongfei 已提交
843 844 845 846 847 848 849
    """
    Optimizer of the stochastic gradient descent algorithm.

    .. math::

        param\_out = param - learning\_rate * grad

850 851 852
    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.
853 854 855
        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.
856 857 858 859 860
        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.
861 862 863 864
        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 None, meaning there is no gradient clipping.
865 866
        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
867 868 869 870

    Examples:
        .. code-block:: python

871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895
            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
896 897
    """

898 899 900 901
    def __init__(self,
                 learning_rate,
                 parameter_list=None,
                 regularization=None,
902
                 grad_clip=None,
903
                 name=None):
Q
Qiao Longfei 已提交
904
        assert learning_rate is not None
Q
Qiao Longfei 已提交
905
        super(SGDOptimizer, self).__init__(
X
Xin Pan 已提交
906
            learning_rate=learning_rate,
907
            parameter_list=parameter_list,
X
Xin Pan 已提交
908
            regularization=regularization,
909
            grad_clip=grad_clip,
X
Xin Pan 已提交
910
            name=name)
Q
Qiao Longfei 已提交
911 912
        self.type = "sgd"

913
    @no_grad
914
    def _append_optimize_op(self, block, param_and_grad):
915
        lr = self._create_param_lr(param_and_grad)
916
        if framework.in_dygraph_mode():
917 918 919
            core.ops.sgd(param_and_grad[0], lr, param_and_grad[1],
                         param_and_grad[0])
            return None
920

921
        assert isinstance(block, framework.Block)
Q
Qiao Longfei 已提交
922 923 924 925 926 927
        # create the optimize op
        sgd_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
928
                "LearningRate": lr
Q
Qiao Longfei 已提交
929
            },
M
minqiyang 已提交
930 931
            outputs={"ParamOut": param_and_grad[0]},
            stop_gradient=True)
Q
Qiao Longfei 已提交
932 933

        return sgd_op
934 935 936


class MomentumOptimizer(Optimizer):
Q
qiaolongfei 已提交
937 938 939 940 941 942 943 944 945 946 947 948 949 950
    """

    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):

951
        &\quad   param = param - (gradient + mu * velocity) * learning\_rate
Q
qiaolongfei 已提交
952 953 954

        & else:

Q
qiaolongfei 已提交
955
        &\quad   param = param - learning\_rate * velocity
Q
qiaolongfei 已提交
956

957 958 959 960
    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
961 962 963
        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.
964
        use_nesterov (bool, optional): Enables Nesterov momentum, default is false.
965 966 967 968 969
        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.
970 971 972 973
        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 None, meaning there is no gradient clipping.
974 975
        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
976 977 978 979

    Examples:
        .. code-block:: python

980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004
            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)

1005 1006 1007
    """
    _velocity_acc_str = "velocity"

X
Xin Pan 已提交
1008 1009 1010
    def __init__(self,
                 learning_rate,
                 momentum,
1011
                 parameter_list=None,
X
Xin Pan 已提交
1012 1013
                 use_nesterov=False,
                 regularization=None,
1014
                 grad_clip=None,
X
Xin Pan 已提交
1015
                 name=None):
1016 1017
        assert learning_rate is not None
        assert momentum is not None
Q
Qiao Longfei 已提交
1018
        super(MomentumOptimizer, self).__init__(
X
Xin Pan 已提交
1019
            learning_rate=learning_rate,
1020
            parameter_list=parameter_list,
X
Xin Pan 已提交
1021
            regularization=regularization,
1022
            grad_clip=grad_clip,
X
Xin Pan 已提交
1023
            name=name)
1024 1025
        self.type = "momentum"
        self._momentum = momentum
1026
        self._use_nesterov = bool(use_nesterov)
1027 1028 1029 1030 1031

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

        for p in parameters:
Q
Qiao Longfei 已提交
1032
            self._add_accumulator(self._velocity_acc_str, p)
1033 1034 1035 1036 1037 1038

    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])
1039 1040 1041 1042 1043 1044 1045 1046
        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
1047

1048
        attrs = {"mu": self._momentum, "use_nesterov": self._use_nesterov}
1049 1050 1051 1052
        inputs = {
            "Param": [param_and_grad[0]],
            "Grad": [param_and_grad[1]],
            "Velocity": [velocity_acc],
1053
            "LearningRate": [lr]
1054 1055 1056 1057 1058 1059
        }

        outputs = {
            "ParamOut": [param_and_grad[0]],
            "VelocityOut": [velocity_acc]
        }
1060 1061 1062
        # create the momentum optimize op
        momentum_op = block.append_op(
            type=self.type,
1063 1064 1065
            inputs=inputs,
            outputs=outputs,
            attrs=attrs,
M
minqiyang 已提交
1066
            stop_gradient=True)
1067 1068

        return momentum_op
1069 1070


1071
class DGCMomentumOptimizer(Optimizer):
1072
    """
S
swtkiwi 已提交
1073 1074
	:api_attr: Static Graph

1075
    DGC (Deep Gradient Compression) Momentum Optimizer. Original paper is https://arxiv.org/abs/1712.01887
1076

G
gongweibao 已提交
1077
    DGC reduces the communication bandwidth by sending only the important gradients (sparse update):\
1078 1079
        only gradients larger than a threshold are transmitted.

G
gongweibao 已提交
1080
    To avoid losing information, DGC accumulates the rest of the gradients locally.
1081 1082 1083

    Eventually, these gradients become large enough to be transmitted.

1084
    Thus, DGC sends the large gradients immediately but eventually sends all of the gradients over time.
1085

G
gongweibao 已提交
1086
    To ensure no loss of accuracy, DGC employs momentum correction and local gradient clipping on top of the gradient sparsification to maintain model performance.
1087 1088 1089 1090

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

    This optimizer will do two things:
1091

1092 1093
        1. Compress the gradient by get TopK import value from tensor \
            and use it for allreduce to reduce network bandwidth.
1094

1095
        2. Call momentum to optimize the cost.
1096 1097

    Args:
1098 1099
        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.
1100
        momentum (float): Momentum factor.
G
gongweibao 已提交
1101
        rampup_begin_step (int): The beginning step from which gradient compression is implemented.
1102 1103 1104 1105 1106 1107 1108
        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.
1109 1110 1111
        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.
1112
        use_nesterov (bool): Enables Nesterov momentum. True means use Nesterov. Default is False.
1113 1114 1115 1116 1117
        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.
1118 1119 1120
        grad_clip (GradientClipByNorm, optional): Gradient cliping strategy. ``DGCMomentumOptimizer`` only support 
            :ref:`api_fluid_clip_GradientClipByNorm` , and if not, it will raise TypeError. Default None, 
            meaning there is no gradient clipping.
1121 1122
        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.
1123 1124 1125 1126

    Examples:
        .. code-block:: python

1127
            import paddle.fluid as fluid
1128
            optimizer = fluid.optimizer.DGCMomentumOptimizer(
G
gongweibao 已提交
1129 1130 1131 1132 1133
                        learning_rate=0.0001,
                        momentum=0.9,
                        rampup_step=1000,
                        rampup_begin_step=1252,
                        sparsity=[0.999, 0.999])
1134 1135

    """
1136 1137
    _u_velocity_acc_str = "_dgc_u_"
    _v_velocity_acc_str = "_dgc_v_"
1138 1139 1140 1141 1142 1143 1144

    def __init__(self,
                 learning_rate,
                 momentum,
                 rampup_begin_step,
                 rampup_step=1,
                 sparsity=[0.999],
1145
                 parameter_list=None,
1146 1147 1148
                 use_nesterov=False,
                 num_trainers=None,
                 regularization=None,
1149
                 grad_clip=None,
1150
                 name=None):
Z
zhongpu 已提交
1151 1152
        if framework.in_dygraph_mode():
            raise Exception("In dygraph, don't support DGCMomentumOptimizer.")
1153 1154 1155 1156

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

1157 1158 1159 1160
        assert learning_rate is not None
        assert momentum is not None
        super(DGCMomentumOptimizer, self).__init__(
            learning_rate=learning_rate,
1161
            parameter_list=parameter_list,
1162
            regularization=regularization,
1163
            grad_clip=grad_clip,
1164 1165 1166 1167
            name=name)
        self.type = "dgc_momentum"
        self._momentum = momentum
        self._use_nesterov = bool(use_nesterov)
1168

