issue with DDIM scheduler for google/ddpm-celebahq-256 example

[murphyk]

Describe the bug

Referring to the example "Unconditional Diffusion with discrete scheduler" at https://github.com/huggingface/diffusers#running-code
The comment says you can use DDIMPipeline but this fails to run on
(got an unexpected keyword argument eta).
Also, the PNDMPipeline runs but just returns noise.

Reproduction

# !pip install diffusers["torch"]
from diffusers import DDPMPipeline, DDIMPipeline, PNDMPipeline

model_id = "google/ddpm-celebahq-256"
device = "cuda"

# load model and scheduler

#ddpm = DDPMPipeline.from_pretrained(model_id)
ddpm = DDIMPipeline.from_pretrained(model_id)
#ddpm = PNDMPipeline.from_pretrained(model_id)

# you can replace DDPMPipeline with DDIMPipeline or PNDMPipeline for faster inference
ddpm.to(device)

# run pipeline in inference (sample random noise and denoise)
image = ddpm().images[0]
image

Logs

---------------------------------------------------------------------------
TypeError Traceback (most recent call last)
Input In [21], in <cell line: 17>()
 14 ddpm.to(device)
 16 # run pipeline in inference (sample random noise and denoise)
---> 17 image = ddpm().images[0]
 18 image

File /usr/local/lib/python3.9/dist-packages/torch/autograd/grad_mode.py:27, in _DecoratorContextManager.__call__.<locals>.decorate_context(*args, **kwargs)
 24 @functools.wraps(func)
 25 def decorate_context(*args, **kwargs):
 26 with self.clone():
---> 27 return func(*args, **kwargs)

File /usr/local/lib/python3.9/dist-packages/diffusers/pipelines/ddim/pipeline_ddim.py:114, in DDIMPipeline.__call__(self, batch_size, generator, eta, num_inference_steps, use_clipped_model_output, output_type, return_dict, **kwargs)
 109 model_output = self.unet(image, t).sample
 111 # 2. predict previous mean of image x_t-1 and add variance depending on eta
 112 # eta corresponds to η in paper and should be between [0, 1]
 113 # do x_t -> x_t-1
--> 114 image = self.scheduler.step(
 115 model_output, t, image, eta=eta, use_clipped_model_output=use_clipped_model_output, generator=generator
 116 ).prev_sample
 118 image = (image / 2 + 0.5).clamp(0, 1)
 119 image = image.cpu().permute(0, 2, 3, 1).numpy()

File /usr/local/lib/python3.9/dist-packages/diffusers/schedulers/scheduling_ddpm.py:259, in DDPMScheduler.step(self, model_output, timestep, sample, generator, return_dict, **kwargs)
 237 """
 238 Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion
 239 process from the learned model outputs (most often the predicted noise).
 (...)
 253 
 254 """
 255 message = (
 256 "Please make sure to instantiate your scheduler with `prediction_type` instead. E.g. `scheduler ="
 257 " DDPMScheduler.from_pretrained(<model_id>, prediction_type='epsilon')`."
 258 )
--> 259 predict_epsilon = deprecate("predict_epsilon", "0.11.0", message, take_from=kwargs)
 260 if predict_epsilon is not None:
 261 new_config = dict(self.config)

File /usr/local/lib/python3.9/dist-packages/diffusers/utils/deprecation_utils.py:43, in deprecate(take_from, standard_warn, *args)
 41 function = call_frame.function
 42 key, value = next(iter(deprecated_kwargs.items()))
---> 43 raise TypeError(f"{function} in {filename} line {line_number-1} got an unexpected keyword argument `{key}`")
 45 if len(values) == 0:
 46 return

TypeError: step in /usr/local/lib/python3.9/dist-packages/diffusers/schedulers/scheduling_ddpm.py line 258 got an unexpected keyword argument `eta`

### System Info



- `diffusers` version: 0.10.2
- Platform: Linux-5.4.0-122-generic-x86_64-with-glibc2.31
- Python version: 3.9.13
- PyTorch version (GPU?): 1.12.0+cu116 (True)
- Huggingface_hub version: 0.11.1
- Transformers version: 4.20.1
- Using GPU in script?: Yes. P6000
- Using distributed or parallel set-up in script?: No

[willdalh]

The unexpected keyword eta error will be fixed in #1932.

[patrickvonplaten]

Hey @willdalh,

Can we close this one or is there still an issue?

[willdalh]

Hi @patrickvonplaten 😄

@murphyk also mentions in this issue that the PNDMPipeline returns noise. This is because the pipeline does not assume that the PNDMScheduler should be used. I have opened a pull request, #2003, that fixes this.