[BT] add BetterTransformer support for ViLT architecture

[ka00ri]

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

[younesbelkada]

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]

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

[younesbelkada]

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]

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

[michaelbenayoun]

LGTM!