1169
        assert rampup_begin_step >= 0, "rampup_begin_step must >= 0"
1170
        self._rampup_begin_step = rampup_begin_step
1171 1172
        self._rampup_step = rampup_step
        self._sparsity = sparsity
1173

1174
        self._rampup_begin_step_var = None
1175
        self._global_step_var = None
1176

1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187
        self._dgc_clip_norm = None
        if grad_clip is not None:
            if not isinstance(grad_clip, GradientClipByNorm):
                raise TypeError(
                    "The type of grad_clip should be 'GradientClipByNorm', because DGCMomentumOptimizer only support GradientClipByNorm"
                )
            assert isinstance(
                num_trainers, int
            ), "The type of num_trainers should be 'int', but received %s" % type(
                value)
            assert num_trainers > 0, "The value of num_trainers should be greater than 0!"
1188 1189

            self._num_trainers = num_trainers
1190
            self._dgc_clip_norm = grad_clip.clip_norm * (num_trainers**-0.5)
1191

1192 1193
        self.regular_type, self.regular_coeff = self._get_regularization_param(
            self.regularization)
1194

1195 1196 1197
    def _get_regularization_param(self, regularization):
        regular_type = 0
        regular_coeff = 0.0
1198

1199 1200
        if regularization is not None:
            regular_coeff = regularization._regularization_coeff
1201
            from .regularizer import L1Decay, L2Decay
1202 1203 1204 1205
            if isinstance(regularization, L1Decay):
                regular_type = 1
            elif isinstance(regularization, L2Decay):
                regular_type = 2
1206 1207
            else:
                assert False, 'regularization must be None|L1Decay|L2Deacy'
1208
        return regular_type, regular_coeff
1209

1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220
    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)
1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235
        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}
1236 1237

        if not self._is_use_dgc(param_and_grad[0], param_and_grad[1]):
1238 1239 1240
            type = "momentum"
        else:
            type = "dgc_momentum"
1241 1242 1243 1244 1245
            inputs.update({
                "current_step": self._global_step_var,
                "nranks": self._nranks_var
            })
            outputs.update({'Grad_out': param_and_grad[1]})
1246
            attrs.update({"rampup_begin_step": float(self._rampup_begin_step)})
1247 1248 1249

        # create the dgc momentum optimize op
        dgc_momentum_op = block.append_op(
1250 1251 1252 1253
            type=type,
            inputs=inputs,
            outputs=outputs,
            attrs=attrs,
1254 1255 1256
            stop_gradient=True)
        return dgc_momentum_op

1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275
    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

1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288
    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

1289 1290 1291 1292 1293 1294
    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
1295
            counter_name=core.dgc.kDGCCounterName(), begin=0)
1296

1297 1298 1299
        self._nranks_var = self._add_nranks_var(
            name=core.dgc.kDGCNRanksName(), value=-1)

1300 1301 1302 1303 1304
        # 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
1305
            name=core.dgc.kDGCRampUpBeginStepName(),
1306 1307 1308
            value=self._rampup_begin_step * 1.0,
            force_cpu=True)

1309 1310
        self.helper = LayerHelper(self.__class__.__name__)

1311
        for param_var, grad_var in param_and_grads:
1312 1313 1314
            # reuse velocity in dgc_op and dgc_momentum_op
            u_var = self._add_accumulator(self._u_velocity_acc_str, param_var)

1315
            if not self._is_use_dgc(param_var, grad_var):
1316 1317
                continue

1318
            v_var = self._add_accumulator(self._v_velocity_acc_str, param_var)
1319 1320 1321 1322 1323

            k_var = tensor.create_global_var(
                shape=[1],
                dtype=param_var.dtype,
                persistable=True,
G
gongweibao 已提交
1324
                name=param_var.name + core.dgc.kDGCKName(),
1325 1326 1327 1328 1329 1330 1331
                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
1332
                name=param_var.name + core.dgc.kDGCEncodedName(),
1333 1334 1335
                value=0.0,
                force_cpu=False)

1336 1337 1338 1339 1340 1341 1342 1343
            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)

1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362
            # 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
1363 1364
            if self._dgc_clip_norm is not None:
                clip_var = self._append_clip_norm(grad_var, self._dgc_clip_norm)
1365
            self._dgc_op(param_var, clip_var, grad_var, u_var, v_var, k_var,
1366
                         encoded_var, gather_var)
1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381

    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:
1382 1383
            name = unique_name.generate_with_ignorable_key(".".join(
                [helper.name, 'tmp']))
1384 1385 1386 1387 1388

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

        helper.append_op(
G
gongweibao 已提交
1389
            type="dgc_clip_by_norm",
1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401
            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
1402
                x=grad_var, max_norm=clip_norm, name=grad_var.name)
1403 1404

    def _dgc_op(self, param_var, clip_var, grad_var, u_var, v_var, k_var,
1405
                encoded_var, gather_var):
1406 1407
        block = framework.default_main_program().global_block()
        op_maker = core.op_proto_and_checker_maker
1408

1409 1410 1411 1412 1413 1414 1415
        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)

1416 1417 1418 1419 1420 1421
        dgc_op = block.append_op(
            type="dgc",
            inputs={
                "U": u_var,
                "V": v_var,
                "Grad": clip_var,
1422
                "Param": param_var,
1423 1424
                "current_step": self._global_step_var,
                "nranks": self._nranks_var,
1425 1426 1427 1428 1429 1430
            },
            outputs={
                "U_out": u_var,
                "V_out": v_var,
                "EncodeGrad": encoded_var,
                "k": k_var,
1431 1432
                "Grad_out": grad_var,
                "GatherBuff": gather_var,
1433 1434 1435 1436 1437 1438
            },
            attrs={
                "m": self._momentum,
                "sparsity": self._sparsity,
                "use_nesterov": self._use_nesterov,
                "rampup_begin_step": float(self._rampup_begin_step),
1439
                "rampup_step": float(self._rampup_step),
1440 1441
                "regular_coeff": float(regular_coeff),
                "regular_type": int(regular_type),
1442 1443 1444 1445 1446 1447 1448 1449
            },
            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])

1450
    @imperative_base.no_grad
1451 1452 1453 1454 1455 1456 1457
    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 = []
1458
        # DGC clip and regularization in optimizer.backward
1459 1460 1461 1462 1463 1464
        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))

1465 1466 1467 1468 1469 1470
        # '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)
1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484

        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

1485

1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500
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

1501 1502 1503 1504 1505 1506
    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.
1507 1508 1509
        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.
1510 1511 1512 1513 1514
        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.
1515 1516 1517 1518
        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 None, meaning there is no gradient clipping.
1519 1520
        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.
1521 1522 1523 1524

    Examples:
        .. code-block:: python

1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540
            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])
1541 1542 1543 1544 1545 1546 1547 1548
    """
    _velocity_acc_str = "velocity"

    def __init__(self,
                 learning_rate,
                 momentum,
                 lars_coeff=0.001,
                 lars_weight_decay=0.0005,
1549
                 parameter_list=None,
1550
                 regularization=None,
1551
                 grad_clip=None,
1552 1553 1554 1555 1556
                 name=None):
        assert learning_rate is not None
        assert momentum is not None
        super(LarsMomentumOptimizer, self).__init__(
            learning_rate=learning_rate,
1557
            parameter_list=parameter_list,
1558
            regularization=regularization,
1559
            grad_clip=grad_clip,
1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593
            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
1594 1595
            },
            stop_gradient=True)
1596 1597 1598 1599

        return momentum_op


1600
class AdagradOptimizer(Optimizer):
Q
qiaolongfei 已提交
1601
    """
1602 1603
    The Adaptive Gradient optimizer (Adagrad for short) can adaptively assign
    different learning rates to individual parameters.
Q
qiaolongfei 已提交
1604

1605
    The parameter ``param_out`` update rule with gradient ``grad``:
Q
qiaolongfei 已提交
1606 1607 1608 1609 1610 1611 1612

    .. math::

        moment\_out &= moment + grad * grad

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

1613 1614 1615 1616 1617 1618
    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
1619 1620 1621
    for numerical stability to avoid the division by zero error.

