Unused parameters in scheduler `init`'s

[natolambert]

What API design would you like to have changed or added to the library? Why?

I see some of the schedulers had these parameters added to init

        trained_betas: Optional[np.ndarray] = None,
        timestep_values: Optional[np.ndarray] = None,

I'm guessing the first is an option to directly pass betas when creating the scheduler, but what is timestep_values for? I didn't see any uses.

I would like to regularize these (e.g. seems like all DDPM -like schedulers can use trained_betas pretty easily)

Can easily address this while completing #376.

[patrickvonplaten]

Is this issues still relevant? Think you are right indeed and we can remove such parameters

[hafiidz]

To add, I keep getting this warning/log with default example txt2img and img2img pipe. {'trained_betas'} was not found in config. Values will be initialized to default values.

[natolambert]

I think I have updated all the unused parameters throughout documentation changes.

RE @Hafiidz 's point, I think that is because the config on the hub is missing the entry? I'm not 100% sure. Future model versions will likely address this.

[patrickvonplaten]

@Hafiidz,

That's good feedback, I think we should relax the warning to a logger.info statement to make for better user experience. @anton-l do you have a spare minute to take a look here?

[patrickvonplaten]

@anton-l @patil-suraj thoughts on this issue? Should we remove those parameters (timestep_values, trained_betas) for now?

[patil-suraj]

@patrickvonplaten Yes, those were added for BDDM which we don't have now, so I think it's fine to remove them now.

[patrickvonplaten]

@anton-l could you take a look?

[anton-l]

Not sure about removing trained_betas, because they are functioning as designed, and someone might be using them already. Gently deprecating them is an option, but only if we don't have plans for pipelines that need them. @patil-suraj is BDDM on our roadmap?

[patil-suraj]

Have we seen any example using trained_betas? If not then I think should be fine to remove it.

BDDM would be nice to have, but I don't have much bandwidth for next 2 weeks to work on it. Maybe @kashif after you are done with notes2audio :) Would be happy to help

[natolambert]

What's BDDM?

[kashif]

https://github.com/tencent-ailab/bddm speech model

[github-actions[bot]]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.