run the original YOLOv3 training on the COCO 2014 dataset using Darknet's pretrained weights that correspond to the original YOLOv3 paper

[martin0310]

Search before asking

• [x] I have searched the YOLOv3 issues and discussions and found no similar questions.

Question

How can I run the original YOLOv3 training on the COCO 2014 dataset using Darknet's pretrained weights that correspond to the original YOLOv3 paper ? It seems that the docs does not provide the details.

Additional

No response

[UltralyticsAssistant]

👋 Hello @martin0310, thank you for submitting a ultralytics/yolov3 🚀 Issue. To help us address your concern efficiently, please ensure you've provided the following information:

1. For bug reports:

   • A clear and concise description of the bug
   • A minimum reproducible example MRE that demonstrates the issue
   • Your environment details (OS, Python version, package versions)
   • Expected behavior vs. actual behavior
   • Any error messages or logs related to the issue

2. For feature requests:

   • A clear and concise description of the proposed feature
   • The problem this feature would solve
   • Any alternative solutions you've considered

3. For questions:

   • Provide as much context as possible about your question
   • Include any research you've already done on the topic
   • Specify which parts of the documentation or discussion threads, if any, you've already consulted

From your description, it sounds like you wish to replicate the original YOLOv3 training workflow using COCO 2014 and Darknet's pretrained weights. While our documentation primarily focuses on YOLOv5/YOLOv8 workflows, the YOLOv3 repo includes support for pretrained weights and COCO training. If you have any specifics you'd like us to clarify or encounter any unexpected issues, let us know! 🚀

Please make sure you've searched existing issues to verify this hasn't already been addressed. If you'd like to add additional details or updates, feel free to comment here.

This is an automated response, but no worries—an Ultralytics engineer will review this and assist you further soon. Thank you for engaging with our project and helping us improve! 😊

[pderrenger]

@martin0310 thanks for your question! For original YOLOv3 training with Darknet weights, we recommend using the official Darknet framework: https://github.com/pjreddie/darknet. The Ultralytics YOLOv3 implementation is PyTorch-based with architectural improvements (see our YOLOv3 docs for details). For exact paper replication, Joseph Redmon's Darknet repo remains the best resource.

[martin0310]

Thank you for your quick reply. I have two new questions:

1. Can I run the original YOLOv3 training on the COCO 2014 dataset without using Darknet's pretrained weights, as described in the original YOLOv3 paper?

2. Can the code in the archive branch run the original YOLOv3 training on the COCO 2014 dataset, either using or not using Darknet's pretrained weights, as specified in the original YOLOv3 paper?

I really need the PyTorch version of the code to train YOLOv3 either from scratch or using Darknet's pretrained weights.

[pderrenger]

Thank you for your follow-up.

1. Training from scratch: Yes, you can train YOLOv3 on COCO 2014 from scratch using PyTorch with our implementation. Example:

   from ultralytics import YOLO
   model = YOLO("yolov3.yaml")  # Build from scratch
   model.train(data="coco2014.yaml", epochs=273, pretrained=False)
   

   Results may differ from the original Darknet implementation due to framework differences.

2. Archive branch: The archive branch contains legacy code that is not actively maintained. For original paper replication, we still recommend using Darknet. Our PyTorch implementation includes architectural improvements (see YOLOv3 docs) and isn't identical to the original Darknet version.