    Args:
1622 1623 1624 1625
        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.
1626 1627 1628
        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.
1629 1630 1631 1632 1633
        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.
1634 1635 1636 1637
        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 None, meaning there is no gradient clipping.
1638 1639 1640 1641 1642
        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
1643 1644 1645 1646

    Examples:
        .. code-block:: python

1647
            import numpy as np
1648
            import paddle.fluid as fluid
1649 1650

            np_inp = np.array([[1.0, 2.0], [3.0, 4.0]], dtype=np.float32)
1651
            inp = fluid.data(name="inp", shape=[2, 2])
1652 1653
            out = fluid.layers.fc(inp, size=3)
            out = fluid.layers.reduce_sum(out)
1654
            optimizer = fluid.optimizer.AdagradOptimizer(learning_rate=0.2)
1655 1656 1657 1658 1659 1660 1661
            optimizer.minimize(out)

            exe = fluid.Executor(fluid.CPUPlace())
            exe.run(fluid.default_startup_program())
            exe.run(
                feed={"inp": np_inp},
                fetch_list=[out.name])
1662 1663 1664
    """
    _moment_acc_str = "moment"

X
Xin Pan 已提交
1665 1666 1667
    def __init__(self,
                 learning_rate,
                 epsilon=1.0e-6,
1668
                 parameter_list=None,
X
Xin Pan 已提交
1669
                 regularization=None,
1670
                 grad_clip=None,
1671
                 name=None,
X
xuezhong 已提交
1672
                 initial_accumulator_value=0.0):
1673 1674
        assert learning_rate is not None
        assert epsilon is not None
Q
Qiao Longfei 已提交
1675
        super(AdagradOptimizer, self).__init__(
X
Xin Pan 已提交
1676
            learning_rate=learning_rate,
1677
            parameter_list=parameter_list,
X
Xin Pan 已提交
1678
            regularization=regularization,
1679
            grad_clip=grad_clip,
X
Xin Pan 已提交
1680
            name=name)
1681 1682
        self.type = "adagrad"
        self._epsilon = epsilon
1683
        self.initial_accumulator_value = initial_accumulator_value
1684 1685 1686 1687 1688

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

        for p in parameters:
Z
zhongpu 已提交
1689 1690 1691 1692
            self._add_accumulator(
                self._moment_acc_str,
                p,
                fill_value=self.initial_accumulator_value)
1693 1694 1695 1696 1697 1698

    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])
1699
        # Create the adagrad optimizer op
1700 1701 1702 1703 1704 1705
        adagrad_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
                "Moment": moment_acc,
1706
                "LearningRate": self._create_param_lr(param_and_grad)
1707 1708 1709
            },
            outputs={"ParamOut": param_and_grad[0],
                     "MomentOut": moment_acc},
M
minqiyang 已提交
1710 1711
            attrs={"epsilon": self._epsilon},
            stop_gradient=True)
1712 1713

        return adagrad_op
1714 1715 1716


class AdamOptimizer(Optimizer):
Q
qiaolongfei 已提交
1717
    """
T
tianshuo78520a 已提交
1718
    The Adam optimizer uses an optimization described at the end
1719 1720 1721 1722 1723
    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
1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737

    .. 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}

1738 1739
    Related paper: `Adam: A Method for Stochastic Optimization <https://arxiv.org/abs/1412.6980>`_

Q
qiaolongfei 已提交
1740
    Args:
1741 1742
        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.
1743 1744
        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.
1745
            The default value is 0.9.
1746 1747
        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.
1748 1749 1750
            The default value is 0.999.
        epsilon (float, optional): A small float value for numerical stability.
            The default value is 1e-08.
1751 1752 1753
        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.
1754 1755 1756 1757 1758
        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.
1759 1760 1761 1762
        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 None, meaning there is no gradient clipping.
1763 1764 1765 1766 1767 1768 1769 1770 1771 1772
        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
1773 1774 1775 1776

    Examples:
        .. code-block:: python

1777 1778 1779 1780 1781 1782
            import paddle
            import paddle.fluid as fluid

            place = fluid.CPUPlace()
            main = fluid.Program()
            with fluid.program_guard(main):
1783 1784
                x = fluid.data(name='x', shape=[None, 13], dtype='float32')
                y = fluid.data(name='y', shape=[None, 1], dtype='float32')
1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799
                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
1800

1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817
        .. 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
1818
                def get_decayed_betas(beta1_init, beta2_init, decay_steps, decay_rate):
1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846
                    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,
1847
                                                    beta1=beta1,
1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858
                                                    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)
1859 1860 1861
    """
    _moment1_acc_str = "moment1"
    _moment2_acc_str = "moment2"
Q
qiaolongfei 已提交
1862 1863
    _beta1_pow_acc_str = "beta1_pow_acc"
    _beta2_pow_acc_str = "beta2_pow_acc"
1864 1865 1866 1867 1868

    def __init__(self,
                 learning_rate=0.001,
                 beta1=0.9,
                 beta2=0.999,
1869
                 epsilon=1e-8,
1870
                 parameter_list=None,
X
Xin Pan 已提交
1871
                 regularization=None,
1872
                 grad_clip=None,
Q
Qiao Longfei 已提交
1873
                 name=None,
Q
Qiao Longfei 已提交
1874
                 lazy_mode=False):
1875 1876 1877 1878
        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
1879
        super(AdamOptimizer, self).__init__(
X
Xin Pan 已提交
1880
            learning_rate=learning_rate,
1881
            parameter_list=parameter_list,
X
Xin Pan 已提交
1882
            regularization=regularization,
1883
            grad_clip=grad_clip,
X
Xin Pan 已提交
1884
            name=name)
1885 1886 1887 1888
        self.type = "adam"
        self._beta1 = beta1
        self._beta2 = beta2
        self._epsilon = epsilon
Q
Qiao Longfei 已提交
1889
        self._lazy_mode = lazy_mode
1890 1891 1892 1893 1894 1895

    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
1896 1897
            self._add_accumulator(self._moment1_acc_str, p)
            self._add_accumulator(self._moment2_acc_str, p)
Q
qiaolongfei 已提交
1898 1899 1900
            self._add_accumulator(
                name=self._beta1_pow_acc_str,
                param=p,
1901 1902
                fill_value=0.9 if isinstance(self._beta1, Variable) \
                        else self._beta1,
1903
                shape=[1],
1904
                type=core.VarDesc.VarType.LOD_TENSOR, device='cpu')
Q
qiaolongfei 已提交
1905 1906 1907
            self._add_accumulator(
                name=self._beta2_pow_acc_str,
                param=p,
1908 1909
                fill_value=0.999 if isinstance(self._beta2, Variable) \
                        else self._beta2,
1910
                shape=[1],
1911
                type=core.VarDesc.VarType.LOD_TENSOR, device='cpu')
1912 1913 1914 1915 1916 1917 1918 1919

    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
1920 1921 1922 1923
        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])
1924
        lr = self._create_param_lr(param_and_grad)
1925
        # create the adam optimize op
1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940

        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

1941
        inputs = {
1942 1943
            "Param": [param_and_grad[0]],
            "Grad": [param_and_grad[1]],
1944
            "LearningRate": [lr],
1945 1946 1947 1948
            "Moment1": [moment1],
            "Moment2": [moment2],
            "Beta1Pow": [beta1_pow_acc],
            "Beta2Pow": [beta2_pow_acc]
1949 1950
        }
        outputs = {
1951 1952 1953 1954 1955
            "ParamOut": [param_and_grad[0]],
            "Moment1Out": [moment1],
            "Moment2Out": [moment2],
            "Beta1PowOut": [beta1_pow_acc],
            "Beta2PowOut": [beta2_pow_acc],
1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971
        }
        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

1972 1973
        adam_op = block.append_op(
            type=self.type,
1974 1975 1976
            inputs=inputs,
            outputs=outputs,
            attrs=attrs,
M
minqiyang 已提交
1977
            stop_gradient=True)
1978 1979 1980

        return adam_op

1981 1982

class AdamaxOptimizer(Optimizer):
Q
qiaolongfei 已提交
1983
    """
1984 1985 1986 1987
    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
1988

1989
    The parameter ``param_out`` update rule with gradient ``grad``:
Q
qiaolongfei 已提交
1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002

    .. 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}

2003
    Related paper: `Adam: A Method for Stochastic Optimization <https://arxiv.org/abs/1412.6980>`_
Q
qiaolongfei 已提交
2004

2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016
    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.
2017 2018 2019
        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.
2020 2021 2022 2023 2024
        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.
2025 2026 2027 2028
        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 None, meaning there is no gradient clipping.
2029 2030 2031 2032 2033 2034
        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
2035

