espnet3.components.multiple_scheduler.MultipleScheduler

source

class espnet3.components.multiple_scheduler.MultipleScheduler(multiple_optimizer: MultipleOptim, lr_scheduler: LRScheduler, optimizer_idx: int)

Bases: _LRScheduler

Wrap a scheduler so Lightning accepts the MultipleOptim wrapper.

Lightning expects a scheduler to expose an optimizer attribute. This shim forwards everything to the wrapped scheduler while reporting the MultipleOptim instance. It allows one scheduler per optimizer inside MultipleOptim without triggering Lightning’s per-optimizer training_step logic.

• Parameters:
  • multiple_optimizer (MultipleOptim) – The wrapper that owns the underlying optimizers.
  • lr_scheduler (torch.optim.lr_scheduler.LRScheduler) – Scheduler configured for a single optimizer inside multiple_optimizer.
  • optimizer_idx (int) – Index of the optimizer the scheduler controls.

NOTE

Only minimal delegation is implemented. Any attribute not explicitly intercepted is passed through to the wrapped scheduler.

Example

>>> ms = MultipleScheduler(multi_optim, StepLR(opt_a, step_size=10), 0)
>>> lr = ms.get_last_lr()