`nexuml_library.training.schedulers.warmup_cosine`¶

nexuml_library.training.schedulers.warmup_cosine ¶

Learning-rate schedulers for NexuML library scenarios.

Bases: LRScheduler

Linear warmup followed by cosine decay to an absolute LR floor.

Parameters:

Name	Type	Description	Default
`optimizer`	`Optimizer`	Optimizer whose parameter groups will be scheduled.	required
`warmup_epochs`	`int`	Number of initial epochs spent linearly increasing from `min_lr` to `base_lr`. Use `0` to disable warmup.	required
`max_epochs`	`int`	Total scheduler horizon. At `epoch == max_epochs` the LR is `min_lr`.	required
`min_lr`	`float`	Absolute minimum learning rate, not a multiplier.	required
`last_epoch`	`int`	Last scheduler epoch, forwarded to PyTorch's scheduler base.	`-1`

get_lr() -> list[float | torch.Tensor]

Return scheduled learning rates for the current epoch.