2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048
    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):
2049
              data = fluid.data(name='X', shape=[None, 1], dtype='float32')
2050 2051
              hidden = fluid.layers.fc(input=data, size=10)
              loss = fluid.layers.mean(hidden)
2052
              adam = fluid.optimizer.AdamaxOptimizer(learning_rate=0.2)
2053 2054 2055 2056 2057 2058 2059 2060 2061
              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])
2062 2063 2064
    """
    _moment_acc_str = "moment"
    _inf_norm_acc_str = "inf_norm"
Q
qiaolongfei 已提交
2065
    _beta1_pow_acc_str = "beta1_pow_acc"
2066 2067 2068 2069 2070

    def __init__(self,
                 learning_rate=0.001,
                 beta1=0.9,
                 beta2=0.999,
2071
                 epsilon=1e-8,
2072
                 parameter_list=None,
X
Xin Pan 已提交
2073
                 regularization=None,
2074
                 grad_clip=None,
X
Xin Pan 已提交
2075
                 name=None):
2076 2077 2078 2079
        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
2080
        super(AdamaxOptimizer, self).__init__(
X
Xin Pan 已提交
2081
            learning_rate=learning_rate,
2082
            parameter_list=parameter_list,
X
Xin Pan 已提交
2083
            regularization=regularization,
2084
            grad_clip=grad_clip,
X
Xin Pan 已提交
2085
            name=name)
2086 2087 2088 2089 2090 2091 2092 2093
        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
2094 2095
            self._add_accumulator(self._moment_acc_str, p)
            self._add_accumulator(self._inf_norm_acc_str, p)
Q
qiaolongfei 已提交
2096 2097 2098 2099 2100
            self._add_accumulator(
                name=self._beta1_pow_acc_str,
                param=p,
                fill_value=self._beta1,
                shape=[1])
2101 2102 2103 2104 2105 2106 2107

    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
2108 2109
        beta1_pow_acc = self._get_accumulator(self._beta1_pow_acc_str,
                                              param_and_grad[0])
2110 2111 2112 2113 2114 2115
        # create the adamax optimize op
        adamax_op = block.append_op(
            type=self.type,
            inputs={
                "Param": param_and_grad[0],
                "Grad": param_and_grad[1],
2116
                "LearningRate": self._create_param_lr(param_and_grad),
2117 2118
                "Moment": moment,
                "InfNorm": inf_norm,
Q
qiaolongfei 已提交
2119
                "Beta1Pow": beta1_pow_acc
2120 2121 2122 2123 2124 2125 2126 2127 2128 2129
            },
            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
2130 2131
            },
            stop_gradient=True)
2132 2133 2134

        return adamax_op

2135
    def _finish_update(self, block, parameters_and_grads):
2136 2137 2138
        """Update Beta1 Power accumulator
        """
        assert isinstance(block, framework.Block)
2139
        for param, grad in parameters_and_grads:
C
chengduo 已提交
2140
            if grad is None or param.trainable is False:
2141
                continue
X
Xin Pan 已提交
2142 2143
            with param.block.program._optimized_guard(
                [param, grad]), name_scope('adamx'):
2144 2145
                beta1_pow_acc = self._get_accumulator(self._beta1_pow_acc_str,
                                                      param)
2146
                block.append_op(
2147 2148 2149
                    type="scale",
                    inputs={"X": beta1_pow_acc},
                    outputs={"Out": beta1_pow_acc},
M
minqiyang 已提交
2150 2151
                    attrs={"scale": self._beta1},
                    stop_gradient=True)
2152 2153


2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191
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.
2192 2193 2194
        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.
2195 2196 2197 2198 2199 2200 2201 2202
    Notes:
       Currently, DpsgdOptimizer doesn't support sparse parameter optimization.
    """

    def __init__(self,
                 learning_rate=0.001,
                 clip=0.9,
                 batch_size=0.999,
2203 2204
                 sigma=1e-8,
                 parameter_list=None):
2205 2206 2207 2208
        assert learning_rate is not None
        assert clip is not None
        assert batch_size is not None
        assert sigma is not None
2209 2210
        super(DpsgdOptimizer, self).__init__(
            learning_rate=learning_rate, parameter_list=parameter_list)
2211 2212 2213 2214
        self.type = "dpsgd"
        self._clip = clip
        self._batch_size = batch_size
        self._sigma = sigma
Z
zhongpu 已提交
2215 2216 2217 2218 2219 2220 2221
        '''
        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
2222 2223 2224 2225 2226

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

        # create the dpsgd optimize op
Z
zhongpu 已提交
2227 2228 2229
        if self._seed == None:
            self._seed = 0

2230 2231 2232 2233 2234 2235 2236 2237 2238 2239 2240
        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
2241 2242
                "sigma": self._sigma,
                "seed": self._seed
2243 2244 2245 2246 2247 2248
            },
            stop_gradient=True)

        return dpsgd_op


2249
class DecayedAdagradOptimizer(Optimizer):
2250
    """
2251 2252 2253
    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.
2254

2255
    The parameter ``param_out`` update rule with gradient ``grad``:
2256 2257 2258 2259 2260 2261 2262

    .. math::

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

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

2263 2264 2265 2266
    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
2267 2268 2269
    stability to avoid the division by zero error.

    Args:
2270 2271 2272 2273 2274
        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.
2275 2276 2277
        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.
2278 2279 2280 2281 2282
        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.
2283 2284 2285 2286
        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 None, meaning there is no gradient clipping.
2287 2288 2289 2290 2291 2292
        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.**
2293 2294 2295 2296

    Examples:
        .. code-block:: python

2297 2298
            import paddle.fluid as fluid

2299 2300 2301 2302
            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)
2303
            optimizer.minimize(cost)
2304 2305 2306
    """
    _moment_acc_str = "moment"

X
Xin Pan 已提交
2307 2308 2309 2310
    def __init__(self,
                 learning_rate,
                 decay=0.95,
                 epsilon=1.0e-6,
2311
                 parameter_list=None,
X
Xin Pan 已提交
2312
                 regularization=None,
2313
                 grad_clip=None,
X
Xin Pan 已提交
2314
                 name=None):
2315 2316 2317 2318
        assert learning_rate is not None
        assert decay is not None
        assert epsilon is not None

Q
Qiao Longfei 已提交
2319
        super(DecayedAdagradOptimizer, self).__init__(
X
Xin Pan 已提交
2320
            learning_rate=learning_rate,
2321
            parameter_list=parameter_list,
X
Xin Pan 已提交
2322
            regularization=regularization,
2323
            grad_clip=grad_clip,
X
Xin Pan 已提交
2324
            name=name)
2325 2326 2327 2328 2329 2330 2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349 2350 2351
        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},
2352 2353
            attrs={"epsilon": self._epsilon,
                   "decay": self._decay},
M
minqiyang 已提交
2354
            stop_gradient=True)
2355 2356

        return decayed_adagrad_op
2357 2358


2359
class AdadeltaOptimizer(Optimizer):
2360
    """
Z
Zeng Jinle 已提交
2361
    **Notes: This API does not support sparse parameter optimization.**
Q
qiaolongfei 已提交
2362

Z
Zeng Jinle 已提交
2363
    Adadelta Optimizer. Please refer to this for details:
Z
Zeng Jinle 已提交
2364 2365 2366
    `ADADELTA: AN ADAPTIVE LEARNING RATE METHOD <https://arxiv.org/abs/1212.5701>`_.

