Source code for mindarmour.diff_privacy.train.model

# Copyright 2020 Huawei Technologies Co., Ltd
#
# 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.
"""
Differential privacy model.
"""
from easydict import EasyDict as edict

from mindspore.train.model import Model
from mindspore._checkparam import Validator as validator
from mindspore._checkparam import Rel
from mindspore.train import amp
from mindspore.train.amp import _config_level
from mindspore.common import dtype as mstype
from mindspore.nn.wrap.cell_wrapper import _VirtualDatasetCell
from mindspore.parallel._utils import _get_parallel_mode
from mindspore.train.model import ParallelMode
from mindspore.train.amp import _do_keep_batchnorm_fp32
from mindspore.train.amp import _add_loss_network
from mindspore import context
from mindspore import nn
from mindspore import Tensor
from mindspore.ops import composite as C
from mindspore.ops import operations as P
from mindspore.ops import functional as F
from mindspore.ops.operations import NPUGetFloatStatus
from mindspore.ops.operations import NPUAllocFloatStatus
from mindspore.ops.operations import NPUClearFloatStatus
from mindspore.ops.operations import ReduceSum
from mindspore.ops.operations import LessEqual
from mindspore.ops.operations import ControlDepend
from mindspore.parallel._utils import _get_mirror_mean
from mindspore.parallel._utils import _get_device_num
from mindspore.nn.wrap.grad_reducer import DistributedGradReducer
from mindspore.common.parameter import Parameter
from mindspore.nn.wrap.loss_scale import _grad_overflow
from mindspore.nn import Cell
from mindspore import ParameterTuple

from mindarmour.utils.logger import LogUtil
from mindarmour.diff_privacy.mechanisms.mechanisms import \
    _MechanismsParamsUpdater
from mindarmour.utils._check_param import check_value_positive, check_param_type
from mindarmour.utils._check_param import check_int_positive

LOGGER = LogUtil.get_instance()
TAG = 'DP model'

GRADIENT_CLIP_TYPE = 1
_grad_scale = C.MultitypeFuncGraph("grad_scale")
_reciprocal = P.Reciprocal()


@_grad_scale.register("Tensor", "Tensor")
def tensor_grad_scale(scale, grad):
    """ grad scaling """
    return grad*F.cast(_reciprocal(scale), F.dtype(grad))


[docs]class DPModel(Model): """ This class is overload mindspore.train.model.Model. Args: micro_batches (int): The number of small batches split from an original batch. Default: 2. norm_bound (float): Use to clip the bound, if set 1, will return the original data. Default: 1.0. noise_mech (Mechanisms): The object can generate the different type of noise. Default: None. clip_mech (Mechanisms): The object is used to update the adaptive clip. Default: None. Raises: ValueError: If DPOptimizer and noise_mecn are both None or not None. ValueError: If noise_mech or DPOtimizer's mech method is adaptive while clip_mech is not None. Examples: >>> norm_bound = 1.0 >>> initial_noise_multiplier = 0.01 >>> network = LeNet5() >>> batch_size = 32 >>> batches = 128 >>> epochs = 1 >>> micro_batches = 2 >>> loss = nn.SoftmaxCrossEntropyWithLogits(is_grad=False, sparse=True) >>> factory_opt = DPOptimizerClassFactory(micro_batches=micro_batches) >>> factory_opt.set_mechanisms('Gaussian', >>> norm_bound=norm_bound, >>> initial_noise_multiplier=initial_noise_multiplier) >>> net_opt = factory_opt.create('Momentum')(network.trainable_params(), >>> learning_rate=0.1, momentum=0.9) >>> clip_mech = ClipMechanismsFactory().create('Gaussian', >>> decay_policy='Linear', >>> learning_rate=0.01, >>> target_unclipped_quantile=0.9, >>> fraction_stddev=0.01) >>> model = DPModel(micro_batches=micro_batches, >>> norm_bound=norm_bound, >>> clip_mech=clip_mech, >>> noise_mech=None, >>> network=network, >>> loss_fn=loss, >>> optimizer=net_opt, >>> metrics=None) >>> ms_ds = ds.GeneratorDataset(dataset_generator(batch_size, batches), >>> ['data', 'label']) >>> ms_ds.set_dataset_size(batch_size*batches) >>> model.train(epochs, ms_ds, dataset_sink_mode=False) """ def __init__(self, micro_batches=2, norm_bound=1.0, noise_mech=None, clip_mech=None, **kwargs): if micro_batches: self._micro_batches = check_int_positive('micro_batches', micro_batches) else: self._micro_batches = None norm_bound = check_param_type('norm_bound', norm_bound, float) norm_bound = check_value_positive('norm_bound', norm_bound) norm_bound = Tensor(norm_bound, mstype.float32) self._norm_bound = Parameter(norm_bound, 'norm_bound') opt = kwargs['optimizer'] opt_name = opt.