Improve/simplify object detection task guide

[NielsRogge]

Feature request

The object detection task guide has a bunch for issues so it'd be great to solve them.

Motivation

Several people have reported the following issues regarding the object detection task guide:

• the model is not predicting any boxes at inference time when you follow the guide for training: https://discuss.huggingface.co/t/example-detr-object-detectors-not-predicting-after-fine-tuning/41824. This may be explained by the slow convergence of DETR. The guide uses 10 epochs, but that may not be enough (cc @MKhalusova who wrote the guide might have some more insights). There's a model on the hub with relatively good metrics which is trained for 100 epochs, so we might need to update this.
• unnecessary complicated evaluation: https://discord.com/channels/879548962464493619/897391062975385640/1207009510171091025
• people need to overwrite the Trainer in order to leverage compute_metrics: https://discuss.huggingface.co/t/possible-fix-for-trainer-evaluation-with-object-detection/72307

Moreover, when inspecting the notebook, I saw:

• there's no GPU leveraged during evaluation
• the timm dependency is no longer required as you can now pass revision="no_timm". Moreover, Accelerate is a dependency but is not included in the guide
• the post_process method needs to be updated to post_process_object_detection, passing a threshold of 0 for computing metrics

Your contribution

I could work on this but would be great if someone else could take this up

[rootonchair]

Hi @NielsRogge , do you still need help on this?

[NielsRogge]

@rootonchair yes, the first 3 points still need to be addressed. We would need to simplify the evaluation of the model during training

[rootonchair]

Sure, let me work on this

[NielsRogge]

Also related: #28893

[g1y5x3]

Here is a training run that I did a while back using the original repo, 10 epochs is roughly around 200k steps in this figure and I doubt the model would give any meaningful predictions at that point. I haven't ran the training through HF's example yet. Will try it to understand the problem better. The problem is just that this model is awfully slow to train even with A100s...

[NielsRogge]

Closing as this is now fixed by #30683