You need to sign in or sign up before continuing.
momentum.py 26.8 KB
Newer Older
J
Jiawei Wang 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14
# Copyright (c) 2020 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.

J
Jiangxinz 已提交
15 16
import warnings

J
Jiawei Wang 已提交
17 18 19
from .optimizer import Optimizer
from ..fluid import core
from ..fluid import framework
20
from ..fluid.layer_helper import LayerHelper
H
huangxu96 已提交
21 22 23
from ..fluid import unique_name
from ..fluid import layers
from paddle.fluid.regularizer import L2DecayRegularizer
24
from paddle import _C_ops, _legacy_C_ops
25
import paddle
26
from paddle.fluid.framework import in_dygraph_mode, _in_legacy_dygraph
J
Jiawei Wang 已提交
27

28 29
__all__ = []

J
Jiawei Wang 已提交
30 31

class Momentum(Optimizer):
32
    r"""
J
Jiawei Wang 已提交
33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56

    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):

        &\quad   param = param - (gradient + mu * velocity) * learning\_rate

        & else:

        &\quad   param = param - learning\_rate * velocity

    Parameters:

        learning_rate (float|Tensor|LearningRateDecay, optional): The learning rate used to update ``Parameter``.
            It can be a float value, a ``Tensor`` with a float type or a LearningRateDecay. The default value is 0.001.
        momentum (float): Momentum factor. The default value is 0.9.
57 58 59 60 61
        parameters (list|tuple, optional): List|Tuple of ``Tensor`` to update to minimize ``loss``. \
            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. \
J
Jiawei Wang 已提交
62 63
            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. \
64 65 66 67 68 69
            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.
J
Jiawei Wang 已提交
70 71 72 73
        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.
H
huangxu96 已提交
74 75 76
        multi_precision (bool, optional): Whether to use multi-precision during weight updating. Default is false.
        rescale_grad (float, optional): Multiply the gradient with `rescale_grad` before updating. \
            Often choose to be ``1.0/batch_size``.
77
        use_multi_tensor (bool, optional): Whether to use multi-tensor strategy to update all parameters at once . Default is false.
J
Jiawei Wang 已提交
78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97
        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` .

    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)
            beta1 = paddle.to_tensor([0.9], dtype="float32")
            beta2 = paddle.to_tensor([0.99], dtype="float32")
            momentum = paddle.optimizer.Momentum(learning_rate=0.1, parameters=linear.parameters(), weight_decay=0.01)
            back = out.backward()
            momentum.step()
            momentum.clear_grad()
98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115

            #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)
            momentum = paddle.optimizer.Momentum(
                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,
116
                momentum=0.9)
117 118 119 120
            out.backward()
            momentum.step()
            momentum.clear_grad()

J
Jiawei Wang 已提交
121 122 123
    """
    _velocity_acc_str = "velocity"

124 125 126 127 128 129 130 131 132 133 134 135 136
    def __init__(
        self,
        learning_rate=0.001,
        momentum=0.9,
        parameters=None,
        use_nesterov=False,
        weight_decay=None,
        grad_clip=None,
        multi_precision=False,
        rescale_grad=1.0,
        use_multi_tensor=False,
        name=None,
    ):
J
Jiawei Wang 已提交
137 138 139 140
        if learning_rate is None:
            raise ValueError("learning_rate is not set")
        if momentum is None:
            raise ValueError("momentum is not set")
141

142 143 144
        predicate = lambda regular: isinstance(
            regular, (L2DecayRegularizer, float)
        )
145 146 147
        if isinstance(parameters, list):
            if isinstance(parameters[0], dict):
                for param_group in parameters:
148 149 150 151 152
                    decay = (
                        param_group['weight_decay']
                        if 'weight_decay' in param_group
                        else weight_decay
                    )
153 154 155 156 157 158
                    reg_method, reg_coeff = self._update_regularization(decay)
                    param_group['regularization_method'] = reg_method
                    param_group['regularization_coeff'] = reg_coeff
                    py_regular = None if predicate(decay) else decay
                    param_group['weight_decay'] = py_regular

H
huangxu96 已提交
159
        py_regular = None if predicate(weight_decay) else weight_decay
160 161 162 163 164 165 166
        super(Momentum, self).__init__(
            learning_rate=learning_rate,
            parameters=parameters,
            weight_decay=py_regular,
            grad_clip=grad_clip,
            name=name,
        )
J
Jiawei Wang 已提交
167 168 169
        self.type = "momentum"
        self._momentum = momentum
        self._use_nesterov = bool(use_nesterov)
170 171 172 173
        (
            self._regularization_method,
            self._regularization_coeff,
        ) = self._update_regularization(weight_decay)
H
huangxu96 已提交
174 175 176 177
        self._multi_precision = multi_precision
        self._rescale_grad = rescale_grad
        self._master_weights = {}

