Fix squared optimization steps bug

[twerkmeister]

Currently, when increasing the number of epochs, the number of steps per epoch is also increased, leading to a number of optimization steps thats roughly (examples / batch size) * epochs * epochs instead of (examples / batch size) * epochs. The proposed change should fix this.

[tomaarsen]

Hello!

This bug was discovered #268 (comment), and a fix was proposed. It involves simply removing this line:

setfit/src/setfit/trainer.py

Line 378 in f777c2c

steps_per_epoch=train_steps,

When steps_per_epoch is not specified, the number of steps per epoch will be equivalent to the number of datapoints in the dataloader. See the docs on the fit method for more info.

The remainder of the code would not need any changes.
Would you agree that this is a preferable solution over the changes that you have proposed? I think it's best if we make a new PR for this.

• Tom Aarsen

[twerkmeister]

Hey @tomaarsen, I removed the argument as you mentioned. Beyond that, this version also fixes the fact that for the triplet losses self.num_epochs was used to calculate total number of optimization steps for reporting when in fact num_epochs should be used because self.num_epochs can be shadowed by an argument to the train method

[tomaarsen]

You're very right! Thanks for catching that. I'll merge this once the tests go green 🎉

[tomaarsen]

Suggested change

	warmup_steps = math.ceil(train_steps * self.warmup_proportion)
	warmup_steps = math.ceil(train_steps * self.warmup_proportion)

This still uses the train_steps variable that you removed.

[tomaarsen]

Wonderful! Thank you for this fix, and for the edits on the PR @twerkmeister!