DDPM changes to support v diffusion #1121
Conversation
|
The documentation is not available anymore as the PR was closed or merged. |
| generator: Optional[torch.Generator] = None, | ||
| output_type: Optional[str] = "pil", | ||
| return_dict: bool = True, | ||
| prediction_type: Literal["epsilon", "sample", "v"] = "epsilon", |
There was a problem hiding this comment.
This should be retrieved from the config of the scheduler / model IMO - I don't think this is something that the user would ever change during runtime.
| model_output: torch.FloatTensor, | ||
| timestep: int, | ||
| sample: torch.FloatTensor, | ||
| prediction_type: str = "epsilon", |
There was a problem hiding this comment.
Actually I think we should deprecate this argument and just have it in the config instead - I don't it's something that differs for every scheduler call. @anton-l what do you think?
There was a problem hiding this comment.
That makes sense! That means the pipeline doesn’t have to change at all either
There was a problem hiding this comment.
I agree, the prediction type doesn't need to change during sampling
|
By the way, here is the butterflies example samples trained on this implementation. I think it looks good compared to the original example, but wanted to put it out there for others to evaluate.
|
5 participants
Trying to keep this a small PR that just adds necessary changes in DDPM scheduler to support v diffusion. I have a supporting example script using the butterflies dataset I can add either here or as a separate PR