feat: Ultralytics Integration

[joydeep049]

Added basic structure for Integration of Ultralytics into Nyuntam Vision.

-> Flake8 check passed
-> Black check passed
-> Interrogate check passed (100% coverage)
-> PyUpgrade check passed

cc @Arnav0400 @Akash-guna

[Akash-guna]

The solution does not include include loading custom weights. YOLO("{custom_model_path}") should allow loading custom weights other than stock pretrained weights.

[joydeep049]

Just wanted to confirm my understanding:
We either directly load a yolo model like yolo11n.pt or we load custom weights from a cfg file. Is that correct?

@Akash-guna

[joydeep049]

Related documentation at https://github.com/ultralytics/ultralytics/blob/main/ultralytics/models/yolo/model.py

[Arnav0400]

@joydeep049 which models did you test the script on, also are there any experimental numbers?

[joydeep049]

I've been trying to test ultralytics with resnet with distill, but I'm not able to access the nyun docs.
The nyun docs link gives a 404 error
@Arnav0400

[joydeep049]

(Accidentally clicked on close PR)

[Arnav0400]

https://nyunai.github.io/nyuntam-docs/ - Here are the docs

[joydeep049]

Do I need some kind of access to use nyun-vision and nyun-adapt?

@Arnav0400

[Arnav0400]

We recommend using virtual environments instead of dockers for dev cycles - https://github.com/nyunAI/nyuntam?tab=readme-ov-file#git--virtual-environment

Additionally if you want to use dockers, we have fixed the permissions and you can pull now.

[joydeep049]

I'm testing ultralytics on the coco dataset using prune strategy and will share the results soon.
Meanwhile, I wanted to let you know that there are some issues which I had to take care of before I could run my config.

1. In nyuntam-vision/factory.py, in the get_algorithm method the initialize_initialization function is given two arguments, but in the original definition in nyuntam-vision/prune/__init__.py it takes only one argument.

2. Line 57 of nyuntam-vision/factory.py is kw.update(kwargs[algo_type][algorithm]), but in the config we use two prune blocks which is wrong.

3. In the create_from_dict() method in nyuntam/factory.py, the caller iterating over cls always throws an error. But if we do something like:

cls = get_factories(task, job_service)
if(cls[0]):
    return cls[0](args)
else:    
    raise Exception("Factory instance could not be created.")

This can be done because cls only has a single element. Also, it works locallly.

Should I put separate PRs for these issues?

@Arnav0400

[Arnav0400]

Thanks @joydeep049, since nyuntam is a relatively new toolkit we expect many such bugs and errors. @Abhrant is actively looking to resolve them and I would really appreciate if you can raise issues and the corresponding PRs for them.

[joydeep049]

Sure! I will do the aforementioned.

Meanwhile, I request you to take a look at my LinkedIn messages.

@Arnav0400