ultralytics 8.1.23 add YOLOv9-C and E models

[Laughing-q]

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Extended YOLOv9 documentation and model configurations, added new neural network modules, and updated the package version.

📊 Key Changes

• 📝 Tweaked YOLOv9 documentation, including usage examples and supported tasks/modes.
• 💾 Added new model configuration files for YOLOv9c and YOLOv9e.
• 🧠 Introduced new neural network modules like RepNCSPELAN4, ADown, SPPELAN, CBFuse, CBLinear, and Silence.
• 🖥️ Removed a redundant print statement in exporter.py.
• 🔼 Bumped the version in __init__.py from "8.1.22" to "8.1.23".
• 📚 Expanded block.py to include documentation and definitions for the new neural network modules.

🎯 Purpose & Impact

• 📖 Updated documentation empowers users with clearer instructions for training and using YOLOv9 model variants.
• 🌐 New modules expand the neural network architecture possibilities, potentially leading to more efficient or accurate models.
• ⚙️ The version increment reflects these updates, letting users know that new features have been incorporated.
• 🧹 The minor code cleanup (removing the print statement) improves code quality without impacting functionality.

[codecov]

Codecov Report

Attention: Patch coverage is 95.65217% with 4 lines in your changes are missing coverage. Please review.

Project coverage is 75.96%. Comparing base (e138d70) to head (212f793).

Files	Patch %	Lines
ultralytics/nn/modules/block.py	96.42%	3 Missing ⚠️
ultralytics/nn/tasks.py	85.71%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #8571      +/-   ##
==========================================
- Coverage   78.94%   75.96%   -2.98%     
==========================================
  Files         117      117              
  Lines       14709    14797      +88     
==========================================
- Hits        11612    11241     -371     
- Misses       3097     3556     +459     

Flag	Coverage Δ
Benchmarks	36.82% <29.34%> (-0.06%)	⬇️
GPU	38.91% <31.52%> (-0.05%)	⬇️
Tests	70.98% <95.65%> (-3.46%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[glenn-jocher]

@Laughing-q nice work on this!!!

[Laughing-q]

@glenn-jocher do you think we should rename the model to yolov9c/yolov9e?(remove - symbol between yolov9 and the version) to keep consistency with our models.
If so, I could rename it and the pretrained weights then upload weights, if not then I'll upload weights directly.

Currently the links of pretrained weights in yolov9.md are broken, please let me know your thought before merging this PR so I can upload the weights.

[codename-cn]

@Laughing-q please also add/provide a pre-trained yolov9c-seg.pt and yolov9e-seg.pt model.

Thank you very much for your work.

[Burhan-Q]

@Laughing-q I think removing - makes sense to make it similar to the other model naming schema.

[Burhan-Q]

@codename-cn thank you for the feedback! We have started an internal discussion about these, and if we choose to include them, you will see an Issue get added to the YOLO Project so make sure to check back there for updates. It's a lot of work to integrate new models and we all have a lot of other planned work as well, so please be patient with us 🙏 Of course you're also welcome to open a PR to integrate the segmentation pretrained models as well.

[glenn-jocher]

@Laughing-q yes, we should update the naming convention to match ours as you suggested.

[glenn-jocher]

@Laughing-q ok I think this looks good. Once you rename the models can you upload the assets or are you lacking permissions in the assets repo?

Oh I forgot, we need to add the models also the offline assets list here (to avoid overwhelming the GitHub API by only using it if the file is not in this list):

ultralytics/ultralytics/utils/downloads.py

Lines 19 to 30 in 1146bb0

	GITHUB_ASSETS_NAMES = (
	[f"yolov8{k}{suffix}.pt" for k in "nsmlx" for suffix in ("", "-cls", "-seg", "-pose", "-obb")]
	+ [f"yolov5{k}{resolution}u.pt" for k in "nsmlx" for resolution in ("", "6")]
	+ [f"yolov3{k}u.pt" for k in ("", "-spp", "-tiny")]
	+ [f"yolov8{k}-world.pt" for k in "sml"]
	+ [f"yolo_nas_{k}.pt" for k in "sml"]
	+ [f"sam_{k}.pt" for k in "bl"]
	+ [f"FastSAM-{k}.pt" for k in "sx"]
	+ [f"rtdetr-{k}.pt" for k in "lx"]
	+ ["mobile_sam.pt"]
	+ ["calibration_image_sample_data_20x128x128x3_float32.npy.zip"]
	)