    The update is done as follows:
2367

Z
Zeng Jinle 已提交
2368 2369
    .. math::

Z
Zeng Jinle 已提交
2370
        E(g_t^2) &= \\rho * E(g_{t-1}^2) + (1-\\rho) * g^2
2371

Z
Zeng Jinle 已提交
2372
        learning\_rate &= \sqrt{ ( E(dx_{t-1}^2) + \\epsilon ) / ( E(g_t^2) + \\epsilon ) }
Z
Zeng Jinle 已提交
2373

Z
Zeng Jinle 已提交
2374
        E(dx_t^2) &= \\rho * E(dx_{t-1}^2) + (1-\\rho) * (-g*learning\_rate)^2
2375 2376

    Args:
Z
Zeng Jinle 已提交
2377 2378 2379
        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.
2380 2381 2382
        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.
2383 2384 2385 2386 2387
        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.
2388 2389 2390 2391
        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 None, meaning there is no gradient clipping.
2392 2393 2394
        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` .
2395 2396 2397 2398

    Examples:
        .. code-block:: python

2399
            import paddle.fluid as fluid
Z
Zeng Jinle 已提交
2400

2401
            image = fluid.data(name='image', shape=[None, 28], dtype='float32')
Z
Zeng Jinle 已提交
2402 2403
            fc = fluid.layers.fc(image, size=10)
            cost = fluid.layers.reduce_mean(fc)
2404 2405
            optimizer = fluid.optimizer.Adadelta(
                learning_rate=0.0003, epsilon=1.0e-6, rho=0.95)
C
chengduo 已提交
2406

Z
Zeng Jinle 已提交
2407 2408 2409 2410
            # 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)
2411
    """
2412

2413 2414 2415
    _avg_squared_grad_acc_str = "_avg_squared_grad"
    _avg_squared_update_acc_str = "_avg_squared_update"

X
Xin Pan 已提交
2416 2417 2418 2419
    def __init__(self,
                 learning_rate,
                 epsilon=1.0e-6,
                 rho=0.95,
2420
                 parameter_list=None,
X
Xin Pan 已提交
2421
                 regularization=None,
2422
                 grad_clip=None,
X
Xin Pan 已提交
2423
                 name=None):
2424 2425 2426 2427 2428 2429
        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.")
2430
        super(AdadeltaOptimizer, self).__init__(
X
Xin Pan 已提交
2431
            learning_rate=learning_rate,
2432
            parameter_list=parameter_list,
X
Xin Pan 已提交
2433
            regularization=regularization,
2434
            grad_clip=grad_clip,
X
Xin Pan 已提交
2435
            name=name)
2436 2437 2438 2439 2440
        self.type = "adadelta"
        self._epsilon = epsilon
        self._rho = rho

    def _create_accumulators(self, block, parameters):
2441 2442
        if not isinstance(block, framework.Block):
            raise TypeError("block is not instance of framework.Block.")
2443 2444 2445 2446 2447 2448

        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):
2449 2450
        if not isinstance(block, framework.Block):
            raise TypeError("block is not instance of framework.Block.")
2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466 2467 2468 2469 2470 2471

        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
2472 2473
                   "rho": self._rho},
            stop_gradient=True)
2474 2475 2476 2477

        return adadelta_op


Q
qingqing01 已提交
2478 2479 2480 2481 2482 2483 2484 2485 2486 2487
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
2488
        r(w, t) & = \\rho r(w, t-1) + (1 - \\rho)(\\nabla Q_{i}(w))^2
Q
qingqing01 已提交
2489 2490 2491 2492

        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
2493
    each weight. Then dividing the gradient by :math:`sqrt{v(w,t)}`.
Q
qingqing01 已提交
2494 2495 2496 2497 2498 2499

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

    ..  math::

Q
qiaolongfei 已提交
2500
        r(w, t) & = \\rho r(w, t-1) + (1 - \\rho)(\\nabla Q_{i}(w))^2
Q
qingqing01 已提交
2501

2502 2503 2504 2505 2506 2507 2508 2509 2510 2511 2512 2513 2514 2515
        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
2516 2517 2518 2519
            \\epsilon}} \\nabla Q_{i}(w)

        w & = w - v(w, t)

Q
qiaolongfei 已提交
2520
    where, :math:`\\rho` is a hyperparameter and typical values are 0.9, 0.95
Q
qingqing01 已提交
2521 2522 2523 2524 2525
    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.


2526 2527 2528
    Parameters:
        learning_rate(float): Global learning rate.
        rho(float): rho is :math: `\\rho` in equation, default is 0.95.
Q
qingqing01 已提交
2529
        epsilon(float): :math: `\\epsilon` in equation is smoothing term to
2530
            avoid division by zero, default is 1e-6.
Q
qiaolongfei 已提交
2531
        momentum(float): :math:`\\beta` in equation is the momentum term,
2532
            default is 0.0.
2533 2534 2535 2536
        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.
2537 2538 2539
        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.
2540 2541 2542 2543 2544
        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.
2545 2546 2547 2548
        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 None, meaning there is no gradient clipping.
2549 2550
        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
2551 2552 2553 2554 2555 2556 2557

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

    Examples:
          .. code-block:: python

2558 2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582
            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
2583 2584 2585 2586
    """

    _momentum_acc_str = "momentum"
    _mean_square_acc_str = "mean_square"
2587
    _mean_grad_acc_str = "mean_grad"
Q
qingqing01 已提交
2588 2589 2590 2591 2592 2593

    def __init__(self,
                 learning_rate,
                 rho=0.95,
                 epsilon=1.0e-6,
                 momentum=0.0,
2594
                 centered=False,
2595
                 parameter_list=None,
X
Xin Pan 已提交
2596
                 regularization=None,
2597
                 grad_clip=None,
X
Xin Pan 已提交
2598
                 name=None):
Q
qingqing01 已提交
2599
        super(RMSPropOptimizer, self).__init__(
X
Xin Pan 已提交
2600
            learning_rate=learning_rate,
2601
            parameter_list=parameter_list,
X
Xin Pan 已提交
2602
            regularization=regularization,
2603
            grad_clip=grad_clip,
X
Xin Pan 已提交
2604
            name=name)
Q
qingqing01 已提交
2605 2606 2607 2608 2609 2610 2611 2612 2613 2614 2615 2616 2617
        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
2618
        self._centered = centered
Q
qingqing01 已提交
2619 2620 2621 2622 2623 2624 2625 2626

    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)
2627
            self._add_accumulator(self._mean_grad_acc_str, p)
Q
qingqing01 已提交
2628 2629 2630 2631 2632 2633 2634 2635 2636

    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])
2637 2638
        mean_grad_acc = self._get_accumulator(self._mean_grad_acc_str,
                                              param_and_grad[0])
Q
qingqing01 已提交
2639 2640 2641 2642 2643 2644 2645
        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,
2646
                "MeanGrad": mean_grad_acc,
Q
qingqing01 已提交
2647 2648 2649 2650 2651
                "LearningRate": self._create_param_lr(param_and_grad),
            },
            outputs={
                "ParamOut": param_and_grad[0],
                "MomentOut": momentum_acc,
2652 2653
                "MeanSquareOut": mean_square_acc,
                "MeanGradOut": mean_grad_acc
Q
qingqing01 已提交
2654 2655 2656 2657
            },
            attrs={
                "epsilon": self._epsilon,
                "decay": self._rho,
2658 2659
                "momentum": self._momentum,
                "centered": self._centered
M
minqiyang 已提交
2660 2661
            },
            stop_gradient=True)
Q
qingqing01 已提交
2662 2663 2664 2665

        return rmsprop_op


Q
qiaolongfei 已提交
2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704 2705
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

2706 2707 2708 2709 2710
    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.
2711 2712 2713
        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.
2714 2715 2716 2717 2718
        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.
2719 2720 2721 2722
        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 None, meaning there is no gradient clipping.
2723 2724
        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
2725 2726 2727 2728 2729 2730 2731

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

    Examples:
          .. code-block:: python

2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748 2749 2750 2751 2752 2753 2754 2755
            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
2756

2757
    NOTE:
C
chengduo 已提交
2758
       Currently, FtrlOptimizer doesn't support sparse parameter optimization.
Q
qiaolongfei 已提交
2759 2760 2761 2762 2763
    """

    _squared_acc_str = "squared"
    _linear_acc_str = "linear"

X
Xin Pan 已提交
2764 2765 2766 2767 2768
    def __init__(self,
                 learning_rate,
                 l1=0.0,
                 l2=0.0,
                 lr_power=-0.5,
2769
                 parameter_list=None,
X
Xin Pan 已提交
2770
                 regularization=None,
2771
                 grad_clip=None,
X
Xin Pan 已提交
2772
                 name=None):
Q
qiaolongfei 已提交
2773
        super(FtrlOptimizer, self).__init__(
X
Xin Pan 已提交
2774
            learning_rate=learning_rate,
2775
            parameter_list=parameter_list,
X
Xin Pan 已提交
2776
            regularization=regularization,
2777
            grad_clip=grad_clip,
X
Xin Pan 已提交
2778
            name=name)
Q
qiaolongfei 已提交
2779 2780 2781 2782 2783 2784 2785 2786 2787 2788 2789 2790 2791 2792 2793 2794 2795 2796 2797 2798 2799 2800 2801 2802 2803 2804 2805 2806 2807 2808 2809 2810 2811 2812 2813 2814 2815 2816 2817 2818
        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
2819 2820
                   "lr_power": self._lr_power},
            stop_gradient=True)
Q
qiaolongfei 已提交
2821 2822 2823 2824

        return ftrl_op


Y
Yibing Liu 已提交
2825 2826 2827 2828 2829 2830
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
2831 2832
    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
2833 2834 2835 2836 2837