178 179 180 181 182 183 184
        self._default_dict = {
            'momentum': momentum,
            'use_nesterov': use_nesterov,
            'rescale_grad': rescale_grad,
            'regularization_method': self._regularization_method,
            'regularization_coeff': self._regularization_coeff,
        }
185 186
        self._use_multi_tensor = use_multi_tensor
        if self._use_multi_tensor:
187 188 189 190 191 192
            self._param_dict = self._create_multi_tensor_dict()
            self._velocity_dict = self._create_multi_tensor_dict()
            self._master_weight_dict = self._create_multi_tensor_dict()
            self._master_weight_dict['FP32_LODTensor'] = None
            self._regularization_method_dict = self._create_multi_tensor_dict()
            self._regularization_coeff_dict = self._create_multi_tensor_dict()
193 194 195

    def _update_regularization(self, weight_decay):
        reg_method = ""
196
        reg_coeff = 0.0
197

198
        if isinstance(weight_decay, L2DecayRegularizer):
199 200
            reg_method = "l2_decay"
            reg_coeff = weight_decay._regularization_coeff
201
        if isinstance(weight_decay, float):
202 203 204
            reg_method = "l2_decay"
            reg_coeff = weight_decay
        return reg_method, reg_coeff
J
Jiawei Wang 已提交
205

H
huangxu96 已提交
206
    def _create_master_weight(self, param):
207 208 209 210 211 212 213
        if param.name in self._master_weights:
            var = self._master_weights[param.name]
        else:
            assert isinstance(self.helper, LayerHelper)

            var_name = param.name + "_fp32_master"
            var_name = unique_name.generate(var_name)
214 215 216 217 218 219 220
            var = layers.create_global_var(
                name=var_name,
                shape=param.shape,
                value=0,
                dtype='float32',
                persistable=True,
            )
221
            block = self.helper.startup_program.global_block()
222 223 224 225 226 227 228 229 230
            block.append_op(
                type="cast",
                inputs={"X": [param]},
                outputs={"Out": [var]},
                attrs={
                    "in_dtype": param.dtype,
                    "out_dtype": core.VarDesc.VarType.FP32,
                },
            )
231
            self._master_weights[param.name] = var
H
huangxu96 已提交
232 233 234 235 236 237 238 239 240 241 242 243 244 245
        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 variable for which accumulator is to be fetched

        Returns:
            accumulator variable for the parameter
        """
        if self._name is not None:
            name = self._name + "_" + name
246 247 248 249 250 251
        find_master = (
            self._multi_precision and param.dtype == core.VarDesc.VarType.FP16
        )
        target_param = (
            self._master_weights[param.name] if find_master else param
        )
H
huangxu96 已提交
252
        target_name = target_param.name
253 254 255 256
        if (
            name not in self._accumulators
            or target_name not in self._accumulators[name]
        ):
257 258
            raise Exception(
                "Accumulator {} does not exist for parameter {}".format(
259 260 261
                    name, target_name
                )
            )
H
huangxu96 已提交
262 263
        return self._accumulators[name][target_name]

J
Jiawei Wang 已提交
264
    def _create_accumulators(self, block, parameters):
265
        '''
J
Jiabin Yang 已提交
266
        if framework._non_static_mode():
267
            return
268
        '''
J
Jiawei Wang 已提交
269
        assert isinstance(block, framework.Block)
270 271 272 273

        if isinstance(parameters, dict):
            parameters = self._update_param_group(parameters)

274 275 276 277 278
        for p in parameters:
            if self._multi_precision and p.dtype == core.VarDesc.VarType.FP16:
                master_p = self._create_master_weight(p)
                self._add_accumulator(self._velocity_acc_str, master_p)
                continue
279 280 281 282
            if (
                p.dtype == core.VarDesc.VarType.FP16
                and not self._multi_precision
            ):
283 284 285 286 287
                warnings.warn(
                    "Accumulating with FP16 in optimizer can lead to poor accuracy or slow convergence."
                    "Consider using multi_precision=True option of the Momentum optimizer."
                )
            self._add_accumulator(self._velocity_acc_str, p)
J
Jiawei Wang 已提交
288

289
    def _create_regularization_of_grad(self, param, grad, regularization=None):
290
        """Create and add backward regularization Operators
291

