Add visualbert onnx suport

[tedasdf]

What does this PR do?

Fixes #555 (issue)

Support VisualBERT model config.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

Who can review?

[echarlaix]

Thanks for your contribution @tedasdf !

[echarlaix]

should be

Suggested change

	"visualbert": supported_tasks_mapping(
	"visual-bert": supported_tasks_mapping(

https://huggingface.co/hf-internal-testing/tiny-random-VisualBertModel/blob/main/config.json#L26

[tedasdf]

wouldnt that be visual_bert

[echarlaix]

model_type is defined as follows

optimum/optimum/exporters/tasks.py

Line 1318 in 01110ad

model_type = model_type.lower().replace("_", "-")

so here "visual-bert"

[tedasdf]

Hi @echarlaix, I wanted to thank you for reviewing the pull request. Frankly, I’m struggling a bit. I understand the files that need to be altered and the function to be implemented, but I’m having trouble finding the information related to the model, especially given that I’m not very familiar with such a large codebase. I would like to know if there are ways I can contribute more to this issue. I understand you’re busy, but if possible, could you guide me further?

I found the structure of inputs and outputs from the forward function of the model here. However, I’m not sure if the model needs the normalized text configuration or where to locate the appropriate _task_mapping attribute. I chose these areas based on the tasks the model has been fine-tuned for, but I could use some guidance on where to look.

[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.

[echarlaix]

Hi @tedasdf, thanks for your contribution! I see that you closed this PR yesterday but happy to help if you want to re-open it and continue working on the integration.

I found the structure of inputs and outputs from the forward function of the model here. However, I’m not sure if the model needs the normalized text configuration or where to locate the appropriate _task_mapping attribute. I chose these areas based on the tasks the model has been fine-tuned for, but I could use some guidance on where to look.

The inputs used to export the model needs to be a subset of the model's inputs (a check is done here https://github.com/huggingface/optimum/blob/v1.23.3/optimum/exporters/onnx/convert.py#L96 for this). So you need to first make sure that the inputs defined in your VisualBertOnnxConfig is a subset of https://github.com/huggingface/transformers/blob/v4.46.3/src/transformers/models/visual_bert/modeling_visual_bert.py#L700-L712

In your case I see that you have pixel_values which is not part of the forward signature, do you mean visual_embeds ? If yes, you'll need to rename it using the rename_ambiguous_inputs method. Also pixel_values inputs are generated as tensors of shape [batch_size, num_channels, height, width]

optimum/optimum/utils/input_generators.py

Line 829 in d6de676

shape=[self.batch_size, self.num_channels, self.height, self.width],

which I don't think is what you expect from visual_embeds. You can add one using an DummyInputGenerator