optimizer.py 50.0 KB
Newer Older
M
MRXLT 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# 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
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# 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.

from __future__ import print_function

import numpy as np
import six
import logging
from collections import defaultdict

22
import paddle
M
MRXLT 已提交
23 24 25 26 27 28 29 30
from paddle.fluid.distribute_lookup_table import find_distributed_lookup_table
from paddle.fluid.framework import Program, Variable, name_scope, default_main_program, default_startup_program, device_guard

from ..fluid import framework
from ..fluid import layers
from ..fluid import unique_name
from ..fluid.backward import append_backward, _some_in_set_, _append_grad_suffix_, _get_no_grad_set_name
from ..fluid.clip import GradientClipBase, GradientClipByNorm, error_clip_callback, append_gradient_clip_ops
31
from ..fluid.framework import program_guard, Parameter
M
MRXLT 已提交
32 33 34 35 36 37 38 39 40 41
from ..fluid.initializer import Constant
from ..fluid.layer_helper import LayerHelper
from ..fluid.layers import ops
from ..fluid.dygraph import base as imperative_base
from ..fluid.dygraph import no_grad
from paddle.fluid import core
from paddle.fluid.layers import tensor
from functools import reduce
from ..fluid.wrapped_decorator import signature_safe_contextmanager
from .. import compat as cpt
42
from .lr import LRScheduler
43
import copy
M
MRXLT 已提交
44

45 46
__all__ = []

M
MRXLT 已提交
47 48

class Optimizer(object):
49
    r"""Optimizer Base class.
M
MRXLT 已提交
50 51 52 53 54 55

    Define the common interface of an optimizer.
    User should not use this class directly,
    but need to use one of it's implementation.

    Args:
56 57
        learning_rate (float|LRScheduler): The learning rate used to update ``Parameter``.
            It can be a float value or any subclass of ``LRScheduler`` .
58
        parameters (list|tuple, optional): List/Tuple of ``Tensor`` names to update to minimize ``loss``. \
59 60 61 62
            This parameter is required in dygraph mode. And you can specify different options for \
            different parameter groups such as the learning rate, weight decay, etc, \
            then the parameters are list of dict. Note that the learning_rate in paramter groups \
            represents the scale of base learning_rate. \
M
MRXLT 已提交
63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87
            The default value is None in static mode, at this time all parameters will be updated.
        weight_decay (float|WeightDecayRegularizer, optional): The strategy of regularization. \
            It canbe a float value as coeff of L2 regularization or \
            :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.
        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.
        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.

    Returns:
       Base class for optimizer. 
    
    Examples:
        .. code-block:: python

            #Take the subclass adam as an example
            import paddle
            linear = paddle.nn.Linear(10, 10)
88
            inp = paddle.uniform(shape=[10, 10], min=-0.1, max=0.1)
M
MRXLT 已提交
89 90 91 92 93 94 95 96
            out = linear(inp)
            loss = paddle.mean(out)
            adam = paddle.optimizer.Adam(learning_rate=0.1,
                    parameters=linear.parameters())
            out.backward()
            adam.step()
            adam.clear_grad()

97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119
            #Take the subclass sgd as an example
            #optimize parameters in linear_1 and linear2 in different options. 
            #Note that the learning_rate of linear_2 is 0.01.
            linear_1 = paddle.nn.Linear(10, 10)
            linear_2 = paddle.nn.Linear(10, 10)
            inp = paddle.uniform(shape=[10, 10], min=-0.1, max=0.1)
            out = linear_1(inp)
            out = linear_2(out)
            loss = paddle.mean(out)
            sgd = paddle.optimizer.SGD(
                learning_rate=0.1,
                parameters=[{
                    'params': linear_1.parameters()
                }, {
                    'params': linear_2.parameters(),
                    'weight_decay': 0.001,
                    'learning_rate': 0.1
                }],
                weight_decay=0.01)                   
            out.backward()
            sgd.step()
            sgd.clear_grad()

M
MRXLT 已提交
120 121
    """

122
    @imperative_base.no_grad
M
MRXLT 已提交
123 124 125 126 127 128
    def __init__(self,
                 learning_rate,
                 parameters=None,
                 weight_decay=None,
                 grad_clip=None,
                 name=None):
129

130 131 132 133 134 135 136 137 138
        if parameters is not None:
            # paddle.Tensor is also iterable, so here we don't check whether
            # the input is iterable, if the input is paddle.Tensor, the
            # list(paddle.Tensor) will be a error value
            if isinstance(parameters, paddle.Tensor):
                raise TypeError(
                    "`parameters` argument given to the optimizer should be "
                    "an iterable of paddle Tensors, but got argument type is `{}`.".
                    format(type(parameters)))
139 140 141 142 143
            if isinstance(parameters, dict):
                raise TypeError(
                    "`parameters` argument should not get dict type, "
                    "if parameter groups is needed, please set `parameters`"
                    " as list of dict")
144 145 146 147
            self._parameter_list = list(parameters)
        else:
            self._parameter_list = None