292 293 294 295
        Function helper of append_regularization_ops.
        """
        # If ParamAttr is set to L2Decay, we skip doing regularization here. And then we fused
        # L2Decay with momentum which can refer to _append_optimize_op below.
296 297 298
        if hasattr(param, 'regularizer') and isinstance(
            param.regularizer, L2DecayRegularizer
        ):
299 300
            return grad
        return super(Momentum, self)._create_regularization_of_grad(
301 302
            param, grad, regularization
        )
303

J
Jiawei Wang 已提交
304 305
    def _append_optimize_op(self, block, param_and_grad):
        assert isinstance(block, framework.Block)
306 307
        if isinstance(param_and_grad, dict):
            param_and_grad = self._update_param_group(param_and_grad)
J
Jiawei Wang 已提交
308

309 310 311
        velocity_acc = self._get_accumulator(
            self._velocity_acc_str, param_and_grad[0]
        )
J
Jiawei Wang 已提交
312 313
        lr = self._create_param_lr(param_and_grad)

314
        # For fusion of momentum and l2decay
315 316 317 318 319 320 321 322 323 324 325
        param = param_and_grad[0]
        regularization_method = self._regularization_method
        regularization_coeff = self._regularization_coeff
        if hasattr(param, 'regularizer'):
            # we skip param's l2decay before, so fuse it with momentum here.
            if isinstance(param.regularizer, L2DecayRegularizer):
                regularization_method = "l2_decay"
                regularization_coeff = param.regularizer._regularization_coeff
            # the param's regularization has been done before, we avoid do l2decay in momentum.
            elif param.regularizer is not None:
                regularization_method = ""
326
                regularization_coeff = 0.0
327

328 329 330 331 332 333 334 335 336
        find_master = (
            self._multi_precision
            and param_and_grad[0].dtype == core.VarDesc.VarType.FP16
        )
        master_weight = (
            self._master_weights[param_and_grad[0].name]
            if find_master
            else None
        )
337

338
        if _in_legacy_dygraph():
339 340
            if isinstance(param_and_grad, dict):
                self._update_regularization(param_and_grad['weight_decay'])
341
            _, _, _ = _legacy_C_ops.momentum(
342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360
                param_and_grad[0],
                param_and_grad[1],
                velocity_acc,
                lr,
                master_weight,
                param_and_grad[0],
                velocity_acc,
                master_weight,
                'mu',
                self._momentum,
                'use_nesterov',
                self._use_nesterov,
                'regularization_method',
                regularization_method,
                'regularization_coeff',
                regularization_coeff,
                'multi_precision',
                find_master,
            )
361
            return None
362 363 364
        if in_dygraph_mode():
            if isinstance(param_and_grad, dict):
                self._update_regularization(param_and_grad['weight_decay'])
365 366 367 368 369 370 371 372 373 374 375 376 377
            return _C_ops.momentum_(
                param_and_grad[0],
                param_and_grad[1],
                velocity_acc,
                lr,
                master_weight,
                self._momentum,
                self._use_nesterov,
                regularization_method,
                regularization_coeff,
                find_master,
                self._rescale_grad,
            )
378

H
huangxu96 已提交
379 380 381
        attrs = {
            "mu": self._momentum,
            "use_nesterov": self._use_nesterov,
382 383
            "regularization_method": regularization_method,
            "regularization_coeff": regularization_coeff,
H
huangxu96 已提交
384
            "multi_precision": find_master,
385
            "rescale_grad": self._rescale_grad,
H
huangxu96 已提交
386 387
        }

J
Jiawei Wang 已提交
388 389 390 391
        inputs = {
            "Param": [param_and_grad[0]],
            "Grad": [param_and_grad[1]],
            "Velocity": [velocity_acc],
392
            "LearningRate": [lr],
J
Jiawei Wang 已提交
393 394 395 396
        }

        outputs = {
            "ParamOut": [param_and_grad[0]],
397
            "VelocityOut": [velocity_acc],
J
Jiawei Wang 已提交
398
        }
H
huangxu96 已提交
399 400 401 402 403

        if find_master:
            inputs["MasterParam"] = master_weight
            outputs["MasterParamOut"] = master_weight

J
Jiawei Wang 已提交
404
        # create the momentum optimize op
405 406 407 408 409 410 411
        momentum_op = block.append_op(
            type=self.type,
            inputs=inputs,
            outputs=outputs,
            attrs=attrs,
            stop_gradient=True,
        )
J
Jiawei Wang 已提交
412 413

        return momentum_op
414

415
    def _multi_tensor_init(self, target_block, parameters, param_group_idx):
416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432
        """
        All parameters used for optimizer (such as: parameters, master_weight, velocity_acc for momentum) calculations are grouped into a python list by data type (float16, float32).
        This function will be overridden in the corresponding optimizer file.

