[Tests] Fix SD slow tests #364
Conversation
|
The documentation is not available anymore as the PR was closed or merged. |
| @@ -917,7 +917,7 @@ def test_stable_diffusion_fast_ddim(self): | |||
| image_slice = image[0, -3:, -3:, -1] | |||
|
|
|||
| assert image.shape == (1, 512, 512, 3) | |||
| expected_slice = np.array([0.8354, 0.83, 0.866, 0.838, 0.8315, 0.867, 0.836, 0.8584, 0.869]) | |||
There was a problem hiding this comment.
Any idea why this was changed (maybe I need to double check here real quick with SD github)
There was a problem hiding this comment.
Could be all the way back from the generator's cpu->gpu move, so please check if you have the script nearby!
There was a problem hiding this comment.
Results on master are 1-to-1 the same as original CompVis repo still (checked with scripts from: #182) -> so change is good to go!
| pipe = StableDiffusionImg2ImgPipeline.from_pretrained(model_id, use_auth_token=True) | ||
| pipe = StableDiffusionImg2ImgPipeline.from_pretrained( | ||
| model_id, | ||
| revision="fp16", # fp16 to infer 768x512 images with 16GB of VRAM |
There was a problem hiding this comment.
think we can test with smaller images here, no ?
768x512 i would too big for tests.
There was a problem hiding this comment.
Will update these in another PR, gotta draw a small owl or smth 😂
There was a problem hiding this comment.
(also would like to rely less on the dataset indices for the reference images)
| pipe.to(torch_device) | ||
| pipe.set_progress_bar_config(disable=None) | ||
|
|
||
| prompt = "A fantasy landscape, trending on artstation" | ||
|
|
||
| generator = torch.Generator(device=torch_device).manual_seed(0) | ||
| output = pipe(prompt=prompt, init_image=init_image, strength=0.75, guidance_scale=7.5, generator=generator) | ||
| with torch.autocast("cuda"): |
There was a problem hiding this comment.
should we do device check here, before doing autocast ?
There was a problem hiding this comment.
But skipIf(torch_device == "cpu") should be ok for now, no?
| expected_array = np.array(output_image) / 255.0 | ||
| sampled_array = np.array(image) / 255.0 |
There was a problem hiding this comment.
Maybe we could set output_type=numpy in pipe instead of converting it to numpy here again
There was a problem hiding this comment.
Running into a couple of rounding errors when comparing numpy and a PIL-loaded png:
Expected :0.0001
Actual :0.0022058823529411686
TODO: check the DDIM output against the original repo