Fix BaseOutput initialization from dict #570
Conversation
|
The documentation is not available anymore as the PR was closed or merged. |
|
Now I see the need for this code 😅 . We need to adapt it though for But essentially, the code as is on this branch will currently break the following: from diffusers import DDIMPipeline
pipeline = DDIMPipeline.from_pretrained("google/ddpm-cifar10-32")
outputs = pipeline()
outputs["images"]@anton-l could you see how this can be fixed and then maybe also add a test here :-) |
| # check every way of getting the attribute | ||
| assert isinstance(outputs.images, np.ndarray) | ||
| assert outputs.images.shape == (1, 3, 4, 4) | ||
| assert isinstance(outputs["images"], np.ndarray) |
There was a problem hiding this comment.
The transformers version (initial commit from this PR) was failing on this line
cc @patrickvonplaten
| outputs = CustomOutput({"images": np.random.rand(1, 3, 4, 4)}) | ||
|
|
||
| # check every way of getting the attribute | ||
| assert isinstance(outputs.images, np.ndarray) |
There was a problem hiding this comment.
the pre-PR diffusers version was failing on this line
|
@patrickvonplaten I've implemented a simpler version that works for your case as well. |
|
|
||
|
|
||
| class ConfigTester(unittest.TestCase): | ||
| def test_outputs_single_attribute(self): |
| outputs = CustomOutput(images=[PIL.Image.new("RGB", (4, 4))]) | ||
|
|
||
| # check every way of getting the attribute | ||
| assert isinstance(outputs.images, list) |
There was a problem hiding this comment.
Cool that this is all tested now!
patrickvonplaten
left a comment
patrickvonplaten
left a comment
There was a problem hiding this comment.
Great! Thanks a lot for fixing it!
* Fix BaseOutput initialization from dict * style * Simplify post-init, add tests * remove debug
This will allow
accelerateto reinitialize model outputs like so:output = UNet2DOutput({"sample": torch.randn(3, 4)})and avoid nesting attributes.Closes #560