Change the scheduler interval in CLI

[ForJadeForest]

I just trying to use StepLR in my .yaml config.

lr_scheduler:
  class_path: StepLR
  init_args:
    step_size : 10
    gamma : 0.5

I guess the CLI will set the interval as "epoch", but I want to use the "step" to control my learning rate. I do not find anything about it in the docs. How can achieve it just in my .yaml file?

[rohitgr7]

cc: @carmocca

[carmocca]

This is not possible from the yaml file. If you need to customize the learning rate scheduler configuration, you can do so by overriding:

class MyLightningCLI(LightningCLI):
    @staticmethod
    def configure_optimizers(lightning_module, optimizer, lr_scheduler=None):
        if lr_scheduler is None:
            return optimizer
        return {
            "optimizer": optimizer,
            "lr_scheduler": {"scheduler": lr_scheduler, "interval": "step"},
        }

[ForJadeForest]

Thanks for your reply. Do you mean that I can change the interval in the configure_optimizers method? But if using it, can I change the optimizer from the yaml file?

[carmocca]

Yes, but notice this is the configure_optimizers method in the LightningCLI that will replace the method in the LightningModule. You can still change the optimizer instance itself from the config file. It will be passed as the optimizer argument in the snippet above.

[ForJadeForest]

That's what I need. Thanks a lot!