M
MRXLT 已提交
148 149 150 151 152 153 154
        self._name = name
        if framework.in_dygraph_mode():
            if self._parameter_list is None:
                raise AttributeError(
                    "parameters argument given to the Optimizer should not be None in dygraph mode."
                )
            if weight_decay is not None:
155 156 157 158 159 160 161 162 163 164 165
                if not isinstance(self._parameter_list[0], dict):
                    for param in self._parameter_list:
                        if hasattr(
                                param,
                                'regularizer') and param.regularizer is not None:
                            logging.info(
                                "If regularizer of a Parameter has been set by 'paddle.ParamAttr' or 'static.WeightNormParamAttr' already. "
                                "The weight_decay[%s] in Optimizer will not take effect, and it will only be applied to other Parameters!"
                                % weight_decay.__str__())
                            break

166
        if not isinstance(learning_rate, (float, LRScheduler)):
167
            raise TypeError(
168
                "learning rate should be float or LRScheduler, got %s here" %
169
                type(learning_rate))
M
MRXLT 已提交
170 171 172 173 174 175 176 177 178 179 180 181
        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"
                )
        if isinstance(weight_decay, float):
            from ..fluid.regularizer import L2Decay
            self.regularization = L2Decay(weight_decay)
        else:
            self.regularization = weight_decay
        self._grad_clip = grad_clip
        self._learning_rate = learning_rate
L
Leo Chen 已提交
182

M
MRXLT 已提交
183
        self._dtype = None
L
Leo Chen 已提交
184 185
        # Infer the dtype form parameter
        if self._parameter_list:
186 187 188 189 190 191 192
            if isinstance(self._parameter_list[0], dict):
                for param_group in self._parameter_list:
                    assert 'params' in param_group, \
                        'params should be set in parameters if parameter groups are optimized in different options'
                self._dtype = self._parameter_list[0]['params'][0].dtype
            else:
                self._dtype = self._parameter_list[0].dtype
L
Leo Chen 已提交
193

M
MRXLT 已提交
194 195 196 197 198 199 200 201 202 203 204 205 206
        # each program should have a independent learning rate
        # program -> tensor(learning_rate)
        self._learning_rate_map = dict()
        # Dictionary of accumulators. Some optimizer subclasses need to
        # allocate and manage extra tensors associated with the parameters
        # to train. These tensors are called accumulators.
        # {accum_name : { paramter_name : accumulator_for_parameter, ...}, ...}
        self._accumulators = defaultdict(lambda: dict())
        self.helper = None
        self._opti_name_list = []
        self._accumulators_holder = {}
        self._param_device_map = dict()
        self.clear_gradients = self.clear_grad
207 208 209 210 211 212 213 214 215 216 217 218
        self._default_dict = {
            'learning_rate': self._learning_rate,
            'weight_decay': self.regularization,
            'grad_clip': self._grad_clip
        }

        self._param_groups = []
        if self._parameter_list and isinstance(self._parameter_list[0], dict):
            for param_group in self._parameter_list:
                self._add_param_group(param_group.copy())
        else:
            self._param_groups = self._parameter_list
M
MRXLT 已提交
219 220 221 222

    @framework.dygraph_only
    def state_dict(self):
        '''
223
        Get state dict information from optimizer. It contain all the tensor used by optimizer. For Adam optimizer, contains beta1, beta2, momentum etc. If LRScheduler have been used, global_step will be include in state dict.
M
MRXLT 已提交
224 225 226 227 228 229 230 231 232 233 234 235
        If the optimizer never be called(minimize function), the state_dict is empty.

        Args: 
            None

        Returns:
            state_dict(dict) : dict contains all the Tensor used by optimizer
        
        Examples:
            .. code-block:: python

                import paddle
M
MRXLT 已提交
236
                emb = paddle.nn.Embedding(10, 10)
M
MRXLT 已提交
237 238 239 240 241 242 243 244 245 246

                adam = paddle.optimizer.Adam(0.001, parameters=emb.parameters())
                state_dict = adam.state_dict()

        '''
        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
247
        if isinstance(self._learning_rate, LRScheduler):
M
MRXLT 已提交
248 249 250 251 252 253
            state_dict["LR_Scheduler"] = self._learning_rate.state_dict()
        return state_dict

    @framework.dygraph_only
    def set_state_dict(self, state_dict):
        '''
254
        Load optimizer state dict. For Adam optimizer, contains beta1, beta2, momentum etc. If LRScheduler have been used, global_step will be changed.
M
MRXLT 已提交
255 256 257 258 259 260 261 262 263 264 265

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

                import paddle

266
                emb = paddle.nn.Embedding(10, 10)
M
MRXLT 已提交
267

268 269
                layer_state_dict = emb.state_dict()
                paddle.save(layer_state_dict, "emb.pdparams")
M
MRXLT 已提交
270

