In Efficient_ad, Dose the lr not need to be updated?

[rachelYY12]

def configure_optimizers(self) -> optim.Optimizer:
    optimizer = optim.Adam(
        list(self.model.student.parameters()) + list(self.model.ae.parameters()),
        lr=self.lr,
        weight_decay=self.weight_decay,
    )
    num_steps = min(
        self.trainer.max_steps, self.trainer.max_epochs * len(self.trainer.datamodule.train_dataloader())
    )
    scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=int(0.95 * num_steps), gamma=0.1)
    return {"optimizer": optimizer, "lr_scheduler": scheduler}

I find lr upgrade after one epoch in pytorch_lightning，but step_size=int(0.95 * num_steps) is larger than max_epochs，Does the lr not need to be updatated？

[blaz-r]

It looks like you are using an older version of the anomalib where this was not properly addressed, as the code for EfficientAD should look like this:

anomalib/src/anomalib/models/image/efficient_ad/lightning_model.py

Lines 217 to 231 in 14ab566

	# lightning stops training when either 'max_steps' or 'max_epochs' is reached (earliest),
	# so actual training steps need to be determined here
	if self.trainer.max_epochs < 0:
	# max_epochs not set
	num_steps = self.trainer.max_steps
	elif self.trainer.max_steps < 0:
	# max_steps not set -> determine steps as 'max_epochs' * 'steps in a single training epoch'
	num_steps = self.trainer.max_epochs * len(self.trainer.datamodule.train_dataloader())
	else:
	num_steps = min(
	self.trainer.max_steps,
	self.trainer.max_epochs * len(self.trainer.datamodule.train_dataloader()),
	)
	scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=int(0.95 * num_steps), gamma=0.1)