        Args:
            target_block: the block in which the loss tensor is present
            parameters: list of parameter tensors for the optimizer
        """
        self._create_accumulators(target_block, parameters)
        for param in parameters:
            velocity_acc = self._get_accumulator(self._velocity_acc_str, param)
            regularization_method = self._regularization_method
            regularization_coeff = self._regularization_coeff
            if hasattr(param, 'regularizer'):
                # we skip param's l2decay before, so fuse it with momentum here.
                if isinstance(param.regularizer, L2DecayRegularizer):
                    regularization_method = "l2_decay"
433 434 435
                    regularization_coeff = (
                        param.regularizer._regularization_coeff
                    )
436
                elif param.regularizer is not None:
437 438 439
                    regularization_method = ""
                    regularization_coeff = 0.0
            if param.dtype == paddle.float32:
440 441
                self._param_dict['FP32_LODTensor'][param_group_idx].append(
                    param
442
                )
443 444
                self._velocity_dict['FP32_LODTensor'][param_group_idx].append(
                    velocity_acc
445
                )
446 447 448 449 450 451 452
                # fp32 no master weight
                self._regularization_method_dict['FP32_LODTensor'][
                    param_group_idx
                ].append(regularization_method)
                self._regularization_coeff_dict['FP32_LODTensor'][
                    param_group_idx
                ].append(regularization_coeff)
453
            elif param.dtype == paddle.float16:
454 455
                self._param_dict['FP16_LODTensor'][param_group_idx].append(
                    param
456
                )
457 458
                self._velocity_dict['FP16_LODTensor'][param_group_idx].append(
                    velocity_acc
459
                )
460 461 462 463 464 465 466 467 468 469 470 471 472 473
                if self._multi_precision:
                    self._master_weight_dict['FP16_LODTensor'][
                        param_group_idx
                    ].append(self._master_weights[param.name])
                else:
                    self._master_weight_dict['FP16_LODTensor'][
                        param_group_idx
                    ] = None
                self._regularization_method_dict['FP16_LODTensor'][
                    param_group_idx
                ].append(regularization_method)
                self._regularization_coeff_dict['FP16_LODTensor'][
                    param_group_idx
                ].append(regularization_coeff)
474 475 476 477 478
            else:
                raise ValueError(
                    "Now multi_tensor_momentum only support fp32 and fp16 parameters and grad is LOD_TENSOR."
                )

479
    def _append_optimize_multi_tensor_op(
480 481 482 483
        self,
        target_block,
        parameters_and_grads,
        param_group_idx,
484
    ):
485
        """
486 487 488 489 490 491 492 493 494 495 496 497
        For Multi Tensor, append optimize merged_operator to block.
        """
        assert isinstance(target_block, framework.Block)

        grad_dict = {'FP32_LODTensor': [], 'FP16_LODTensor': []}
        lr_dict = {'FP32_LODTensor': [], 'FP16_LODTensor': []}

        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:
498 499 500 501 502
                    if (
                        param_and_grad[0].dtype == paddle.float32
                        and param_and_grad[1].type
                        == core.VarDesc.VarType.LOD_TENSOR
                    ):
503 504 505
                        grad_dict['FP32_LODTensor'].append(param_and_grad[1])
                        lr = self._create_param_lr(param_and_grad)
                        lr_dict['FP32_LODTensor'].append(lr)
506 507 508 509 510
                    elif (
                        param_and_grad[0].dtype == paddle.float16
                        and param_and_grad[1].type
                        == core.VarDesc.VarType.LOD_TENSOR
                    ):
511 512 513 514 515 516 517 518 519 520
                        grad_dict['FP16_LODTensor'].append(param_and_grad[1])
                        lr = self._create_param_lr(param_and_grad)
                        lr_dict['FP16_LODTensor'].append(lr)
        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
521 522 523 524 525 526 527
                    param_grad_dict.update(
                        {
                            k: v
                            for k, v in parameters_and_grads.items()
                            if k != 'params'
                        }
                    )
528
                    param_and_grad = self._update_param_group(param_grad_dict)
529 530 531 532 533
                    if (
                        param_and_grad[0].dtype == paddle.float32
                        and param_and_grad[1].type
                        == core.VarDesc.VarType.LOD_TENSOR
                    ):
534 535 536
                        grad_dict['FP32_LODTensor'].append(param_and_grad[1])
                        lr = self._create_param_lr(param_and_grad)
                        lr_dict['FP32_LODTensor'].append(lr)
537 538 539 540 541
                    elif (
                        param_and_grad[0].dtype == paddle.float16
                        and param_and_grad[1].type
                        == core.VarDesc.VarType.LOD_TENSOR
                    ):
542 543 544 545 546 547
                        grad_dict['FP16_LODTensor'].append(param_and_grad[1])
                        lr = self._create_param_lr(param_and_grad)
                        lr_dict['FP16_LODTensor'].append(lr)

        multi_tensor_list = ['FP32_LODTensor', 'FP16_LODTensor']
        for key in multi_tensor_list:
548
            if len(self._param_dict[key][param_group_idx]) > 0:
549
                find_master = self._multi_precision and key == 'FP16_LODTensor'
550