271 272 273 274 275 276 277
                scheduler = paddle.optimizer.lr.NoamDecay(	
                    d_model=0.01, warmup_steps=100, verbose=True)
                adam = paddle.optimizer.Adam(
                    learning_rate=scheduler,
                    parameters=emb.parameters())
                opt_state_dict = adam.state_dict()
                paddle.save(opt_state_dict, "adam.pdopt")
M
MRXLT 已提交
278

279
                opti_state_dict = paddle.load("adam.pdopt")
M
MRXLT 已提交
280 281 282
                adam.set_state_dict(opti_state_dict)

        '''
283
        if isinstance(self._learning_rate, LRScheduler):
M
MRXLT 已提交
284
            self._learning_rate.set_dict(state_dict["LR_Scheduler"])
M
MRXLT 已提交
285

286
        if isinstance(self._learning_rate, LRScheduler):
287
            self._learning_rate.set_state_dict(state_dict["LR_Scheduler"])
M
MRXLT 已提交
288

289 290 291 292 293
        # NOTE: exclude learning rate scheduler's state from 
        # _accumulators_holder.
        state_dict = state_dict.copy()
        if "LR_Scheduler" in state_dict:
            state_dict.pop("LR_Scheduler")
M
MRXLT 已提交
294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316
        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 Tensor {} not found".format( var_tmp.name )
                var = var_tmp.value()
                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):
                    load_para_np = load_para.numpy()
                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(
J
Jiangxinz 已提交
317
                                                 model_np.name, model_np.shape, load_para_np.shape)
M
MRXLT 已提交
318 319 320

                assert model_np.dtype == load_para_np.dtype, \
                                          "Parameter dtype not match, Dygraph Parameter [ {} ] need tensor with dtype {}  but load tensor with dtype {}".format(
J
Jiangxinz 已提交
321
                                                model_np.name, model_np.dtype, load_para_np.dtype)
M
MRXLT 已提交
322 323 324 325 326 327 328

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

    def get_opti_var_name_list(self):
        return self._opti_name_list

    def _create_global_learning_rate(self):
329
        if isinstance(self._learning_rate, LRScheduler):
330 331 332 333 334 335 336 337 338 339 340 341 342 343 344
            lr_var = self._global_learning_rate()
            # only create global lr_var once
            if not isinstance(lr_var, framework.Variable):
                lr_name = unique_name.generate('learning_rate')
                self._learning_rate._var_name = lr_name
                lr_var = self.helper.create_global_variable(
                    name=lr_name,
                    shape=[1],
                    persistable=True,
                    stop_gradient=True,
                    dtype=paddle.get_default_dtype()
                    if self._dtype is None else self._dtype)
                main_prog = framework.default_main_program()
                main_prog.lr_sheduler = self._learning_rate
                main_prog.lr_var = lr_var
M
MRXLT 已提交
345

M
MRXLT 已提交
346
                self._learning_rate_map[framework.default_main_program(
347
                )] = lr_var
M
MRXLT 已提交
348

349 350 351 352 353 354
            lr_value = float(self._learning_rate())
            self.helper.set_variable_initializer(
                lr_var, initializer=Constant(value=lr_value))
        elif isinstance(self._learning_rate, float):
            # only create global lr_var once
            lr = self._global_learning_rate()
M
MRXLT 已提交
355 356 357
            if isinstance(lr, framework.Variable):
                return
            else:
358 359 360 361 362 363 364 365
                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=paddle.get_default_dtype()
                    if self._dtype is None else self._dtype,
                    persistable=True)
M
MRXLT 已提交
366 367 368 369 370 371

    @framework.dygraph_only
    def set_lr(self, value):
        """
        :api_attr: imperative
        
372
        Set the value of the learning rate manually in the optimizer. If the optimizer use LRScheduler,
M
MRXLT 已提交
373 374 375
        this API cannot be invoked, because it will lead to conflict.

        Args:
M
MRXLT 已提交
376
            value (float): the value of learning rate
M
MRXLT 已提交
377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402

        Returns:
            None
          
        Examples:
            .. code-block:: python

                import paddle
                linear = paddle.nn.Linear(10, 10)

                adam = paddle.optimizer.Adam(0.1, parameters=linear.parameters())

                # set learning rate manually by python float value
                lr_list = [0.2, 0.3, 0.4, 0.5, 0.6]
                for i in range(5):
                    adam.set_lr(lr_list[i])
                    lr = adam.get_lr()
                    print("current lr is {}".format(lr))
                # Print:
                #    current lr is 0.2
                #    current lr is 0.3
                #    current lr is 0.4
                #    current lr is 0.5
                #    current lr is 0.6

