RLPack provides easy-to-use interface to integrate LR Schedulers in your training. All the LR Schedulers use PyTorch implementation under-the-hood. To use Learning Rate Schedulers easily, you can pass lr_scheduler_name: <keyword>
in the config dictionary. For example, if we set lr_scheduler_name: "step_lr"
we will select StepLR
scheduler in our training. To pass additional arguments to the desired scheduler, we must pass lr_scheduler_args
as a dictionary containing keyword arguments to the config dict. Further, you can limit the influence of LR Scheduler by setting lr_threshold
in agent_args
. Once this value is reached, the LR Scheduler is not called further.
LR Scheduler | Description | Keyword |
---|---|---|
StepLR | Step Learning Rate Scheduler. For mandatory arguments and further details, please refer here. | "step_lr" |
LinearLR | Linear Learning Rate Scheduler. For mandatory arguments and further details, please refer here. | "linear_lr" |
CyclicLR | Cyclic Learning Rate Scheduler. For mandatory arguments and further details, please refer here. | "cyclic_lr" |