    The updating of parameters follows:

    ..  math::

Y
Yibing Liu 已提交
2838
        m_t &= \\beta_1 m_{t - 1}+ (1 - \\beta_1)g_t 
Y
Yibing Liu 已提交
2839

Y
Yibing Liu 已提交
2840
        v_t &= \\beta_2 v_{t - 1}  + (1 - \\beta_2)g_t^2
Y
Yibing Liu 已提交
2841

Y
Yibing Liu 已提交
2842
        r_t &= \\frac{m_t}{\\sqrt{v_t}+\\epsilon}
Y
Yibing Liu 已提交
2843

Y
Yibing Liu 已提交
2844
        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
2845 2846 2847 2848 2849 2850


    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
2851 2852 2853 2854 2855 2856 2857 2858
        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.
2859 2860 2861
        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.
2862 2863 2864 2865 2866
        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.
2867 2868 2869 2870
        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 None, meaning there is no gradient clipping.
Y
Yibing Liu 已提交
2871 2872 2873 2874 2875
        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
2876 2877 2878 2879 2880 2881

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

Y
Yibing Liu 已提交
2882
            data = fluid.data(name='x', shape=[-1, 5], dtype='float32')
Y
Yibing Liu 已提交
2883 2884 2885
            hidden = fluid.layers.fc(input=data, size=10)
            cost = fluid.layers.mean(hidden)

Y
Yibing Liu 已提交
2886 2887 2888 2889 2890
            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
2891 2892 2893 2894
            optimizer.minimize(cost)
    """
    _moment1_acc_str = "moment1"
    _moment2_acc_str = "moment2"
Y
Yibing Liu 已提交
2895
    # these two not used in op temporarily
Y
Yibing Liu 已提交
2896 2897 2898 2899 2900 2901 2902 2903 2904
    _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,
2905
                 parameter_list=None,
Y
Yibing Liu 已提交
2906
                 regularization=None,
2907
                 grad_clip=None,
Y
Yibing Liu 已提交
2908
                 exclude_from_weight_decay_fn=None,
Y
Yibing Liu 已提交
2909 2910 2911 2912 2913 2914 2915 2916
                 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,
2917
            parameter_list=parameter_list,
Y
Yibing Liu 已提交
2918
            regularization=regularization,
2919
            grad_clip=grad_clip,
Y
Yibing Liu 已提交
2920 2921 2922 2923 2924 2925
            beta1=beta1,
            beta2=beta2,
            epsilon=epsilon,
            name=name)
        self.type = "lamb"
        self._weight_decay = lamb_weight_decay
Y
Yibing Liu 已提交
2926
        self._exclude_from_weight_decay_fn = exclude_from_weight_decay_fn
Y
Yibing Liu 已提交
2927 2928 2929

    def _append_optimize_op(self, block, param_and_grad):
        assert isinstance(block, framework.Block)
2930
        block.program._use_lamb = True
Y
Yibing Liu 已提交
2931 2932 2933 2934 2935 2936 2937 2938 2939 2940

        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
2941 2942 2943 2944 2945 2946
        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
2947 2948 2949 2950 2951 2952 2953 2954 2955 2956 2957 2958 2959 2960 2961 2962 2963 2964 2965 2966 2967
        # 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
2968
                "weight_decay": weight_decay
Y
Yibing Liu 已提交
2969 2970 2971 2972 2973 2974
            },
            stop_gradient=True)

        return lamb_op


2975 2976 2977 2978 2979 2980 2981 2982 2983 2984 2985 2986 2987
# 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
2988
Dpsgd = DpsgdOptimizer
2989
DecayedAdagrad = DecayedAdagradOptimizer
2990
Adadelta = AdadeltaOptimizer
Q
qingqing01 已提交
2991
RMSProp = RMSPropOptimizer
Q
qiaolongfei 已提交
2992
Ftrl = FtrlOptimizer
2993
LarsMomentum = LarsMomentumOptimizer
Y
Yibing Liu 已提交
2994
Lamb = LambOptimizer
2995 2996 2997


class ModelAverage(Optimizer):
2998
    """
S
swtkiwi 已提交
2999 3000
	:api_attr: Static Graph

3001 3002 3003 3004 3005 3006 3007 3008 3009 3010 3011 3012 3013 3014 3015 3016 3017 3018
    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:

    ::
3019

3020 3021 3022 3023 3024 3025 3026 3027 3028
        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.
3029 3030

    Args:
3031 3032 3033
        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.
3034 3035 3036 3037 3038
        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.
3039 3040 3041
        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.
3042

3043
    Examples:
Q
qiaolongfei 已提交
3044 3045 3046

      .. code-block:: python

3047 3048 3049 3050 3051 3052
        import paddle.fluid as fluid
        import numpy

        # First create the Executor.
        place = fluid.CPUPlace()  # fluid.CUDAPlace(0)
        exe = fluid.Executor(place)
3053

3054 3055 3056 3057
        train_program = fluid.Program()
        startup_program = fluid.Program()
        with fluid.program_guard(train_program, startup_program):
            # build net
3058
            data = fluid.data(name='X', shape=[None, 1], dtype='float32')
3059 3060 3061 3062 3063 3064 3065 3066
            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,
3067
                                                         max_average_window=12500)
3068 3069

            exe.run(startup_program)
3070 3071 3072 3073 3074
            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])
3075 3076

            # apply ModelAverage
3077
            with model_average.apply(exe):
3078 3079 3080 3081
                x = numpy.random.random(size=(10, 1)).astype('float32')
                exe.run(program=train_program,
                        feed={'X': x},
                        fetch_list=[loss.name])
3082 3083 3084
    """

    def __init__(self,
W
wanghaoshuang 已提交
3085
                 average_window_rate,
3086 3087
                 min_average_window=10000,
                 max_average_window=10000,
X
Xin Pan 已提交
3088 3089
                 regularization=None,
                 name=None):
Z
zhongpu 已提交
3090 3091
        if framework.in_dygraph_mode():
            raise Exception("In dygraph, don't support ModelAverage.")
X
Xin Pan 已提交
3092 3093
        super(ModelAverage, self).__init__(
            0.0, regularization=regularization, name=name)
3094 3095 3096
        self.average_window = average_window_rate
        self.min_average_window = min_average_window
        self.max_average_window = max_average_window
3097

3098
        self.params_grads = []
3099 3100
        for param in framework.default_main_program().global_block(
        ).all_parameters():
3101
            if param.do_model_average != False:
3102
                grad = param.block.create_var(
3103 3104
                    name=unique_name.generate_with_ignorable_key(".".join(
                        [param.name, 'tmp'])),
3105 3106
                    dtype=param.dtype,
                    persistable=False,
W
wanghaoshuang 已提交
3107
                    stop_gradient=True)
3108
                self.params_grads.append((param, grad))
3109

3110
        for param, grad in self.params_grads:
3111 3112
            if grad is None:
                continue
X
Xin Pan 已提交
3113 3114
            with param.block.program._optimized_guard(
                [param, grad]), name_scope('move_average'):
3115
                self._append_average_accumulate_op(param)
3116

3117 3118 3119 3120
        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:
3121
                self._add_average_apply_op(block, param_grad)
3122 3123 3124 3125 3126

        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:
3127
                self._add_average_restore_op(block, param_grad)
3128

3129
    def _add_average_apply_op(self, block, param_grad):
L
Luo Tao 已提交
3130 3131 3132 3133 3134 3135
        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(
3136
            self._get_accumulator('num_accumulates', param))
L
Luo Tao 已提交
3137
        old_num_accumulates = block._clone_variable(
3138
            self._get_accumulator('old_num_accumulates', param))
L
Luo Tao 已提交
3139
        num_updates = block._clone_variable(
3140 3141 3142 3143 3144 3145
            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
3146 3147 3148 3149
        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
3150
        ops._elementwise_div(x=sum, y=tmp, out=param)
3151 3152

    def _add_average_restore_op(self, block, param_grad):
L
Luo Tao 已提交
3153 3154
        param = block._clone_variable(param_grad[0])
        grad = block._clone_variable(param_grad[1])
3155 3156 3157 3158 3159 3160 3161 3162 3163 3164 3165 3166 3167 3168 3169 3170 3171 3172 3173 3174 3175 3176 3177 3178 3179 3180 3181 3182 3183 3184 3185 3186 3187 3188 3189 3190 3191
        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
3192 3193
            },
            stop_gradient=True)
3194