        """
403
        if not isinstance(value, (int, float)):
M
MRXLT 已提交
404
            raise TypeError(
405
                "The type of 'value' in optimizer.set_lr must be float, but received %s."
M
MRXLT 已提交
406
                % (type(value)))
407
        if isinstance(self._learning_rate, LRScheduler):
M
MRXLT 已提交
408
            raise RuntimeError(
409
                "optimizer's learning rate can't be LRScheduler when invoke this API, because this will lead to conflict."
M
MRXLT 已提交
410
            )
411 412 413 414 415 416 417 418 419 420 421 422 423
        self._learning_rate = float(value)
        current_lr = self._global_learning_rate()
        if current_lr is not None:
            global_block = framework.default_main_program().global_block()
            global_block.append_op(
                type='fill_constant',
                outputs={'Out': [current_lr]},
                attrs={
                    'dtype': current_lr.dtype,
                    'shape': list(current_lr.shape),
                    'value': float(value)
                },
                stop_gradient=True)
M
MRXLT 已提交
424 425 426

    def get_lr(self):
        """
427 428 429
        Get current learning rate of optimizer. 
        If 'LRScheduler' is not used, the return value is all the same.
        If 'LRScheduler' is used, the return value is the current scheduled learing rete.
M
MRXLT 已提交
430

M
MRXLT 已提交
431
        Returns:
432
            float: The current learning rate of optimizer.
M
MRXLT 已提交
433 434 435 436

        Examples:
            .. code-block:: python

437
                # train on default dynamic graph mode
M
MRXLT 已提交
438
                import paddle
439 440 441 442 443 444 445 446 447 448 449
                import numpy as np
                emb = paddle.nn.Embedding(10, 3)

                ## example1: LRScheduler is not used, return the same value is all the same
                adam = paddle.optimizer.Adam(0.01, parameters = emb.parameters())
                for batch in range(10):
                    input = paddle.randint(low=0, high=5, shape=[5])
                    out = emb(input)
                    out.backward()
                    print("Learning rate of step{}: {}".format(batch, adam.get_lr())) # 0.01
                    adam.step()
M
MRXLT 已提交
450

451 452 453 454 455 456 457 458
                ## example2: StepDecay is used, return the scheduled learning rate
                scheduler = paddle.optimizer.lr.StepDecay(learning_rate=0.5, step_size=2, gamma=0.1)
                adam = paddle.optimizer.Adam(scheduler, parameters = emb.parameters())
                for batch in range(10):
                    input = paddle.randint(low=0, high=5, shape=[5])
                    out = emb(input)
                    out.backward()
                    print("Learning rate of step{}: {}".format(batch, adam.get_lr())) # 0.5->0.05...
M
MRXLT 已提交
459
                    adam.step()
460
                    scheduler.step()
461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479

                # train on static graph mode
                paddle.enable_static()
                main_prog = paddle.static.Program()
                start_prog = paddle.static.Program()
                with paddle.static.program_guard(main_prog, start_prog):
                    x = paddle.static.data(name='x', shape=[None, 10])
                    z = paddle.static.nn.fc(x, 100)
                    loss = paddle.mean(z)
                    scheduler = paddle.optimizer.lr.StepDecay(learning_rate=0.5, step_size=2, gamma=0.1)
                    adam = paddle.optimizer.Adam(learning_rate=scheduler)
                    adam.minimize(loss)

                exe = paddle.static.Executor()
                exe.run(start_prog)
                for batch in range(10):
                    print("Learning rate of step{}: {}", adam.get_lr())     # 0.5->0.05->0.005...
                    out = exe.run(main_prog, feed={'x': np.random.randn(3, 10).astype('float32')})
                    scheduler.step()
M
MRXLT 已提交
480 481 482 483 484

        """
        if isinstance(self._learning_rate, float):
            return self._learning_rate
        else:
485
            return self._learning_rate()
M
MRXLT 已提交
486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505

    def _global_learning_rate(self, program=None):
        """
        get global decayed learning rate
        :return:
        """
        if program is None:
            program = framework.default_main_program()
        return self._learning_rate_map.get(program, None)

    def _append_optimize_op(self, block, param_and_grad):
        """ append optimize operator to block and return all the added optimize_op
        """
        raise NotImplementedError(
            "Class \"Optimizer\" connot be used directly as an optimizer, please use its subclasses such as \"Adam\""
        )

    def _create_param_lr(self, param_and_grad):
        # create learning rate tensor for every parameter
        param = param_and_grad[0]
506 507 508 509
        if hasattr(param, 'optimize_attr'):
            param_lr = param.optimize_attr['learning_rate']
            if type(param_lr) == Variable:
                return param_lr
M
MRXLT 已提交
510
            else:
511 512 513 514 515 516 517 518 519
                if param_lr == 1.0:
                    return self._global_learning_rate()
                else:
                    with default_main_program()._lr_schedule_guard(
                            is_with_opt=True), framework.name_scope(
                                'scale_with_param_lr'):
                        return self._global_learning_rate() * param_lr
        else:
            return self._global_learning_rate()
M
MRXLT 已提交
520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617

    def _create_accumulators(self, block, parameters):
        """Create all accumulators needed by the parameters

        Args:
            block: the block in which the loss tensor is present
            parameters: list of parameter tensors for the optimizer
        """
        pass

    def _finish_update(self, block, parameters_and_grads):
        """Finish any custom updates needed
           before completing an optimization step

        Args:
            block: the block in which the loss tensor is present
            parameters: list of parameter tensors for the optimizer

