Fix TFViT

[NielsRogge]

What does this PR do?

Fix the code example of TFViTForImageClassification + fix the integration test.

cc @ydshieh

[sgugger]

Looks good, thanks for uploading the checkpoint!

[ydshieh]

LGTM, thanks!

[ydshieh]

I open an issue #14403 for the following.

---

Hi, I found that there is something wrong (not introduced in this PR).

model.save_pretrained("tmp", saved_model=True) will give an error AttributeError: 'ViTConfig' object has no attribute 'items'.
(I found this when running test_saved_model_creation_extended in TF ViT test script)

Maybe it's better to open a new issue for this.

Update
Not specific to TFViT, even occurs for TFBert

from transformers import TFBertModel
model = TFBertModel.from_pretrained("bert-base-uncased")

# failed
for x in model.config.items():
    print(x)

# failed
model.save_pretrained("tmp", saved_model=True)

This probably suggests that test_saved_model_creation_extended (@tooslow) in test_modeling_tf_common.py will fail for all TF models.

Full example:

from transformers import TFViTForImageClassification
model = TFViTForImageClassification.from_pretrained('google/vit-base-patch16-224')
model.save_pretrained("tmp", saved_model=True)