BaseImageProcessor #26
Conversation
|
The documentation is not available anymore as the PR was closed or merged. |
amyeroberts
changed the base branch from
base-image-processor-class
to
image-processor-mixin
Co-authored-by: Sylvain Gugger <Sylvain.gugger@gmail.com> Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
| data (`dict`): | ||
| Dictionary of lists/arrays/tensors returned by the __call__/pad methods ('pixel_values', 'attention_mask', | ||
| etc.). | ||
| tensor_type (`Union[None, str, TensorType]`, *optional*): |
There was a problem hiding this comment.
A quick question - are we going to default to "np" ? If so, maybe we can remove None from the accepted argument types or make it a non-optional argument
There was a problem hiding this comment.
It depends partly on whether we merge in: huggingface#18499
Defaulting to "np" isn't necessary to be able to use different combinations of e.g. do_resize and do_normalize. As we're aliasing the previous feature extractors with the new image processors, change the default would still be a breaking change.
If we decided to default to "np", then we'd have to include additional checks on the processed images. At the moment, because resize resizes the images to multiples of size_divisor, they are not guaranteed to all be the same size. This means calls the BatchFeature will fail if any of "np", "tf", "pt"or"jax"` are passed in as the images can't be batched together.
My preference would be to keep return_tensors=None as this more closely matches the behaviour of our tokenizers. However, our tokenizer provides arguments such that batches can be created e.g. padding=True. Not sure if an equivalent makes sense here.
What do you think? If we want to set "np" as default we should discuss how to handle introducing the image processors versus introducing that change.
cc @NielsRogge
Co-authored-by: Alara Dirik <8944735+alaradirik@users.noreply.github.com>
Image processor glpn
What does this PR do?
Introduces the
BaseImageProcessorclass for all other model image processors to inherit from.Before submitting
Pull Request section?
to it if that's the case.
documentation guidelines, and
here are tips on formatting docstrings.