        Returns:
            None
        """
        pass

    def _add_accumulator(self,
                         name,
                         param,
                         dtype=None,
                         fill_value=0.0,
                         shape=None,
                         type=None,
                         device=None):
        """Utility function to add an accumulator for a parameter

        Args:
            block: the block in which the loss tensor is present
            name: name of the accumulator
            param: parameter tensor for which accumulator is to be added
            dtype: data type of the accumulator tensor
            fill_value: value to initialize the accumulator tensor
        """
        if self._name is not None:
            name = self._name + "_" + name
        if (name in self._accumulators and
                param.name in self._accumulators[name]):
            if framework.in_dygraph_mode():
                return self._accumulators[name][param.name]
            raise Exception("Accumulator {} already exists for parameter {}".
                            format(name, param.name))
        if shape == None:
            shape = param.shape
        assert isinstance(self.helper, LayerHelper)

        var_name = param.name + "_" + name
        var_name = unique_name.generate(var_name)
        self._opti_name_list.append(var_name)

        var = self.helper.create_global_variable(
            name=var_name,
            persistable=True,
            dtype=dtype or param.dtype,
            type=param.type if type is None else type,
            shape=shape,
            belong_to_optimizer=True)
        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)))

        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])

        self._accumulators[name][param.name] = var
        return var

    def _get_accumulator(self, name, param):
        """Utility function to fetch an accumulator for a parameter

        Args:
            name: name of the accumulator
            param: parameter tensor for which accumulator is to be fetched

        Returns:
            accumulator tensor for the parameter
        """
        if self._name is not None:
            name = self._name + "_" + name
        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]

    def _update_param_device_map(self, parameters_and_grads, target_block):
        for param_and_grad in parameters_and_grads:
618
            if param_and_grad[0].stop_gradient is False:
M
MRXLT 已提交
619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670
                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)
                        break

    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

    def _create_optimization_pass(self, parameters_and_grads):
        """Add optimization operators to update gradients to tensors.

        Args:
          parameters_and_grads(list(tuple(Tensor, Tensor))):
            a list of (tensor, gradient) pair to update.

        Returns:
          return_op_list: a list of operators that will complete one step of
            optimization. This will include parameter update ops, global step
            update ops and any other custom ops required by subclasses to manage
            their internal state.
        """
        # 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
        # for parameters and extend _finish_update method to add custom ops.

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

        global_block = framework.default_main_program().global_block()
        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)
        self.helper = LayerHelper(self.__class__.__name__)
671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686
        params_grads_device_map = parameters_and_grads['params'] if isinstance(
            parameters_and_grads, dict) else parameters_and_grads
        self._update_param_device_map(params_grads_device_map, target_block)
        if isinstance(parameters_and_grads, list):
            self._create_accumulators(
                target_block,
                [p[0] for p in parameters_and_grads if not p[0].stop_gradient])

        else:
            params_acc_dict = parameters_and_grads.copy()
            params_acc_dict['params'] = [
                p[0] for p in params_acc_dict['params']
                if not p[0].stop_gradient
            ]
            self._create_accumulators(target_block, params_acc_dict)

M
MRXLT 已提交
687 688 689
        self._create_global_learning_rate()

        if framework.in_dygraph_mode():
690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709

            if isinstance(parameters_and_grads, list):
                for param_and_grad in parameters_and_grads:
                    if param_and_grad[1] is None:
                        continue
                    if param_and_grad[0].stop_gradient is False:
                        self._append_optimize_op(target_block, param_and_grad)
            else:
                for param_and_grad in parameters_and_grads['params']:
                    if param_and_grad[1] is None:
                        continue
                    if param_and_grad[0].stop_gradient is False:
                        param_grad_dict = dict()
                        param_grad_dict['params'] = param_and_grad
                        param_grad_dict.update({
                            k: v
                            for k, v in parameters_and_grads.items()
                            if k != 'params'
                        })
                        self._append_optimize_op(target_block, param_grad_dict)
M
MRXLT 已提交
710 711 712 713 714 715
        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"):
716
                    if param_and_grad[0].stop_gradient is False:
M
MRXLT 已提交
717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766
                        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)

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

        end = len(target_block.ops)
        return target_block._slice_ops(start, end)

    def _append_dgc_ops(self, param_and_grad):
        pass

    def backward(self,
                 loss,
                 startup_program=None,
                 parameters=None,
                 no_grad_set=None,
                 callbacks=None):
        """
        The first part of ``minimize``, do auto-diff to append backward operations for
        the current program.

        Args:
            loss (Tensor): ``loss`` tensor to run optimizations.
            startup_program (Program, optional): :ref:`api_fluid_Program` for
                initializing parameters in ``parameters``. The default value
                is None, at this time :ref:`api_fluid_default_startup_program` will be used.
            parameters (list, optional): List of ``Tensor`` or ``Tensor.name`` to update
                to minimize ``loss``. The default value is None, at this time all parameters
                will be updated.
            no_grad_set (set, optional): Set of ``Tensor``  or ``Tensor.name`` that don't need
                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.

