Source code for mindspore.nn.optim.lars

# Copyright 2020-2022 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.
# ============================================================================
"""lars optimizer"""
from __future__ import absolute_import

from mindspore.ops import operations as P
from mindspore.ops import composite as C
from mindspore.ops import functional as F
from mindspore._checkparam import Validator as validator
from mindspore.common import Tensor, Parameter, dtype as mstype
from mindspore.common.api import jit
from mindspore.nn.optim.optimizer import _grad_scale, Optimizer
from mindspore.nn.optim.optimizer import opt_init_args_register

_lars_opt = C.MultitypeFuncGraph("lars_opt")


@_lars_opt.register("Function", "Number", "Tensor", "Tensor", "Tensor", "Tensor", "Bool", "Bool")
def _tensor_run_opt(lars, loss_scale, learning_rate, weight_decay, gradient, weight, decay_flag, lars_flag):
    """Apply lars optimizer to the weight parameter."""
    if lars_flag:
        op_reduce_sum = P.SquareSumAll()
        w_square_sum, grad_square_sum = op_reduce_sum(weight, gradient)
        if decay_flag:
            grad_t = lars(weight, gradient, w_square_sum, grad_square_sum, weight_decay / loss_scale, learning_rate)
        else:
            num_zero = 0.0
            grad_t = lars(weight, gradient, w_square_sum, grad_square_sum, num_zero, learning_rate)
        return grad_t

    return gradient


def _check_param_value(optimizer, epsilon, coefficient, use_clip, prim_name):
    validator.check_value_type("optimizer", optimizer, Optimizer, prim_name)
    validator.check_value_type("epsilon", epsilon, [float], prim_name)
    validator.check_value_type("coefficient", coefficient, [float], prim_name)
    validator.check_value_type("use_clip", use_clip, [bool], prim_name)


