Add: doc page for the object detection task

[MKhalusova]

This is a PR for the #20805 issue.
The guide has content and working code examples for:

• 
Introduction
• 
Loading CPPE-5 dataset from Hub
• 
Preprocessing both images and annotations. Images are augmented, and annotations are reformatted to be in the format DETR expects
• 
Training with Trainer
• 
Evaluation
• 
Inference

[MKhalusova]

This PR replaces the #20874

[MKhalusova]

To preserve the discussion, here's @sayakpaul 's comment relevant to the CI issue: #20874 (comment)

[HuggingFaceDocBuilderDev]

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

[sayakpaul]

Excellent work @MKhalusova! I like the language and how the code is structured. I really really liked the code :)

I left a couple of requests for changes. Feel free to ask questions about them.

Do you have a Colab Notebook where this code has been tested (preferably with outputs)?

[sayakpaul]

We also need to embed the resultant image after this. Refer here for a reference.

We keep the documentation images here: https://huggingface.co/datasets/huggingface/documentation-images. Once all the images have been collated, we open a PR to that repo (documentation-images). Before opening the PR, we usually compress the images (ImageOptim is a good tool) and add them as Git LFS objects.

Once you raise the PR and until it's merged, you can host the images on IMGUR.

[MKhalusova]

Makes sense, I'll create a separate PR to add images and meanwhile I'll add it to the page as an imgur-hosted image.

[MKhalusova]

I added an image in 1a38260
Creating a PR to (documentation-images) is on my TODO.

[stevhliu]

I think it'd be easier to add the images to the documentation-images repo yourself. I believe opening a PR is more for external contributors who can't commit directly 🙂

[sayakpaul]

Spend a few sentences discussing the results and also append the output of print(results).

[sgugger]

Thanks a lot for working on this guide, it looks great!
My main comments are:

• use the auto API everywhere. The convention for the docs is to only use one specific model/config/tokenizer/processor class in the docstring of that class, everywhere else should be the auto APIs as they will work without change for any other checkpoint.
• Make sure to split long code samples in several smaller ones with text introducing each step. The evaluation in particular is too long and may be too technical for this guide.

[sgugger]

This is indeed way too long for this guide. It should be moved to a more advanced guide or juse be part in an example (this whole guide would make a great example once converted to a Python script) that we can link here.

[MKhalusova]

• Make sure to split long code samples in several smaller ones with text introducing each step. The evaluation in particular is too long and may be too technical for this guide.

I agree that the evaluation part is a bit too technical, but unfortunately at the moment, there's no simpler way (hopefully soon there will be an easier way to have coco evaluation metrics). But I can certainly split it up somewhat.

[NielsRogge]

Please include a guide on how people can use their own data here. Using a HuggingFace dataset is nice, but the first question people always come up with is "but how I can use my own custom dataset?".

Here I'd recommend adding a link to ImageFolder + metadata, which allows people to turn their own images + bounding box annotations into a 🤗 Dataset: https://huggingface.co/docs/datasets/v2.8.0/en/image_dataset#object-detection

[sgugger]

This is out of scope of the tasks guide. None of the other one does that.

[MKhalusova]

If this is a common question, perhaps, this can be a note or a tip in this doc? https://huggingface.co/docs/transformers/training
Something along the lines "In the Transformers documentation we use datasets from the Hugging Face Hub. If you would like to use your own custom data, check out the Datasets documentation on how to create a dataset" ?

[sayakpaul]

That would be a good addition IMO.

[NielsRogge]

Suggested change

	1. Finetune [DETR](https://huggingface.co/docs/transformers/model_doc/detr), a model that combines a convolutional
	1. Fine-tune [DETR](https://huggingface.co/docs/transformers/model_doc/detr), a model that combines a convolutional

[MKhalusova]

Technically, both spellings are correct, however, I see that some folks have a preference for one, and some for the other. Both can be seen throughout the docs. Perhaps, we should agree on a single spelling variant, and update it throughout the whole doc.

[sgugger]

In an opensource repo like Transformers, it's going to be very tedious to maintain a consistent spelling since we have so many contributors (for this or pre-trained/pretrained). The positive impact of forcing one spelling is exactly zero since both are perfectly understandable, while taking time from reviewers/maintainers, so overall a negative impact.

[NielsRogge]

Suggested change

	2. Use your finetuned model for inference.
	2. Use your fine-tuned model for inference.

[NielsRogge]

Suggested change

	To finetune a model, you must preprocess the data you plan to use to match precisely the approach used for the pre-trained model.
	To fine-tune a model, you must preprocess the data you plan to use to match precisely the approach used for the pre-trained model.

[NielsRogge]

Suggested change

	Instantiate the image processor from the same checkpoint as the model you want to finetune.
	Instantiate the image processor from the same checkpoint as the model you want to fine-tune.

[NielsRogge]

Suggested change

	4. Call [`~Trainer.train`] to finetune your model.
	4. Call [`~Trainer.train`] to fine-tune your model.

[NielsRogge]

Suggested change

	Now that you have finetuned a DETR model, evaluated it, and uploaded it to the Hugging Face Hub, you can use it for inference.
	The simplest way to try out your finetuned model for inference is to use it in a `pipeline()`. Instantiate a pipeline
	for object detection with your model, and pass an image to it:
	Now that you have fine-tuned a DETR model, evaluated it, and uploaded it to the Hugging Face Hub, you can use it for inference.
	The simplest way to try out your fine-tuned model for inference is to use it in a `pipeline()`. Instantiate a pipeline
	for object detection with your model, and pass an image to it:

[sgugger]

@NielsRogge note that finetune/fine-tune has no decided standard in the doc/transformers and both are used equally.

[stevhliu]

Great job, and thanks for all your work on this guide! It'll be huge for our computer vision efforts :)

[sayakpaul]

Thanks a LOT for your hard work @MKhalusova! Everything looks good to me. What a huge effort you have put in here and I am sure the community will be benefitted from it.