        Return:
            list: list of (param, grad) tensor pairs, param is ``Parameter``,
                grad is the gradient value corresponding to the parameter.

        Examples:
            .. code-block:: python

                import paddle
                import numpy as np
                value = np.arange(26).reshape(2, 13).astype("float32")
                a = paddle.to_tensor(value)
M
MRXLT 已提交
767
                linear = paddle.nn.Linear(13, 5)
M
MRXLT 已提交
768 769 770 771 772 773 774 775 776 777 778 779 780 781
                # This can be any optimizer supported by dygraph.
                adam = paddle.optimizer.Adam(learning_rate = 0.01, 
                                            parameters = linear.parameters())
                out = linear(a)
                out.backward()
                adam.step()
                adam.clear_grad()
        """
        act_no_grad_set = None
        if framework.in_dygraph_mode():
            pass
        else:
            act_no_grad_set = self._get_no_grad_set(loss, no_grad_set)

L
Leo Chen 已提交
782 783 784 785
        # Infer dtype by loss if None
        if self._dtype is None:
            self._dtype = loss.dtype

M
MRXLT 已提交
786
        if framework.in_dygraph_mode():
787 788 789
            parameter_list = parameters if parameters \
                else self._parameter_list

M
MRXLT 已提交
790
            params_grads = []
791
            for param in parameter_list:
792
                if param.stop_gradient:
M
MRXLT 已提交
793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856
                    continue
                if param._grad_ivar() is not None:
                    # create gradient tensor
                    grad_var = param._grad_ivar()
                    params_grads.append((param, grad_var))
        else:
            if callbacks is None:
                callbacks = [error_clip_callback]
            else:
                assert (isinstance(callbacks, list))
            program = loss.block.program
            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 paddle.mean to process the current loss.".format(
                    loss.shape)
            parameter_list = parameters if parameters \
                else self._parameter_list
            with program_guard(program, startup_program):
                params_grads = append_backward(loss, parameter_list,
                                               act_no_grad_set, callbacks)
                # 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

    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.

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

        Examples:
            .. code-block:: python

                import paddle
                import numpy as np

                inp = np.random.uniform(-0.1, 0.1, [10, 10]).astype("float32")
                linear = paddle.nn.Linear(10, 10)
                inp = paddle.to_tensor(inp)
                out = linear(inp)
                loss = paddle.mean(out)
                optimizer = paddle.optimizer.Adam(learning_rate=0.1,
                        parameters=linear.parameters())
                params_grads = optimizer.backward(loss)
                optimizer.apply_gradients(params_grads)

        """

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

        # 'optimizer(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)

        # Add regularization if any
857 858
        params_grads = self.append_regularization_ops(params_grads,
                                                      self.regularization)
M
MRXLT 已提交
859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877

        optimize_ops = self._create_optimization_pass(params_grads)
        return optimize_ops

    def _apply_optimize(self, loss, startup_program, params_grads):
        """
        Second part of `minimize`, appending optimization operators for
        given `params_grads` pairs.
        Args:
            loss (Tensor): loss tensor to run optimizations.
            startup_program (Program): startup_program for initializing parameters
                in `parameters`.
            params_grads (list): list of (param, grad) pair to do optimization.
        Returns:
            list: A list of operators appended to the current program.
        """
        if framework.in_dygraph_mode():
            with program_guard(framework.default_main_program(),
                               framework.default_startup_program()):
878 879 880
                if isinstance(params_grads, list):
                    if self._grad_clip is not None:
                        params_grads = self._grad_clip(params_grads)
881
                    params_grads = self.append_regularization_ops(
882 883 884 885 886 887 888
                        params_grads, self.regularization)
                else:
                    grad_clip = params_grads['grad_clip']
                    if grad_clip is not None:
                        params_grads['params'] = grad_clip(params_grads[
                            'params'])

889
                    params_grads['params'] = self.append_regularization_ops(
890
                        params_grads['params'], self.regularization)
M
MRXLT 已提交
891 892 893 894 895 896 897
                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

898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917
    def _create_regularization_of_grad(self, param, grad, regularization=None):
        """ Create and add backward regularization Operators
    
        Function helper of append_regularization_ops.
        """
        # If no gradient or no regularization is specified,  then we don't need to do anything
        if grad is None or ((not hasattr(param, 'regularizer') or
                             (hasattr(param, 'regularizer') and
                              param.regularizer is None)) and
                            regularization is None):
            return grad
        regularization_term = None
        if hasattr(param, 'regularizer') and param.regularizer is not None:
            # Add variable for regularization term in grad block
            regularization_term = param.regularizer(param, grad, grad.block)
        elif regularization is not None:
            regularization_term = regularization(param, grad, grad.block)

        assert regularization_term is not None

918 919 920
        if framework.in_dygraph_mode():
            return core.ops.sum([grad, regularization_term])

