Difference b.w calling self.log(..., on_step=False, on_epoch=True) in training_step() and training_epoch_end()

[janghyuk-choi]

lightning-hydra-template/src/models/mnist_module.py

Lines 75 to 76 in 75b44ff

	self.log("train/loss", self.train_loss, on_step=False, on_epoch=True, prog_bar=True)
	self.log("train/acc", self.train_acc, on_step=False, on_epoch=True, prog_bar=True)

In the src/models/mnist_module.py, self.log() is called inside training_step() with on_step=False and on_epoch=True.

I wonder what the difference is compared to using self.log(..., on_step=False, on_epoch=True) inside training_epoch_end().

It seems that the resulting values in either way are the same.

Is there any difference in the internal mechanism?

[ashleve]

No difference as far as I'm aware, as long as we're logging torchmetrics object directly. Not sure if there won't be slight differences when we log through value though.

• Inside training_epoch_end() flags on_step=False, on_epoch=True don't matter as you're always logging only once per epoch.
• Inside training_step() setting on_epoch=True will make lightning average given value over logs from all steps.

[janghyuk-choi]

Thank you for your kind reply!

It seems like just a convention by PL.

And there has been a discussion on the purpose or function of that code (Lightning-AI/pytorch-lightning#3228).

[adosar]

No difference as far as I'm aware, as long as we're logging torchmetrics object directly. Not sure if there won't be slight differences when we log through value though.

• Inside training_epoch_end() flags on_step=False, on_epoch=True don't matter as you're always logging only once per epoch.
• Inside training_step() setting on_epoch=True will make lightning average given value over logs from all steps.

Is this behavior somewhere documented that "on_epoch=True" doesn't matter for "on_train_epoch_end"?