__class__.__name__ # Check whether noise_mech and DPOptimizer are both None or not None, if so, raise ValueError. # And check whether noise_mech or DPOtimizer's mech method is adaptive while clip_mech is not None, # if so, rasie ValuerError too. if noise_mech is not None and "DPOptimizer" in opt_name: msg = 'DPOptimizer is not supported while noise_mech is not None' LOGGER.error(TAG, msg) raise ValueError(msg) if noise_mech is None: if "DPOptimizer" in opt_name: if context.get_context('mode') != context.PYNATIVE_MODE: msg = 'DPOptimizer just support pynative mode currently.' LOGGER.error(TAG, msg) raise ValueError(msg) if 'Ada' in opt._mech.__class__.__name__ and clip_mech is not None: msg = "When DPOptimizer's mech method is adaptive, clip_mech must be None." LOGGER.error(TAG, msg) raise ValueError(msg) else: msg = 'DPModel should set noise_mech or DPOptimizer configure, ' \ 'please refer to example.' LOGGER.error(TAG, msg) raise ValueError(msg) self._noise_mech = noise_mech if noise_mech is not None: if 'Ada' in noise_mech.__class__.__name__ and clip_mech is not None: msg = 'When noise_mech is Adaptive, clip_mech must be None.' LOGGER.error(TAG, msg) raise ValueError(msg) if clip_mech is None or isinstance(clip_mech, Cell): self._clip_mech = clip_mech super(DPModel, self).__init__(**kwargs) def _amp_build_train_network(self, network, optimizer, loss_fn=None, level='O0', **kwargs): """ Build the mixed precision training cell automatically. Args: network (Cell): Definition of the network. loss_fn (Union[None, Cell]): Definition of the loss_fn. If None, the `network` should have the loss inside. Default: None. optimizer (Optimizer): Optimizer to update the Parameter. level (str): Supports [O0, O2]. Default: "O0". - O0: Do not change. - O2: Cast network to float16, keep batchnorm and `loss_fn` (if set) run in float32, using dynamic loss scale. cast_model_type (:class:`mindspore.dtype`): Supports `mstype.float16` or `mstype.float32`. If set to `mstype.float16`, use `float16` mode to train. If set, overwrite the level setting. keep_batchnorm_fp32 (bool): Keep Batchnorm run in `float32`. If set, overwrite the level setting. loss_scale_manager (Union[None, LossScaleManager]): If None, not scale the loss, or else scale the loss by LossScaleManager. If set, overwrite the level setting. """ validator.check_value_type('network', network, nn.Cell, None) validator.check_value_type('optimizer', optimizer, nn.Optimizer, None) validator.check('level', level, "", ['O0', 'O2'], Rel.IN, None) self._check_kwargs(kwargs) config = dict(_config_level[level], **kwargs) config = edict(config) if config.cast_model_type == mstype.float16: network.to_float(mstype.float16) if config.keep_batchnorm_fp32: _do_keep_batchnorm_fp32(network) if loss_fn: network = _add_loss_network(network, loss_fn, config.cast_model_type) if _get_parallel_mode() in ( ParallelMode.SEMI_AUTO_PARALLEL, ParallelMode.AUTO_PARALLEL): network = _VirtualDatasetCell(network) loss_scale = 1.0 if config.loss_scale_manager is not None: loss_scale_manager = config.loss_scale_manager loss_scale = loss_scale_manager.get_loss_scale() update_cell = loss_scale_manager.get_update_cell() if update_cell is not None: # only cpu not support `TrainOneStepWithLossScaleCell` for control flow. if not context.get_context("enable_ge") and context.get_context( "device_target") == "CPU": msg = "Only `loss_scale_manager=None` and " \ "`loss_scale_manager=FixedLossScaleManager(drop_overflow" \ "_update=False)` are supported in current version. " \ "If you use `O2` option, please use " \ "`loss_scale_manager=None` or `FixedLossScaleManager`" LOGGER.error(TAG, msg) raise ValueError(msg) network = _TrainOneStepWithLossScaleCell(network, optimizer, scale_update_cell=update_cell, micro_batches=self._micro_batches, norm_bound=self._norm_bound, clip_mech=self._clip_mech, noise_mech=self._noise_mech).set_train() return network network = _TrainOneStepCell(network, optimizer, self._norm_bound, loss_scale, micro_batches=self._micro_batches, clip_mech=self._clip_mech, noise_mech=self._noise_mech).set_train() return network def _build_train_network(self): """Build train network""" network = self._network if self._micro_batches: if self._optimizer: if self._loss_scale_manager_set: network = self._amp_build_train_network(network, self._optimizer, self._loss_fn, level=self._amp_level, loss_scale_manager=self._loss_scale_manager, keep_batchnorm_fp32=self._keep_bn_fp32) else: network = self._amp_build_train_network(network, self._optimizer, self._loss_fn, level=self._amp_level, keep_batchnorm_fp32=self._keep_bn_fp32) elif self._loss_fn: network = nn.WithLossCell(network, self._loss_fn) else: if self._optimizer: if self._loss_scale_manager_set: network = amp.build_train_network(network, self._optimizer, self._loss_fn, level=self._amp_level, loss_scale_manager=self._loss_scale_manager, keep_batchnorm_fp32=self._keep_bn_fp32) else: network = amp.build_train_network(network, self._optimizer, self._loss_fn, level=self._amp_level, keep_batchnorm_fp32=self._keep_bn_fp32) elif self._loss_fn: network = nn.WithLossCell(network, self._loss_fn) if self._parallel_mode in (ParallelMode.SEMI_AUTO_PARALLEL, ParallelMode.AUTO_PARALLEL): network.set_auto_parallel() return network
class _ClipGradients(nn.Cell): """ Clip gradients. Inputs: grads (tuple[Tensor]): Gradients. clip_type (int): The way to clip, 0 for 'value', 1 for 'norm'. clip_value (float): Specifies how much to clip. Outputs: tuple[Tensor], clipped gradients. """ def __init__(self): super(_ClipGradients, self).__init__() self.clip_by_norm = nn.ClipByNorm() self.dtype = P.DType() def construct(self, grads, clip_type, clip_value): """ construct a compute flow. """ # pylint: disable=consider-using-in if clip_type != 0 and clip_type != 1: return grads new_grads = () for grad in grads: if clip_type == 0: norm = C.clip_by_value(grad, -clip_value, clip_value) else: norm = self.clip_by_norm(grad, clip_value) new_grads = new_grads + (norm,) return new_grads class _TupleAdd(nn.Cell): def __init__(self): super(_TupleAdd, self).__init__() self.add = P.TensorAdd() self.hyper_map = C.HyperMap() def construct(self, input1, input2): """Add two tuple of data.""" out = self.hyper_map(self.add, input1, input2) return out class _TrainOneStepWithLossScaleCell(Cell): r""" Network training with loss scaling. This is a training step with loss scaling. It takes a network, an optimizer and possibly a scale update Cell as args. The loss scale value can be updated in both host side or device side. The TrainOneStepWithLossScaleCell will be compiled to be graph which takes `data`, `label`, `sens` as input data. The `sens` is acting as loss scaling value. If you want to update it on host side, the value should be provided. If `sens` is not given, the loss scale update logic should be provied by `scale_update_cell`. If `scale_update_cell` is not None and `sens` is provided, the `scale_update_cell` will be ignored. Args: network (Cell): The training network. optimizer (Cell): Optimizer for updating the weights. scale_update_cell(Cell): The loss scaling update logic cell. Default: None. micro_batches (int): The number of small batches split from an original batch. Default: None. norm_bound (Tensor): Use to clip the bound, if set 1, will return the original data. Default: 1.0. noise_mech (Mechanisms): The object can generate the different type of noise. Default: None. Inputs: - **inputs** (Tensor) - Tensor of shape :math:`(N, \ldots)`. - **label** (Tensor) - Tensor of shape :math:`(N, \ldots)`. - **scaling_sens** (Tensor) - Tensor of shape :math:`()`. Outputs: Tuple of 3 Tensor, the loss, overflow flag and current loss scaling value. - **loss** (Tensor) - Tensor with shape :math:`()`. - **overflow** (Tensor) - Tensor with shape :math:`()`, type is bool. - **loss_scale** (Tensor) - Tensor with shape :math:`()`. """ def __init__(self, network, optimizer, scale_update_cell=None, micro_batches=None, norm_bound=1.0, noise_mech=None, clip_mech=None): super(_TrainOneStepWithLossScaleCell, self).__init__(auto_prefix=False) self.network = network self.network.set_grad() self.network.add_flags(defer_inline=True) self.weights = ParameterTuple(network.trainable_params()) self.optimizer = optimizer self.grad = C.GradOperation('grad', get_by_list=True, sens_param=True) self.hyper_map = C.HyperMap() if context.get_context("device_target") == "GPU": self.gpu_target = True self.float_status = P.FloatStatus() self.addn = P.AddN() self.reshape = P.Reshape() else: self.gpu_target = False self.alloc_status = NPUAllocFloatStatus() self.get_status = NPUGetFloatStatus() self.clear_status = NPUClearFloatStatus() self.reduce_sum = ReduceSum(keep_dims=False) self.base = Tensor(1, mstype.float32) self.less_equal = LessEqual() self.depend_parameter_use = ControlDepend(depend_mode=1) self.allreduce = P.AllReduce() self.parallel_mode = _get_parallel_mode() self.grad_reducer = F.identity self.reducer_flag = self.parallel_mode in [ParallelMode.DATA_PARALLEL, ParallelMode.HYBRID_PARALLEL] if self.reducer_flag: mean = _get_mirror_mean() degree = _get_device_num() self.grad_reducer = DistributedGradReducer(optimizer.parameters, mean, degree) self.is_distributed = self.parallel_mode != ParallelMode.STAND_ALONE self.loss_scale = None self.loss_scaling_manager = scale_update_cell if scale_update_cell: self.loss_scale = Parameter(Tensor(scale_update_cell.get_loss_scale(), dtype=mstype.float32), name="loss_scale") self.add_flags(has_effect=True) # dp params self._micro_batches = micro_batches self._norm_bound = norm_bound self._split = P.Split(0, self._micro_batches) self._clip_by_global_norm = _ClipGradients() self._noise_mech = noise_mech self._clip_mech = clip_mech self._add = P.TensorAdd() self._norm = nn.Norm() self._tuple_add = _TupleAdd() self._hyper_map = C.HyperMap() self._micro_float = Tensor(micro_batches, mstype.float32) self._zero = Tensor(0, mstype.float32) self._assign = P.Assign() self._div = P.Div() self._sqrt = P.Sqrt() self._reduce_sum = P.ReduceSum() self._square_all = P.Square() self._less = P.Less() self._cast = P.Cast() self._noise_mech_param_updater = None if self._noise_mech is not None and self._noise_mech._decay_policy is not None: self._noise_mech_param_updater = _MechanismsParamsUpdater( decay_policy=self._noise_mech._decay_policy, decay_rate=self._noise_mech._noise_decay_rate, cur_noise_multiplier= self._noise_mech._noise_multiplier, init_noise_multiplier= self._noise_mech._initial_noise_multiplier) def construct(self, data, label, sens=None): """ construct a compute flow. """ init = False if not self.gpu_target: # init overflow buffer init = self.alloc_status() # clear overflow buffer self.clear_status(init) if sens is None: scaling_sens = self.loss_scale else: scaling_sens = sens # DP clip weights = self.weights record_datas = self._split(data) record_labels = self._split(label) # first index loss = self.network(record_datas[0], record_labels[0]) scaling_sens_filled = C.ones_like(loss)*F.cast(scaling_sens, F.dtype(loss)) record_grad = self.grad(self.network, weights)(record_datas[0], record_labels[0], scaling_sens_filled) beta = self._zero square_sum = self._zero for grad in record_grad: square_sum = self._add(square_sum, self._reduce_sum(self._square_all(grad))) norm_grad = self._sqrt(square_sum) beta = self._add(beta, self._cast(self._less(norm_grad, self._norm_bound), mstype.float32)) record_grad = self._clip_by_global_norm(record_grad, GRADIENT_CLIP_TYPE, self._norm_bound) grads = record_grad total_loss = loss for i in range(1, self._micro_batches): loss = self.network(record_datas[i], record_labels[i]) scaling_sens_filled = C.ones_like(loss)*F.cast(scaling_sens, F.dtype(loss)) record_grad = self.grad(self.network, weights)(record_datas[i], record_labels[i], scaling_sens_filled) square_sum = self._zero for grad in record_grad: square_sum = self._add(square_sum, self._reduce_sum(self._square_all(grad))) norm_grad = self._sqrt(square_sum) beta = self._add(beta, self._cast(self._less(norm_grad, self._norm_bound), mstype.float32)) record_grad = self._clip_by_global_norm(record_grad, GRADIENT_CLIP_TYPE, self._norm_bound) grads = self._tuple_add(grads, record_grad) total_loss = P.TensorAdd()(total_loss, loss) loss = P.Div()(total_loss, self._micro_float) beta = self._div(beta, self._micro_batches) if self._noise_mech is not None: grad_noise_tuple = () for grad_item in grads: grad_noise = self._mech(grad_item) grad_noise_tuple = grad_noise_tuple + (grad_noise,) grads = self._tuple_add(grads, grad_noise_tuple) grads = self._hyper_map(F.partial(_grad_scale, self._micro_float), grads) # update mech parameters if self._noise_mech_param_updater is not None: multiplier = self._noise_mech_param_updater() loss = F.depend(loss, multiplier) grads = self.hyper_map(F.partial(_grad_scale, scaling_sens), grads) # apply grad reducer on grads grads = self.grad_reducer(grads) # get the overflow buffer if not self.gpu_target: self.get_status(init) # sum overflow buffer elements, 0:not overflow , >0:overflow flag_sum = self.reduce_sum(init, (0,)) else: flag_sum = self.hyper_map(F.partial(_grad_overflow), grads) flag_sum = self.addn(flag_sum) # convert flag_sum to scalar flag_sum = self.reshape(flag_sum, (())) if self.is_distributed: # sum overflow flag over devices flag_reduce = self.allreduce(flag_sum) cond = self.less_equal(self.base, flag_reduce) else: cond = self.less_equal(self.base, flag_sum) overflow = cond if sens is None: overflow = self.loss_scaling_manager(self.loss_scale, cond) # if there is no overflow, do optimize if overflow: opt = False else: opt = self.optimizer(grads) ret = (loss, cond, scaling_sens) if self._clip_mech is not None: next_norm_bound = self._clip_mech(beta, self._norm_bound) P.assign(self._norm_bound, next_norm_bound) return F.depend(ret, opt) class _TrainOneStepCell(Cell): r""" Network training package class. Wraps the network with an optimizer. The resulting Cell be trained with input data and label. Backward graph will be created in the construct function to do parameter updating. Different parallel modes are available to run the training. Args: network (Cell): The training network. optimizer (Cell): Optimizer for updating the weights. sens (Number): The scaling number to be filled as the input of back propagation. Default value is 1.0. micro_batches (int): The number of small batches split from an original batch. Default: None. norm_bound (Tensor): Use to clip the bound, if set 1, will return the original data. Default: 1.0. noise_mech (Mechanisms): The object can generate the different type of noise. Default: None. Inputs: - **data** (Tensor) - Tensor of shape :math:`(N, \ldots)`. - **label** (Tensor) - Tensor of shape :math:`(N, \ldots)`. Outputs: Tensor, a scalar Tensor with shape :math:`()`. """ def __init__(self, network, optimizer, norm_bound=1.0, sens=1.0, micro_batches=None, noise_mech=None, clip_mech=None): super(_TrainOneStepCell, self).