Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[BT] add BetterTransformer support for ViLT architecture #508

Merged
merged 5 commits into from
Nov 24, 2022
Merged

[BT] add BetterTransformer support for ViLT architecture #508

merged 5 commits into from
Nov 24, 2022

Conversation

ka00ri
Copy link
Contributor

@ka00ri ka00ri commented Nov 23, 2022 •
edited
Loading

What does this PR do?

Added BetterTransformer support for ViLT architecture.
Added test model for ViltModel.

Tested the conversion of ViltLayer:

from optimum.bettertransformer import BetterTransformer
model = BetterTransformer.transform("hf-internal-testing/tiny-random-ViltModel")
print(model)
ViltModel(
  (embeddings): ViltEmbeddings(
    (text_embeddings): TextEmbeddings(
      (word_embeddings): Embedding(1124, 32)
      (position_embeddings): Embedding(512, 32)
      (token_type_embeddings): Embedding(16, 32)
      (LayerNorm): LayerNorm((32,), eps=1e-12, elementwise_affine=True)
      (dropout): Dropout(p=0.1, inplace=False)
    )
    (patch_embeddings): ViltPatchEmbeddings(
      (projection): Conv2d(3, 32, kernel_size=(2, 2), stride=(2, 2))
    )
    (token_type_embeddings): Embedding(2, 32)
    (dropout): Dropout(p=0.1, inplace=False)
  )
  (encoder): ViltEncoder(
    (layer): ModuleList(
      (0): ViltLayerBetterTransformer()
      (1): ViltLayerBetterTransformer()
      (2): ViltLayerBetterTransformer()
      (3): ViltLayerBetterTransformer()
      (4): ViltLayerBetterTransformer()
    )
  )
  (layernorm): LayerNorm((32,), eps=1e-12, elementwise_affine=True)
  (pooler): ViltPooler(
    (dense): Linear(in_features=32, out_features=32, bias=True)
    (activation): Tanh()
  )
)

To: @younesbelkada @michaelbenayoun @fxmarty

This was referenced Nov 23, 2022
Closed
Open
younesbelkada
younesbelkada reviewed Nov 23, 2022
View reviewed changes
Copy link
Contributor

@younesbelkada younesbelkada left a comment •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Very nice! Thank you very much for adding BetterTransformer support for ViLT model!
I think we need 2 more additional steps before merging:
1- Could you add a new class on the testing suite since the model expected text + image. So inside tests_bettertransformer_vision.py you can just add something like:

ALL_VISION_TEXT_MODELS_TO_TEST = ("hf-internal-testing/tiny-random-ViltModel")

class BetterTransformersViLTTest(BetterTransformersTestMixin, unittest.TestCase):
    r"""
    Testing suite for Vision Models - tests all the tests defined in `BetterTransformersTestMixin`
    """
    all_models_to_test = ALL_VISION_TEXT_MODELS_TO_TEST

    def prepare_inputs_for_class(self, model_id=None):
        url = "http://images.cocodataset.org/val2017/000000039769.jpg"
        image = Image.open(requests.get(url, stream=True).raw)
        processor = AutoProcessor.from_pretrained("hf-internal-testing/tiny-random-ViltModel")
        .... 
        return inputs

And tweak the function prepare_inputs_for_class based on this example: https://huggingface.co/dandelin/vilt-b32-finetuned-vqa

2- Can you add the new architecture on the documentation? 🙏 this would require adding just a line here (don't forget to respect the alphabetical order ;) )

Also don't forget to run make style before pushing! ;)

Thanks a bunch!

@HuggingFaceDocBuilderDev
Copy link

HuggingFaceDocBuilderDev commented Nov 23, 2022 •
edited
Loading

The documentation is not available anymore as the PR was closed or merged.

@younesbelkada younesbelkada changed the title add BetterTransformer support for ViLT architecture [BT] add BetterTransformer support for ViLT architecture Nov 23, 2022
younesbelkada
younesbelkada approved these changes Nov 23, 2022
View reviewed changes
Copy link
Contributor

@younesbelkada younesbelkada left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Very clean and super fast implementation of ViLT support for BetterTransformer! Thanks also for taking care of slightly refactoring the documentation.
Thanks a lot for adding this, you really did a great job here! 💪
I have nothing to add on my side, gently pinging @fxmarty and @michaelbenayoun for a last review!

@ka00ri
Copy link
Contributor Author

ka00ri commented Nov 23, 2022

@younesbelkada Thank you for your support and making this first contribution so smooth 😄 Glad I could help 💪

michaelbenayoun
michaelbenayoun approved these changes Nov 24, 2022
View reviewed changes
Copy link
Member

@michaelbenayoun michaelbenayoun left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM!

@younesbelkada younesbelkada merged commit c95565d into huggingface:main Nov 24, 2022
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@younesbelkada younesbelkada younesbelkada approved these changes

@michaelbenayoun michaelbenayoun michaelbenayoun approved these changes

@fxmarty fxmarty Awaiting requested review from fxmarty

Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
4 participants
@ka00ri @HuggingFaceDocBuilderDev @michaelbenayoun @younesbelkada