basecls.solver.optimizer.lars#

LARS optimizer

References: rwightman/pytorch-image-models

class basecls.solver.optimizer.lars.LARS(params, lr, momentum=0.0, nesterov=False, weight_decay=0.0, always_adapt=False)[源代码]#

基类：Optimizer

Implements LARS algorithm.

LARS is proposed in “Large Batch Optimization for Deep Learning: Training BERT in 76 minutes”.

参数

  • params (Union[Iterable[Parameter], dict]) – iterable of parameters to optimize or dicts defining parameter groups.

  • lr (float) – learning rate.

  • momentum (float) – momentum factor. Default: 0.0

  • nesterov (bool) – enables Nesterov momentum. Default: False

  • weight_decay (float) – weight decay (L2 penalty). Default: 0.0

  • always_adapt (bool) – apply adaptive lr to 0.0 weight decay parameter. Default: False