Pycocotools best.pt after COCO train

[glenn-jocher]

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Implementation of YOLOv3 and YOLOv3-tiny configurations into the Ultralytics yolov5 repository and refinement of various utilities.

📊 Key Changes

• Added YOLOv3-tiny and YOLOv3 model configurations to the project.
• Removed unnecessary usage of glob in test.py, specifying the exact annotation JSON path instead.
• Enhanced string formatting in Python by utilizing f-strings in test.py.
• Updated train.py with new padding strategy and streamlined optimizer state saving.
• Made a small change in google_utils.py to ensure weight strings are properly formatted.

🎯 Purpose & Impact

• 🏗 The new model configurations allow researchers and developers to train YOLOv3 and YOLOv3-tiny models with the Ultralytics framework, offering more variety and flexibility.
• 🛠 The update enhances code clarity and maintainability by specifying exact file paths and adopting modern Python string formatting strategies.
• 🚀 The train.py adjustments aim to optimize the training process, potentially leading to better performance or faster training times.
• 👩‍💻 The changes in google_utils.py aim to minimize errors related to weight file handling, leading to a smoother user experience when downloading pre-trained models.

[glenn-jocher]

This PR runs best.pt with pycocotools after COCO training. Custom training is unaffected. Additional updates were made to testloader in train.py to match exactly the arguments used in test.py when called directly. This should lead to more reproducible results between train.py and test.py, which a number of users had raised issues about previously.

[glenn-jocher]

Lastly YOLOv3 models are imported to models/hub from the newly updated https://github.com/ultralytics/yolov3.