__init__(auto_prefix=False) self.network = network self.network.set_grad() self.network.add_flags(defer_inline=True) self.weights = optimizer.parameters self.optimizer = optimizer self.grad = C.GradOperation('grad', get_by_list=True, sens_param=True) self.sens = sens self.reducer_flag = False self.grad_reducer = None parallel_mode = _get_parallel_mode() if parallel_mode in ( ParallelMode.DATA_PARALLEL, ParallelMode.HYBRID_PARALLEL): self.reducer_flag = True if self.reducer_flag: mean = _get_mirror_mean() degree = _get_device_num() self.grad_reducer = DistributedGradReducer(optimizer.parameters, mean, degree) # dp params if micro_batches is None: msg = 'micro_batches must give in differential privacy, but got value: {}'.format( micro_batches) LOGGER.error(TAG, msg) raise ValueError(msg) self._micro_batches = micro_batches self._norm_bound = norm_bound self._split = P.Split(0, self._micro_batches) self._clip_by_global_norm = _ClipGradients() self._noise_mech = noise_mech self._clip_mech = clip_mech self._tuple_add = _TupleAdd() self._add = P.TensorAdd() self._norm = nn.Norm() self._hyper_map = C.HyperMap() self._zero = Tensor(0, mstype.float32) self._assign = P.Assign() self._div = P.Div() self._sqrt = P.Sqrt() self._reduce_sum = P.ReduceSum() self._square_all = P.Square() self._less = P.Less() self._cast = P.Cast() self._micro_float = Tensor(micro_batches, mstype.float32) self._noise_mech_param_updater = None if self._noise_mech is not None and self._noise_mech._decay_policy is not None: self._noise_mech_param_updater = _MechanismsParamsUpdater( decay_policy=self._noise_mech._decay_policy, decay_rate=self._noise_mech._noise_decay_rate, cur_noise_multiplier= self._noise_mech._noise_multiplier, init_noise_multiplier= self._noise_mech._initial_noise_multiplier) def construct(self, data, label): """ construct a compute flow. """ weights = self.weights record_datas = self._split(data) record_labels = self._split(label) loss = self.network(record_datas[0], record_labels[0]) sens = P.Fill()(P.DType()(loss), P.Shape()(loss), self.sens) record_grad = self.grad(self.network, weights)(record_datas[0], record_labels[0], sens) beta = self._zero # calcu beta if self._clip_mech is not None: square_sum = self._zero for grad in record_grad: square_sum = self._add(square_sum, self._reduce_sum(self._square_all(grad))) norm_grad = self._sqrt(square_sum) beta = self._add(beta, self._cast(self._less(norm_grad, self._norm_bound), mstype.float32)) record_grad = self._clip_by_global_norm(record_grad, GRADIENT_CLIP_TYPE, self._norm_bound) grads = record_grad total_loss = loss for i in range(1, self._micro_batches): loss = self.network(record_datas[i], record_labels[i]) sens = P.Fill()(P.DType()(loss), P.Shape()(loss), self.sens) record_grad = self.grad(self.network, weights)(record_datas[i], record_labels[i], sens) # calcu beta if self._clip_mech is not None: square_sum = self._zero for grad in record_grad: square_sum = self._add(square_sum, self._reduce_sum(self._square_all(grad))) norm_grad = self._sqrt(square_sum) beta = self._add(beta, self._cast(self._less(norm_grad, self._norm_bound), mstype.float32)) record_grad = self._clip_by_global_norm(record_grad, GRADIENT_CLIP_TYPE, self._norm_bound) grads = self._tuple_add(grads, record_grad) total_loss = P.TensorAdd()(total_loss, loss) loss = self._div(total_loss, self._micro_float) if self._noise_mech is not None: grad_noise_tuple = () for grad_item in grads: grad_noise = self._noise_mech(grad_item) grad_noise_tuple = grad_noise_tuple + (grad_noise,) grads = self._tuple_add(grads, grad_noise_tuple) grads = self._hyper_map(F.partial(_grad_scale, self._micro_float), grads) # update mech parameters if self._noise_mech_param_updater is not None: multiplier = self._noise_mech_param_updater() loss = F.depend(loss, multiplier) if self.reducer_flag: # apply grad reducer on grads grads = self.grad_reducer(grads) if self._clip_mech is not None: beta = self._div(beta, self._micro_batches) next_norm_bound = self._clip_mech(beta, self._norm_bound) self._norm_bound = self._assign(self._norm_bound, next_norm_bound) loss = F.depend(loss, self._norm_bound) return F.depend(loss, self.optimizer(grads))