551 552 553 554 555 556 557
                master_weight = self._master_weight_dict[key]
                master_weight = (
                    master_weight[param_group_idx]
                    if master_weight is not None
                    else None
                )

J
Jiabin Yang 已提交
558
                if framework._non_static_mode():
559
                    if in_dygraph_mode():
560
                        _, _, _ = _C_ops.merged_momentum_(
561
                            self._param_dict[key][param_group_idx],
562
                            grad_dict[key],
563
                            self._velocity_dict[key][param_group_idx],
564
                            lr_dict[key],
565
                            master_weight,
566
                            self._momentum,
567
                            self._use_nesterov,
568 569 570 571 572 573
                            self._regularization_method_dict[key][
                                param_group_idx
                            ],
                            self._regularization_coeff_dict[key][
                                param_group_idx
                            ],
574 575 576
                            find_master,
                            self._rescale_grad,
                        )
577
                    else:
578
                        _, _, _ = _legacy_C_ops.merged_momentum(
579
                            self._param_dict[key][param_group_idx],
580
                            grad_dict[key],
581
                            self._velocity_dict[key][param_group_idx],
582
                            lr_dict[key],
583 584 585 586
                            master_weight,
                            self._param_dict[key][param_group_idx],
                            self._velocity_dict[key][param_group_idx],
                            master_weight,
587 588 589 590
                            'mu',
                            self._momentum,
                            'use_nesterov',
                            self._use_nesterov,
591
                            'regularization_method',
592 593 594
                            self._regularization_method_dict[key][
                                param_group_idx
                            ],
595
                            'regularization_coeff',
596 597 598
                            self._regularization_coeff_dict[key][
                                param_group_idx
                            ],
599 600 601
                            'multi_precision',
                            find_master,
                        )
602 603
                else:
                    inputs = {
604
                        "Param": self._param_dict[key][param_group_idx],
605
                        "Grad": grad_dict[key],
606
                        "Velocity": self._velocity_dict[key][param_group_idx],
607 608 609
                        "LearningRate": lr_dict[key],
                    }
                    outputs = {
610 611 612 613
                        "ParamOut": self._param_dict[key][param_group_idx],
                        "VelocityOut": self._velocity_dict[key][
                            param_group_idx
                        ],
614 615
                    }
                    attrs = {
616 617 618 619
                        "mu": self._momentum,
                        "use_nesterov": self._use_nesterov,
                        "regularization_method": self._regularization_method_dict[
                            key
620 621
                        ][
                            param_group_idx
622 623 624
                        ],
                        "regularization_coeff": self._regularization_coeff_dict[
                            key
625
                        ][param_group_idx],
626
                    }
627
                    if find_master:
628 629 630
                        inputs["MasterParam"] = self._master_weight_dict[key][
                            param_group_idx
                        ]
631
                        outputs["MasterParamOut"] = self._master_weight_dict[
632
                            key
633
                        ][param_group_idx]
634
                        attrs["multi_precision"] = find_master
635 636 637 638 639 640 641
                    target_block.append_op(
                        type="merged_momentum",
                        inputs=inputs,
                        outputs=outputs,
                        attrs=attrs,
                        stop_gradient=True,
                    )
642 643
        return None

644
    def _update_param_group(self, parameters):
645 646 647 648 649 650 651 652 653
        self._momentum = parameters.get(
            'momentum', self._default_dict['momentum']
        )
        self._use_nesterov = parameters.get(
            'use_nesterov', self._default_dict['use_nesterov']
        )
        self._rescale_grad = parameters.get(
            'rescale_grad', self._default_dict['rescale_grad']
        )
654
        self._regularization_method = parameters.get(
655 656
            'regularization_method', self._default_dict['regularization_method']
        )
657
        self._regularization_coeff = parameters.get(
658 659
            'regularization_coeff', self._default_dict['regularization_coeff']
        )
660 661
        parameters = parameters.get('params')
        return parameters