921 922 923 924 925 926 927 928 929 930 931 932 933 934 935
        new_grad = grad
        if grad.type == core.VarDesc.VarType.SELECTED_ROWS:
            # FIXME(zcd): If the grad is SELECTED_ROWS, after regularization,
            # the grad's type and name will be changed. But the gradient's name
            # is used in ParallelExecutor Reduce mode, so I add a flag for
            # the new_grad here.
            new_grad = grad.block.create_var(
                name=grad.name + core.kNewGradSuffix(),
                dtype=param.dtype,
                shape=param.shape,
                lod_level=param.lod_level,
                type=core.VarDesc.VarType.LOD_TENSOR)

        inputs = {"X": [grad, regularization_term]}
        outputs = {"Out": [new_grad]}
936
        grad.block.append_op(type='sum', inputs=inputs, outputs=outputs)
937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984

        return new_grad

    def append_regularization_ops(self,
                                  parameters_and_grads,
                                  regularization=None):
        r"""Create and add backward regularization Operators
    
        Creates and adds backward regularization operators in the BlockDesc.
        This will add gradients of the regularizer function to the gradients
        of the parameters and return these modified gradients. This is the
        same as implementing weight decay in optimizers for regularization.
    
        Args:
            parameters_and_grads: A list of (parameters, gradients) pairs
                                  that need to be regularized.
            regularization: A global regularizer. If the parameter is not
                            set. It will be applied with regularizer.
    
        Returns:
            list[(Variable, Variable)]: list of (parameters, gradients) \
            pair with the regularized gradient
    
        Raises:
            Exception: Unknown regularization type
        """
        params_and_grads = []
        if framework.in_dygraph_mode():
            for param, grad in parameters_and_grads:
                new_grad = self._create_regularization_of_grad(param, grad,
                                                               regularization)
                params_and_grads.append((param, new_grad))
        else:
            repeate_regularizer = False
            with framework.name_scope('regularization'):
                for param, grad in parameters_and_grads:
                    if not repeate_regularizer and param.regularizer is not None and regularization is not None:
                        repeate_regularizer = True
                        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__())
                    with param.block.program._optimized_guard([param, grad]):
                        new_grad = self._create_regularization_of_grad(
                            param, grad, regularization)
                        params_and_grads.append((param, new_grad))
        return params_and_grads

M
MRXLT 已提交
985 986 987
    def _get_no_grad_set(self, loss, no_grad_set=None):
        no_grad_set = _get_no_grad_set_name(no_grad_set)
        parameters = loss.block.program.global_block().all_parameters()
988 989 990
        param_no_trainable = set([
            param.name for param in parameters if param.stop_gradient is True
        ])
M
MRXLT 已提交
991 992 993 994 995 996 997 998 999
        # If the parameter is no trainable, it should not have a gradient.
        no_grad_set.update(param_no_trainable)

        return no_grad_set

    @framework.dygraph_only
    def clear_grad(self):
        """
        Clear the gradients of all optimized parameters for model.
1000 1001

        If not, new gradient will accumulat on previous gradient.
M
MRXLT 已提交
1002 1003 1004 1005 1006 1007 1008 1009 1010
        
        Returns:
            None
        
        Examples:
            .. code-block:: python

                import numpy as np
                import paddle
1011

M
MRXLT 已提交
1012 1013
                value = np.arange(26).reshape(2, 13).astype("float32")
                a = paddle.to_tensor(value)
M
MRXLT 已提交
1014
                linear = paddle.nn.Linear(13, 5)
M
MRXLT 已提交
1015 1016 1017 1018 1019 1020 1021 1022 1023
                # This can be any optimizer supported by dygraph.
                adam = paddle.optimizer.Adam(learning_rate = 0.01, 
                                            parameters = linear.parameters())
                out = linear(a)
                out.backward()
                adam.step()
                adam.clear_grad()

        """
1024 1025 1026 1027 1028 1029 1030 1031 1032 1033
        if self._parameter_list is None or not isinstance(
                self._parameter_list[0], dict):
            for p in self._parameter_list:
                if not p.stop_gradient:
                    p.clear_gradient()
        else:
            for param_group in self._param_groups:
                for p in param_group['params']:
                    if not p.stop_gradient:
                        p.clear_gradient()
M
MRXLT 已提交
1034

1035
    @imperative_base.no_grad
M
MRXLT 已提交
1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058
    def minimize(self,
                 loss,
                 startup_program=None,
                 parameters=None,
                 no_grad_set=None):
        """
        Add operations to minimize ``loss`` by updating ``parameters``.

        Args:
            loss (Tensor): A ``Tensor`` containing the value to minimize.
            startup_program (Program, optional): :ref:`api_fluid_Program` for
                initializing parameters in ``parameters``. The default value
                is None, at this time :ref:`api_fluid_default_startup_program` will be used.
            parameters (list, optional): List of ``Tensor`` or ``Tensor.name`` to update
                to minimize ``loss``. The default value is None, at this time all parameters
                will be updated.
            no_grad_set (set, optional): Set of ``Tensor``  or ``Tensor.name`` that don't need
                to be updated. The default value is None.

