Deprecate predict_epsilon
#1393
Conversation
|
The documentation is not available anymore as the PR was closed or merged. |
| prediction_type: str = "epsilon", | ||
| **kwargs, | ||
| ): | ||
| message = ( | ||
| "Please make sure to instantiate your scheduler with `prediction_type` instead. E.g. `scheduler =" | ||
| " DDPMScheduler.from_pretrained(<model_id>, prediction_type='epsilon')`." | ||
| ) | ||
| predict_epsilon = deprecate("predict_epsilon", "0.10.0", message, take_from=kwargs) | ||
| if predict_epsilon is not None: | ||
| prediction_type = "epsilon" if predict_epsilon else "sample" |
There was a problem hiding this comment.
This doesn't work with from_pretrained:
ddpm = DDPMScheduler.from_pretrained(
"hf-internal-testing/tiny-stable-diffusion-torch",
subfolder="scheduler",
predict_epsilon=False,
)
assert ddpm.prediction_type == "sample" # AssertionErrorThe reason is that from_config extracts predict_epsilon as one unused_kwargs: https://github.com/huggingface/diffusers/blob/main/src/diffusers/configuration_utils.py#L195
How could we handle this case?
There was a problem hiding this comment.
Good catch! Let's correct this in a follow up PR ok?
schedulers using from_pretrained.
| if predict_epsilon is not None: | ||
| new_config = dict(self.scheduler.config) | ||
| new_config["predict_epsilon"] = predict_epsilon | ||
| new_config["prediction_type"] = "epsilon" if predict_epsilon else "sample" |
| model_output = self.unet(image, t).sample | ||
|
|
||
| # 2. compute previous image: x_t -> x_t-1 | ||
| image = self.scheduler.step( |
There was a problem hiding this comment.
good catch we don't need to pass it anymore!
src/diffusers/schedulers/scheduling_dpmsolver_multistep_flax.py
Outdated
Show resolved
Hide resolved
| if predict_epsilon is not None: | ||
| prediction_type = "epsilon" if predict_epsilon else "sample" | ||
|
|
||
| self.prediction_type = prediction_type |
There was a problem hiding this comment.
Here we need to modify the frozen dict as shown here:
There was a problem hiding this comment.
@patrickvonplaten I don't think there's an internal dict yet in the __init__ function. I wrote this test to verify that it works as expected: 01df570#diff-bf88051e4d8be9ab9d5a5f24a0daac59c898ddb1fca8505728f49b3809cd3666R611-R620
There was a problem hiding this comment.
Ah yeah you're right good point! Fixing this on main as well
Co-authored-by: Patrick von Platen <[email protected]>
| def test_inference_deprecated_predict_epsilon(self): | ||
| deprecate("remove this test", "0.10.0", "remove") | ||
| unet = self.dummy_uncond_unet | ||
| scheduler = DDPMScheduler(predict_epsilon=False) |
There was a problem hiding this comment.
This test is now failing because DDPMScheduler is initialized with the deprecated kwarg and then it's replaced with the default prediction_type in register_to_config. @patrickvonplaten should we address this in a followup PR?
|
Cool fixing the failing test now on "main" |
| if "dtype" in unused_kwargs: | ||
| init_dict["dtype"] = unused_kwargs.pop("dtype") | ||
|
|
||
| if "predict_epsilon" in unused_kwargs and "prediction_type" not in init_dict: |
There was a problem hiding this comment.
let's see if we can find a more general solution here
* Adapt ddpm, ddpmsolver to prediction_type. * Deprecate predict_epsilon in __init__. * Bring FlaxDDIMScheduler up to date with DDIMScheduler. * Set prediction_type as an ivar for consistency. * Convert pipeline_ddpm * Adapt tests. * Adapt unconditional training script. * Adapt BitDiffusion example. * Add missing kwargs in dpmsolver_multistep * Ugly workaround to accept deprecated predict_epsilon when loading schedulers using from_pretrained. * make style * Remove import no longer in use. * Apply suggestions from code review Co-authored-by: Patrick von Platen <[email protected]> * Use config.prediction_type everywhere * Add a couple of Flax prediction type tests. * make style * fix register deprecated arg Co-authored-by: Patrick von Platen <[email protected]>
* Adapt ddpm, ddpmsolver to prediction_type. * Deprecate predict_epsilon in __init__. * Bring FlaxDDIMScheduler up to date with DDIMScheduler. * Set prediction_type as an ivar for consistency. * Convert pipeline_ddpm * Adapt tests. * Adapt unconditional training script. * Adapt BitDiffusion example. * Add missing kwargs in dpmsolver_multistep * Ugly workaround to accept deprecated predict_epsilon when loading schedulers using from_pretrained. * make style * Remove import no longer in use. * Apply suggestions from code review Co-authored-by: Patrick von Platen <[email protected]> * Use config.prediction_type everywhere * Add a couple of Flax prediction type tests. * make style * fix register deprecated arg Co-authored-by: Patrick von Platen <[email protected]>
No description provided.