[docs]class LARS(Optimizer): r""" Implements the LARS algorithm. LARS is an optimization algorithm employing a large batch optimization technique. Refer to paper `LARGE BATCH TRAINING OF CONVOLUTIONAL NETWORKS <https://arxiv.org/abs/1708.03888>`_. The updating formulas are as follows, .. math:: \begin{array}{ll} \\ &\newline &\hline \\ &\textbf{Parameters}: \text{base learning rate } \gamma_{0} , \text{ momentum m}, \text{ weight decay } \lambda , \\ &\hspace{5mm}\text{ LARS coefficient } \eta , \text{ number of steps } T \\ &\textbf{Init}: \text{ t=0, v=0, init weight } w_{0}^{l} \text{ for each layer } l \\[-1.ex] &\newline &\hline \\ &\textbf{while} \text{ t<T for each layer } l \textbf{ do} \\ &\hspace{5mm}g_{t}^{l} \leftarrow \nabla L\left(w_{t}^{l}\right) \\ &\hspace{5mm}\gamma_{t} \leftarrow \gamma_{0} *\left(1-\frac{t}{T}\right)^{2} \\ &\hspace{5mm}\gamma^{l} \leftarrow \eta *\frac{\left\|w_{t}^{l}\right\|}{\left\|g_{t}^{l}\right\|+ \lambda\left\|w_{t}^{l}\right\|} \text{(compute the local LR } \gamma^{ l)} \\ &\hspace{5mm}v_{t+1}^{l} \leftarrow m v_{t}^{l}+\gamma_{t+1} * \gamma^{l} *\left(g_{t}^{l}+\lambda w_{t}^{l}\right) \\ &\hspace{5mm}w_{t+1}^{l} \leftarrow w_{t}^{l}-v_{t+1}^{l} \\ &\textbf{ end while } \\[-1.ex] &\newline &\hline \\[-1.ex] \end{array} :math:`w` represents the network parameters, :math:`g` represents `gradients`, :math:`t` represents the current step, :math:`\delta` represents `weight_decay` in `optimizer`, :math:`\gamma` represents `learning_rate` in `optimizer`, :math:`\eta` represents `coefficient`. Args: optimizer (Optimizer): MindSpore optimizer for which to wrap and modify gradients. epsilon (float): Term added to the denominator to improve numerical stability. Default: 1e-05. coefficient (float): Trust coefficient for calculating the local learning rate. Default: 0.001. use_clip (bool): Whether to use clip operation for calculating the local learning rate. Default: False. lars_filter (Function): A function to determine which of the network parameters to use LARS algorithm. Default: lambda x: 'LayerNorm' not in x.name and 'bias' not in x.name. Inputs: - **gradients** (tuple[Tensor]) - The gradients of `params` in the optimizer, the shape is the as same as the `params` in the optimizer. Outputs: Union[Tensor[bool], tuple[Parameter]], it depends on the output of `optimizer`. Supported Platforms: ``Ascend`` Examples: >>> import mindspore as ms >>> from mindspore import nn >>> >>> net = Net() >>> loss = nn.SoftmaxCrossEntropyWithLogits() >>> opt = nn.Momentum(net.trainable_params(), 0.1, 0.9) >>> opt_lars = nn.LARS(opt, epsilon=1e-08, coefficient=0.02) >>> model = ms.Model(net, loss_fn=loss, optimizer=opt_lars, metrics=None) """ @opt_init_args_register def __init__(self, optimizer, epsilon=1e-05, coefficient=0.001, use_clip=False, lars_filter=lambda x: 'LayerNorm' not in x.name and 'bias' not in x.name): super(LARS, self).__init__(0.0, [Parameter(Tensor(0.0), name="fake_param")]) _check_param_value(optimizer, epsilon, coefficient, use_clip, self.cls_name) self.opt = optimizer self.dynamic_decay_flags = optimizer.dynamic_decay_flags self.dynamic_weight_decay = optimizer.dynamic_weight_decay self.weight_decay = optimizer.weight_decay self.global_step = optimizer.global_step self.parameters = optimizer.parameters if optimizer._use_flattened_params: # pylint: disable=W0212 self.opt._use_flattened_params = False # pylint: disable=W0212 self._user_parameters += [param.name for param in self.parameters] self.use_clip = use_clip self.lars_flag = tuple(lars_filter(x) for x in self.parameters) self.is_group = optimizer.is_group self.learning_rate = Parameter(Tensor(0.0, dtype=mstype.float32), name="fake_lr") self.decay_flags = optimizer.decay_flags self.reciprocal_scale = optimizer.reciprocal_scale self.need_scale = optimizer.need_scale self.lars = P.LARSUpdate(epsilon, coefficient, use_clip) self.cast = P.Cast() self.loss_scale = optimizer.loss_scale if use_clip: self.is_group_lr = optimizer.is_group_lr self.dynamic_lr = optimizer.dynamic_lr self.origin_learning_rate = optimizer.learning_rate if self.is_group_lr and self.dynamic_lr: raise ValueError("For 'LARS', if the argument 'use_clip' is set to True, then the dynamic " "learning rate and group learning rate cannot both be true.") if self.is_group: optimizer.dynamic_decay_flags = tuple(map(lambda x: False, self.dynamic_decay_flags)) else: optimizer.dynamic_decay_flags = False optimizer.decay_flags = tuple(map(lambda x: False, self.decay_flags)) optimizer.dynamic_weight_decay = False optimizer.reciprocal_scale = 1.0 optimizer.exec_weight_decay = False def _get_lr(self): """Get the learning rate of current step.""" lr = self.origin_learning_rate if self.dynamic_lr: if self.is_group_lr: lr = () for learning_rate in self.origin_learning_rate: current_dynamic_lr = learning_rate(self.global_step) lr += (current_dynamic_lr,) else: lr = self.origin_learning_rate(self.global_step) return lr @jit def construct(self, gradients): params = self.parameters gradients = self.flatten_gradients(gradients) if self.use_clip: lr = self._get_lr() else: lr = self.learning_rate weight_decay = self.get_weight_decay() if self.need_scale: gradients = self.hyper_map(F.partial(_grad_scale, self.reciprocal_scale), gradients) if self.is_group: if self.is_group_lr: gradients = self.hyper_map(F.partial(_lars_opt, self.lars, self.loss_scale), lr, weight_decay, gradients, params, self.decay_flags, self.lars_flag) else: gradients = self.hyper_map(F.partial(_lars_opt, self.lars, self.loss_scale, lr), weight_decay, gradients, params, self.decay_flags, self.lars_flag) else: gradients = self.hyper_map(F.partial(_lars_opt, self.lars, self.loss_scale, lr, weight_decay), gradients, params, self.decay_flags, self.lars_flag) success = self.opt(gradients) if self._is_dynamic_lr_or_weight_decay() and not self.opt.dynamic_lr and not self.opt.dynamic_weight_decay: self.assignadd(self.global_step, self.global_step_increase_tensor) return success