        Returns:
            tuple: tuple (optimize_ops, params_grads), A list of operators appended
            by minimize and a list of (param, grad) tensor pairs, param is
            ``Parameter``, grad is the gradient value corresponding to the parameter.
M
MRXLT 已提交
1059
            In static graph mode, the returned tuple can be passed to ``fetch_list`` in ``Executor.run()`` to 
M
MRXLT 已提交
1060 1061 1062 1063 1064
            indicate program pruning. If so, the program will be pruned by ``feed`` and 
            ``fetch_list`` before run, see details in ``Executor``.

        Examples:
            .. code-block:: python
M
MRXLT 已提交
1065
 
M
MRXLT 已提交
1066
                import paddle
M
MRXLT 已提交
1067
                linear = paddle.nn.Linear(10, 10)
1068 1069
                input = paddle.uniform(shape=[10, 10], min=-0.1, max=0.1)
                out = linear(input)
M
MRXLT 已提交
1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081
                loss = paddle.mean(out)

                beta1 = paddle.to_tensor([0.9], dtype="float32")
                beta2 = paddle.to_tensor([0.99], dtype="float32")

                adam = paddle.optimizer.Adam(learning_rate=0.1,
                        parameters=linear.parameters(),
                        weight_decay=0.01)
                out.backward()
                adam.minimize(loss)
                adam.clear_grad()

M
MRXLT 已提交
1082 1083 1084 1085 1086
        """
        assert isinstance(loss, Variable), "The loss should be an Tensor."

        parameter_list = parameters if parameters \
            else self._parameter_list
1087

M
MRXLT 已提交
1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098
        params_grads = self.backward(
            loss,
            startup_program=startup_program,
            parameters=parameter_list,
            no_grad_set=no_grad_set)

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

        return optimize_ops, params_grads

L
Leo Chen 已提交
1099
    @imperative_base.no_grad
M
MRXLT 已提交
1100 1101 1102
    @framework.dygraph_only
    def step(self):
        """
M
MRXLT 已提交
1103
        Execute the optimizer and update parameters once.
M
MRXLT 已提交
1104 1105 1106 1107 1108 1109 1110 1111 1112
        
        Returns:
            None

        Examples:
            .. code-block:: python

                import paddle
                import numpy as np
1113

M
MRXLT 已提交
1114 1115
                value = np.arange(26).reshape(2, 13).astype("float32")
                a = paddle.to_tensor(value)
M
MRXLT 已提交
1116
                linear = paddle.nn.Linear(13, 5)
M
MRXLT 已提交
1117 1118 1119 1120 1121 1122 1123 1124
                # This can be any optimizer supported by dygraph.
                adam = paddle.optimizer.Adam(learning_rate = 0.01, 
                                            parameters = linear.parameters())
                out = linear(a)
                out.backward()
                adam.step()
                adam.clear_grad()
        """
1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203

        if not isinstance(self._param_groups[0], dict):
            params_grads = []
            for param in self._param_groups:
                if param.stop_gradient:
                    continue
                if param._grad_ivar() is not None:
                    grad_var = param._grad_ivar()
                    params_grads.append((param, grad_var))

            self._apply_optimize(
                loss=None, startup_program=None, params_grads=params_grads)

        else:
            # optimize parameters in groups
            for param_group in self._param_groups:
                params_grads = defaultdict(lambda: list())
                for param in param_group['params']:
                    if param.stop_gradient:
                        continue
                    if param._grad_ivar() is not None:
                        grad_var = param._grad_ivar()
                        params_grads['params'].append((param, grad_var))
                params_grads.update(
                    {k: v
                     for k, v in param_group.items() if k != 'params'})
                self._apply_optimize(
                    loss=None, startup_program=None, params_grads=params_grads)

    def _add_param_group(self, param_group):
        """
        Add a param group to parameter_list.

        Args:
            param_group (dict): The group of Tensors to be optimzed with
            different optimization options.
        """
        params = param_group['params']
        if isinstance(params, Parameter):
            param_group['params'] = [params]
        elif isinstance(params, set):
            raise TypeError(
                "optimizer parameters should be in ordered collections,"
                "but received set, please use list instead.")
        else:
            param_group['params'] = list(params)

        # Update optimization options for each groups
        for k, v in self._default_dict.items():
            param_group.setdefault(k, v)

        param_set = set()
        for group in self._param_groups:
            param_set.update(set(group['params']))

        if not param_set.isdisjoint(set(param_group['params'])):
            raise ValueError(
                "some parameters appear in more than one parameter group")

        for param in param_group['params']:
            weight_decay = param_group['weight_decay']
            if isinstance(weight_decay, float):
                from ..fluid.regularizer import L2Decay
                regularization = L2Decay(weight_decay)
            else:
                regularization = weight_decay
            param.regularizer = regularization
            param.optimize_attr['learning_rate'] = param_group['learning_rate']

        self._param_groups.append(param_group)

    def _update_param_group(self, parameters):
        """
        Update the param group with new entry
        Args:
            parameters (dict): The extra group of Tensors to be optimzed with
            different optimization options. Only used in child class.
        """
        pass