Integration tests precision improvement for inpainting #1052
Conversation
get tests passing with greater precision using lewington images
|
The documentation is not available anymore as the PR was closed or merged. |
|
@patrickvonplaten I'm submitting this PR a little early, as I am not sure if anyone generally checks dataset PRs on huggingface, and I wanted to get your attention since at the moment you're still assigned to the issue. |
|
Hey @Lewington-pitsos this is a great PR! I'll accepted the change on https://huggingface.co/datasets/hf-internal-testing/diffusers-images/discussions/2 - thanks a mille for adding the datasets directly there. Also, sadly we cannot guarantee a precision of 1e-4 across devices, the best we can do seems to be 1e-3 as pointed out here: |
|
@Lewington-pitsos, if you're interesting we still need to change the integration tests for text2image and img2img from PIL images to numpy just like you've done here. Leaving #914 open for this |
|
On it like a bonnet! |
|
Very cool - it's merged :-) |
) * improve test precision get tests passing with greater precision using lewington images * make old numpy load function a wrapper around a more flexible numpy loading function * adhere to black formatting * add more black formatting * adhere to isort * loosen precision and replace path Co-authored-by: Patrick von Platen <[email protected]>
3 participants
closes #914
Currently all 11 related tests pass with
pytest tests\pipelines\stable_diffusion\test_stable_diffusion_inpaint.py tests\pipelines\stable_diffusion\test_stable_diffusion_inpaint_legacy.py -k test_stable_diffusion_inpainthowever they use my own huggingface dataset.Waiting on this dataset PR to be merged before we can update the URLs and merge this one.
Once this is merged I intend to apply the same strategy to improve other tests.