Allow gradient calculation in DiffusionPipeline.__call__

[JoaoLages]

Nowadays the gradient calculation is disabled when calling the DiffusionPipeline object (example in here).
IMO, this makes the framework less flexible to work with some use cases. For example, this framework had to copy the whole __call__ method just to be able to get the gradients attached to the output image tensor.

[anton-l]

Hi @JoaoLages! I feel like torch.no_grad is a good default for pipelines, as they're intended only for inference by design. Additional functionality would require implementing a custom pipeline, indeed.

For your use case, however, there's a neat python trick to unwrap the nograd-decorated method and keep its contents: pipeline.__call__ = pipeline.__call__.__wrapped__. Could you check if it works for you?

[JoaoLages]

Hi @JoaoLages! I feel like torch.no_grad is a good default for pipelines, as they're intended only for inference by design. Additional functionality would require implementing a custom pipeline, indeed.

For your use case, however, there's a neat python trick to unwrap the nograd-decorated method and keep its contents: pipeline.__call__ = pipeline.__call__.__wrapped__. Could you check if it works for you?

I tried that in an early stage but it is not possible to override __call__ afaik

[github-actions]

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.