Fix DETA save_pretrained

[qubvel]

What does this PR do?

save_pretrained is not working for DetaForObjectDetection. This PR is supposed to fix this error.

A simple script to reproduce:

from transformers import DetaConfig, DetaForObjectDetection

config = DetaConfig()
model = DetaForObjectDetection(config)
model.save_pretrained("output_dir/")

or

from transformers import AutoModelForObjectDetection

model = AutoModelForObjectDetection.from_pretrained("jozhang97/deta-resnet-50")
model.save_pretrained("output_dir/")

The following error occurs:

RuntimeError: The weights trying to be saved contained shared tensors [{'model.decoder.class_embed.0.weight', 'class_embed.0.weight'}, {'model.decoder.class_embed.0.bias', 'class_embed.0.bias'}, {'model.decoder.class_embed.1.weight', 'class_embed.1.weight'}, {'model.decoder.class_embed.1.bias', 'class_embed.1.bias'}, {'model.decoder.class_embed.2.weight', 'class_embed.2.weight'}, {'model.decoder.class_embed.2.bias', 'class_embed.2.bias'}, {'model.decoder.class_embed.3.weight', 'class_embed.3.weight'}, {'class_embed.3.bias', 'model.decoder.class_embed.3.bias'}, {'model.decoder.class_embed.4.weight', 'class_embed.4.weight'}, {'model.decoder.class_embed.4.bias', 'class_embed.4.bias'}, {'model.decoder.class_embed.5.weight', 'class_embed.5.weight'}, {'class_embed.5.bias', 'model.decoder.class_embed.5.bias'}, {'class_embed.6.weight', 'model.decoder.class_embed.6.weight'}, {'class_embed.6.bias', 'model.decoder.class_embed.6.bias'}] that are mismatching the transformers base configuration. Try saving using `safe_serialization=False` or remove this tensor sharing.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[qubvel]

cc @NielsRogge

[NielsRogge]

Thanks for fixing!

[amyeroberts]

Thanks for fixing and adding a test!

Just a small comment on removing the exception

[amyeroberts]

We shouldn't really be raising exceptions in a test. Exceptions are for terminating out of code when there's an incorrect input / code behaviour which we can then choose to handle. In tests, we're really performing sanity checks, which should always be True given the test.

Either we should test this behaviour with an assert or just remove.

[qubvel]

Thanks, Amy, I fixed this test case and the original one too 👍 (replaced with assert)

[amyeroberts]

Thanks for fixing and iterating!