S
rename  
sneaxiy 已提交
3195
    @signature_safe_contextmanager
3196
    def apply(self, executor, need_restore=True):
3197 3198
        """
        Apply the average of the cumulative ``Parameter`` to the parameters of the current model.
3199 3200

        Args:
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 3228 3229 3230 3231 3232 3233 3234 3235 3236 3237 3238 3239 3240 3241 3242 3243 3244
            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])
3245
        """
3246 3247 3248 3249 3250 3251
        executor.run(self.apply_program)
        try:
            yield
        finally:
            if need_restore:
                self.restore(executor)
3252 3253

    def restore(self, executor):
3254 3255
        """
        Restore ``Parameter`` values of current model.
3256 3257
        
        Args:
3258 3259 3260 3261 3262 3263 3264 3265 3266 3267 3268 3269 3270 3271 3272 3273 3274 3275 3276 3277 3278 3279 3280 3281 3282 3283 3284 3285 3286 3287 3288 3289 3290 3291 3292 3293 3294 3295 3296 3297 3298 3299 3300 3301
            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)
3302
        """
3303
        executor.run(self.restore_program)
3304 3305 3306 3307


class ExponentialMovingAverage(object):
    """
S
swtkiwi 已提交
3308 3309
	:api_attr: Static Graph

3310 3311 3312 3313 3314 3315
    Compute the moving average of parameters with exponential decay.
    Given a parameter :math:`\\theta`, its exponential moving average (EMA)
    will be

    ..  math::

3316
        \\text{EMA}_0 & = 0
3317

3318 3319
	\\text{EMA}_t & = \\text{decay} * \\text{EMA}_{t-1} + (1 - \\text{decay}) * \\theta_t

Y
Yibing Liu 已提交
3320 3321 3322 3323
    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.
3324 3325 3326 3327 3328 3329 3330 3331 3332 3333 3334 3335 3336 3337 3338 3339 3340 3341 3342 3343 3344

    **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.
3345 3346 3347


    Args:
Y
Yibing Liu 已提交
3348 3349 3350 3351 3352 3353 3354
	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.
3355 3356 3357 3358 3359


    Examples:

	.. code-block:: python
3360 3361 3362 3363 3364

	    import numpy
	    import paddle
	    import paddle.fluid as fluid

Y
Yibing Liu 已提交
3365
	    data = fluid.data(name='x', shape=[-1, 5], dtype='float32')
3366 3367 3368 3369 3370 3371 3372 3373
	    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)

3374
	    global_steps = fluid.layers.autoincreased_step_counter()
3375 3376 3377 3378 3379 3380 3381 3382 3383 3384 3385 3386 3387 3388 3389 3390 3391 3392 3393 3394 3395 3396 3397 3398 3399 3400 3401 3402 3403
	    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)
3404 3405
    """

3406
    def __init__(self, decay=0.999, thres_steps=None, name=None):
Z
zhongpu 已提交
3407 3408 3409
        if framework.in_dygraph_mode():
            raise Exception(
                "In dygraph, don't support ExponentialMovingAverage.")
3410
        self._decay = decay
3411
        self._thres_steps = thres_steps
3412
        self._name = name if name is not None else ''
3413 3414
        self._decay_var = self._get_ema_decay()

3415
        self._step_counter_name = "@EMA_STEP_COUNTER@"
Y
Yibing Liu 已提交
3416
        self._params_tmps = []
3417
        for param in default_main_program().global_block().all_parameters():
3418 3419 3420 3421 3422 3423 3424
            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
3425
                self._params_tmps.append((param, tmp))
3426

Y
Yibing Liu 已提交
3427 3428
        self._ema_vars = {}
        for param, tmp in self._params_tmps:
3429 3430
            with param.block.program._optimized_guard(
                [param, tmp]), name_scope('moving_average'):
Y
Yibing Liu 已提交
3431
                self._ema_vars[param.name] = self._create_ema_vars(param)
3432 3433 3434 3435

        self.apply_program = Program()
        block = self.apply_program.global_block()
        with program_guard(main_program=self.apply_program):
3436
            decay_pow, global_step = self._get_decay_pow(block)
Y
Yibing Liu 已提交
3437
            for param, tmp in self._params_tmps:
3438 3439
                param = block._clone_variable(param)
                tmp = block._clone_variable(tmp)
Y
Yibing Liu 已提交
3440
                ema = block._clone_variable(self._ema_vars[param.name])
3441
                layers.assign(input=param, output=tmp)
3442
                # bias correction
3443 3444 3445
                with layers.control_flow.Switch() as switch:
                    with switch.case(global_step > 0):
                        layers.assign(output=ema, input=ema / (1.0 - decay_pow))
3446 3447 3448 3449 3450
                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
3451
            for param, tmp in self._params_tmps:
3452 3453 3454 3455
                tmp = block._clone_variable(tmp)
                param = block._clone_variable(param)
                layers.assign(input=tmp, output=param)

3456 3457 3458 3459 3460 3461 3462 3463 3464 3465 3466 3467 3468 3469 3470 3471 3472 3473 3474 3475 3476 3477
    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):
3478 3479 3480 3481 3482 3483 3484
        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")
3485
        decay_var = block._clone_variable(self._decay_var)
3486 3487
        decay_pow_acc = layers.elementwise_pow(decay_var, global_step)
        return decay_pow_acc, global_step
3488

Y
Yibing Liu 已提交
3489
    def _create_ema_vars(self, param):
3490 3491 3492 3493 3494 3495 3496 3497 3498
        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
3499 3500 3501 3502 3503
    def update(self):
        """ 
        Update Exponential Moving Average. Should only call this method in 
        train program.
        """
3504 3505
        global_step = layers.autoincreased_step_counter(
            counter_name=self._step_counter_name)
3506
        param_master_emas = []
Y
Yibing Liu 已提交
3507 3508 3509 3510
        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]
3511
                if param.name + '.master' in self._ema_vars:
3512 3513 3514 3515 3516 3517 3518 3519 3520 3521 3522 3523 3524 3525 3526 3527 3528
                    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
3529

3530 3531 3532 3533 3534 3535 3536
    @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
3537 3538
            need_restore (bool, optional): Whether to restore parameters after 
                applying. Default True.
3539 3540 3541 3542 3543 3544 3545 3546 3547 3548 3549 3550 3551 3552 3553
        """
        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
3554 3555 3556


class PipelineOptimizer(object):
3557
    """
S
swtkiwi 已提交
3558 3559
	:api_attr: Static Graph

3560
    Pipeline Optimizer
H
hutuxian 已提交
3561

T
tianshuo78520a 已提交
3562
    Train with pipeline mode. The program will be split by cut_list. 
H
hutuxian 已提交
3563 3564

    If the len of cut_list is k, then the whole program (including \
T
tianshuo78520a 已提交
3565
    backward part) will be split to 2*k-1 sections. 
H
hutuxian 已提交
3566 3567 3568 3569
    
    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, \
3570
    the final performance depends on the training progress of each pipeline heavily.
H
hutuxian 已提交
3571 3572 3573

    And we will try the synchronous mode in the future.

3574
    Args:
H
hutuxian 已提交
3575 3576 3577 3578
        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.
3579 3580
        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
3581 3582 3583 3584
                        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].

