`DDIMScheduler` does not work unless `set_timesteps` is used

[samedii]

Describe the bug

Need to set number of inference steps since "None" is default. Expected to have a default value.

Adding a very simple PR.

Reproduction

    scheduler = DDIMScheduler(
        beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear"
    )
    # skipping this step will cause an exception later
    # scheduler.set_timesteps(1000)
    pipeline = StableDiffusionPipeline.from_pretrained(
        model_name,
        scheduler=scheduler,
        use_auth_token=True,
    ).to(device)

Logs

>       prev_timestep = timestep - self.config.num_train_timesteps // self.num_inference_steps
E       TypeError: unsupported operand type(s) for //: 'int' and 'NoneType'

System Info

diffusers v0.2.4
python 3.8.8

[pcuenca]

Sorry, I responding to your PR before I saw this issue. @anton-l what do you think?

[natolambert]

If I recall correctly all of the schedulers are not designed to run without calling set_timesteps() because of the heavy dependence of inference on num_inference_timesteps and how the timesteps match with the inference loop.

A quick double check showed DDPM and PNDM have the same behavior.

[patrickvonplaten]

@samedii, actually quick question: How to do you exactly get the above error?
The stable diffusion pipeline should always correctly set the number of timesteps, see:

diffusers/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py

Line 121 in efa773a

self.scheduler.set_timesteps(num_inference_steps, **extra_set_kwargs)

How exactly do you get the error:

>       prev_timestep = timestep - self.config.num_train_timesteps // self.num_inference_steps
E       TypeError: unsupported operand type(s) for //: 'int' and 'NoneType'

If you do:

    scheduler = DDIMScheduler(
        beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear"
    )
    # skipping this step will cause an exception later
    # scheduler.set_timesteps(1000)
    pipeline = StableDiffusionPipeline.from_pretrained(
        model_name,
        scheduler=scheduler,
        use_auth_token=True,
    ).to(device)
   pipeline("an image of a planet")

it should work correctly

[samedii]

@samedii, actually quick question: How to do you exactly get the above error?
...

Sorry I answered in the PR. This is not an issue for anyone using the default pipeline. I built my own pipeline using your components.