trained_betas ignored in some schedulers

[vishnu-anirudh]

Hello

I was going through the code repo and I observed how the self.betas is being assigned. I don't know if what I observed is correct, so I wanted to raise this PR to clarify this.

Without the changes made in this PR, it seems the following code is not useful

if trained_betas is not None:
    self.betas = np.asarray(trained_betas)

This is because self.betas is assigned again as part of the if-statement below. And if beta_schedule doesn't exist then an error is raised. So it doesn't matter if self.betas is assigned to trained_betas.

As a note, certain schedulers such as FlaxDDIMScheduler, DDPMScheduler and FlaxDDPMScheduler seem to have the right implementation.

If you think this is not right, I am happy to close this PR.

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[pcuenca]

This does look like a bug to me, when trained_betas is used instead of one of the beta schedules. I updated the title, I hope that's ok with you @vishnu-anirudh.

Requesting review from others in case I'm missing something.

[kashif]

@vishnu-anirudh nice catch!

[kashif]

Although the issue still remains if trained betas and schedule type is passed no?

[vishnu-anirudh]

Although the issue still remains if trained betas and schedule type is passed no?

From the code, it seems we choose trained_betas by default. Otherwise, we initialise self.betas based on beta_schedule.

[patrickvonplaten]

I think we are intertwining two things here:

• The warning message, which actually already has been downgraded to an info message here:

  diffusers/src/diffusers/configuration_utils.py

  Line 315 in c070e5f

  logger.info(

• This PR which is nice in itself!

This PR looks good to me :-)

[vishnu-anirudh]

Hello (cc @patrickvonplaten)
Can this be merged if there are no more changes (as I don't have the rights to do it).

Thanks in advance.