3585 3586
    Examples:
        .. code-block:: python
H
hutuxian 已提交
3587

3588
            import paddle.fluid as fluid
H
hutuxian 已提交
3589 3590 3591 3592 3593 3594 3595 3596 3597 3598 3599 3600 3601 3602 3603 3604 3605 3606 3607 3608 3609 3610 3611 3612 3613 3614 3615 3616 3617 3618 3619 3620 3621 3622
            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)
3623 3624
    """

H
hutuxian 已提交
3625 3626 3627 3628 3629 3630 3631 3632
    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
3633 3634
        if framework.in_dygraph_mode():
            raise Exception("In dygraph, don't support PipelineOptimizer.")
H
hutuxian 已提交
3635 3636 3637 3638 3639 3640 3641 3642 3643
        # 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
3644
    def _create_vars(self, block, main_program):
H
hutuxian 已提交
3645 3646 3647 3648 3649 3650 3651 3652 3653 3654 3655
        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
3656
    def _extract_section_opt_ops(self, ops, cut_point_name):
H
hutuxian 已提交
3657 3658 3659 3660 3661 3662 3663 3664 3665 3666 3667 3668 3669 3670 3671
        """
        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
3672
    def _find_input_output(self, ops, name, is_forward=True):
H
hutuxian 已提交
3673 3674 3675 3676 3677 3678 3679 3680 3681 3682 3683 3684 3685 3686
        """
        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
3687
    def _find_persistable_vars(self, ops, whole_parameters):
H
hutuxian 已提交
3688 3689 3690 3691 3692 3693 3694 3695 3696 3697 3698 3699 3700 3701 3702 3703 3704 3705 3706 3707 3708 3709 3710 3711 3712 3713 3714
        """
        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
3715
    def _extract_section_ops(self, ops, cut_point_name):
H
hutuxian 已提交
3716 3717 3718 3719 3720 3721 3722 3723 3724 3725 3726 3727 3728 3729 3730 3731 3732 3733 3734
        """
        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
3735 3736
    def _find_section_opt(self, ops, params):
        res = self._extract_section_opt_ops(ops, params)
H
hutuxian 已提交
3737 3738
        return res

H
hutuxian 已提交
3739
    def _split_program(self, main_program, cut_list):
H
hutuxian 已提交
3740 3741 3742 3743 3744 3745 3746 3747 3748 3749 3750 3751 3752 3753 3754 3755 3756 3757 3758 3759
        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
3760
            cur_ops = self._extract_section_ops(ops, cut_vars)
H
hutuxian 已提交
3761 3762 3763 3764 3765 3766
            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
3767
                self._find_input_output(
H
hutuxian 已提交
3768 3769 3770 3771 3772 3773
                    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
3774
                    self._find_persistable_vars(cur_ops, whole_parameters))
H
hutuxian 已提交
3775
            if i >= cut_len - 1:
H
hutuxian 已提交
3776 3777
                opt_ops = self._find_section_opt(
                    ops, sec_params[2 * cut_len - 2 - i])
H
hutuxian 已提交
3778 3779 3780 3781 3782 3783 3784 3785 3786 3787

                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
3788
                self._find_input_output(
H
hutuxian 已提交
3789 3790 3791
                    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
3792
                self._find_input_output(
H
hutuxian 已提交
3793 3794 3795 3796 3797 3798 3799 3800 3801 3802 3803 3804 3805 3806
                    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
3807
            self._find_input_output(
H
hutuxian 已提交
3808 3809 3810 3811 3812 3813 3814 3815 3816 3817 3818 3819 3820 3821 3822 3823 3824 3825 3826 3827
                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
3828 3829 3830 3831 3832 3833 3834 3835
        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
3836 3837 3838 3839 3840 3841 3842 3843 3844 3845 3846 3847 3848 3849 3850 3851 3852 3853 3854 3855
        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
3856 3857


M
mapingshuo 已提交
3858 3859
class RecomputeOptimizer(Optimizer):
    """
S
swtkiwi 已提交
3860 3861
	:api_attr: Static Graph

M
mapingshuo 已提交
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 3919 3920 3921
    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
3922 3923
        if framework.in_dygraph_mode():
            raise Exception("In dygraph, don't support RecomputeOptimizer.")
M
mapingshuo 已提交
3924 3925
        self._optimizer = optimizer
        self._checkpoints = None
M
mapingshuo 已提交
3926 3927
        self._learning_rate = self._optimizer._learning_rate
        self._learning_rate_map = self._optimizer._learning_rate_map
M
mapingshuo 已提交
3928 3929 3930 3931 3932 3933

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

    def load(self, stat_dict):
        """
S
swtkiwi 已提交
3934 3935
	:api_attr: Static Graph

M
mapingshuo 已提交
3936 3937 3938 3939 3940 3941 3942 3943 3944 3945 3946 3947 3948 3949 3950 3951 3952 3953 3954 3955 3956 3957 3958 3959 3960 3961 3962 3963 3964 3965 3966 3967 3968 3969 3970 3971 3972 3973 3974 3975 3976 3977 3978 3979 3980 3981 3982 3983 3984 3985 3986 3987 3988 3989 3990 3991 3992 3993 3994 3995 3996 3997 3998 3999 4000 4001 4002
        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)
4003
                sgd._set_checkpoints([fc_1, pred])
M
mapingshuo 已提交
4004 4005 4006 4007
                params_grads = sgd.backward(
                    cost,
                    startup_program=None,
                    parameter_list=None,
4008
                    no_grad_set=None)
M
mapingshuo 已提交
4009 4010 4011 4012 4013 4014 4015 4016 4017 4018 4019 4020 4021 4022 4023

                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,
4024
                 callbacks=None):
M
mapingshuo 已提交
4025 4026 4027 4028 4029 4030 4031
        """
        call append_backward with checkpoints.

        Args:
            loss (Variable): loss variable to run optimizations.
            startup_program (Program): startup_program for initializing parameters
                in `parameter_list`.
4032 4033
            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
4034 4035 4036 4037 4038 4039 4040 4041 4042 4043 4044 4045 4046 4047 4048 4049 4050 4051 4052 4053 4054 4055 4056 4057
            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)
4058
                sgd._set_checkpoints([fc_1, pred])
M
mapingshuo 已提交
4059 4060 4061 4062
                params_grads = sgd.backward(
                    cost,
                    startup_program=None,
                    parameter_list=None,
4063
                    no_grad_set=None)
M
mapingshuo 已提交
4064 4065 4066 4067 4068 4069 4070 4071 4072 4073 4074 4075 4076 4077 4078
                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)
4079 4080
            # Note: since we can't use all_reduce_op now,
            #  dgc_op should be the last op of one grad.
M
mapingshuo 已提交
4081 4082
            if hasattr(self._optimizer, "_append_dgc_ops"):
                self._optimizer._append_dgc_ops(params_grads)
M
mapingshuo 已提交
4083 4084 4085 4086 4087 4088 4089 4090 4091 4092 4093 4094 4095 4096 4097 4098 4099 4100 4101
        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
4102
                    return sum_cost, fc_1, prediction                
M
mapingshuo 已提交
4103 4104 4105 4106 4107 4108 4109 4110
                
                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)
4111
                sgd._set_checkpoints([fc_1, pred])
M
mapingshuo 已提交
4112 4113 4114 4115
                params_grads = sgd.backward(
                    cost,
                    startup_program=None,
                    parameter_list=None,
4116
                    no_grad_set=None)
M
mapingshuo 已提交
4117 4118 4119 4120 4121 4122 4123 4124 4125 4126 4127 4128 4129 4130
                
                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,
4131
                 no_grad_set=None):
4132
        assert isinstance(loss, Variable), "The loss should be an Variable."
M
mapingshuo 已提交
4133 4134 4135 4136 4137 4138 4139 4140 4141
        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")
        params_grads = self.backward(
            loss,
            startup_program=startup_program,
            parameter_list=parameter_list,
4142
            no_grad_set=no_grad_set)
M
mapingshuo 已提交
4143 4144 4145 4146 4147 4148 4149

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

        return optimize_ops, params_grads


M
mapingshuo 已提交
4150 4151
class LookaheadOptimizer(object):
    """
S
swtkiwi 已提交
4152 4153
	:api_attr: Static Graph

M
mapingshuo 已提交
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
    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
4207 4208
        if framework.in_dygraph_mode():
            raise Exception("In dygraph, don't support LookaheadOptimizer.")
M
mapingshuo 已提交
4209 4210 4211 4212 4213 4214 4215 4216 4217 4218 4219 4220 4221 4222 4223 4224 4225 4226 4227 4228 4229 4230 4231 4232 4233 4234 4235 4236 4237 4238 4239 4240 4241 4242 4243 4244 4245 4246 4247 4248 4249 4250 4251 4252 4253 4254 4255 4256 4257 4258 4259 4260 4261 4262 4263 4264 4265 4266 4267 4268 4269 4270 4271 4272 4273 4274 4275 4276 4277 4278 4279 4280 4281 4282 4283 4284 4285 4286 4287 4288 4289 4290 4291 4292 4293 4294 4295 4296 4297 4298 4299 4300 4301 4302 4303 4304
        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