[Laughing-q]

@glenn-jocher Done! Pretrained-weights are also uploaded

[glenn-jocher]

@Laughing-q ok I think this is all set now. I just wanted to pass the yolov9 weights through the same script I used for the worldv2 weights to strip dill and add the new docs and license keys to the checkpoints dicts. I should be able to do that later today when I get up, or if you want you could try yourself also using the script I saved in the worldv2 PR.

[Laughing-q]

@glenn-jocher sure no problem, I could update the weights later

[Laughing-q]

@glenn-jocher Done!

[glenn-jocher]

@Laughing-q awesome, tested and everything looks good. Merging PR!

[Ilyabasharov]

Hello, @Laughing-q! Will you upload code for Dual training mode?

[ThecoderPinar]

Amazing 👏👏

[Laughing-q]

@Ilyabasharov Hi, we think the model performs pretty good even without any aux branches while training(dual training mode), so currently supporting this is not in our schedule yet. :) But if you would like to implement this mode and open a PR, we're very welcome and will be there to review. :) Thanks!

[WongKinYiu]

Thanks for adding yolov9 support, I create yolov9-*.yaml for supporting more tasks.

This is my first time to use yolov8 codebase, if I have time to read and understand code, I will try to integrate dual training.

[glenn-jocher]

@WongKinYiu hi there! 🎉 Thanks for contributing the yolov9-*.yaml files to support more tasks. Your initiative is greatly appreciated! If you decide to explore dual training integration later on, we'd be excited to see your contributions. Don't hesitate to reach out if you have any questions while navigating the yolov8 codebase. Happy coding!

[iconictg]

Does it work with a custom trained model? Cause when I run a custom train model, I get the error yolo predict model=yolov8n.pt

[glenn-jocher]

@iconictg hi there! 😊 Yes, YOLOv9 does work with custom trained models. The error you're encountering might be due to specifying the wrong model file in your command. Make sure to replace yolov9c.pt with the path to your custom trained model file, like so:

yolo predict model=path/to/your/custom_model.pt

If your custom model file is named best.pt and located in the runs/train/exp directory, your command would look like this:

yolo predict model=runs/train/exp/best.pt

Hope this helps! Let us know if you have any more questions.

[Burhan-Q]

@iconictg also make sure you're updated ultralytics to at least 8.1.23 or higher. Then you'll need to use

from ultralytics import YOLO

model = YOLO("yolov9c.pt")
result = model.train(data="data.yaml", ...,) # use your training arguments here

Once training is completed, you can load your custom trained model using the YOLO class and perform inference. Be sure to visit the docs pages to learn more!

[iconictg]

It the model I train galan-c and thought that would work and it doesn't Yolov9-c works. Unable to train that in colab cause its so large. Thank you guys.

…

________________________________ From: Burhan ***@***.***> Sent: Thursday, March 7, 2024 2:36 PM To: ultralytics/ultralytics ***@***.***> Cc: Timothy Goebel ***@***.***>; Mention ***@***.***> Subject: Re: [ultralytics/ultralytics] `ultralytics 8.1.23` add YOLOv9-C and E models (PR #8571) @iconictg<https://github.com/iconictg> also make sure you're updated ultralytics to at least 8.1.23 or higher. Then you'll need to use from ultralytics import YOLO model = YOLO("yolov9c.pt") result = model.train(data="data.yaml", ...,) # use your training arguments here Once training is completed, you can load your custom trained model using the YOLO class and perform inference. Be sure to visit the docs pages<https://docs.ultralytics.com/> to learn more! — Reply to this email directly, view it on GitHub<#8571 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/BGQD36JUPF2KYTDANTAXUS3YXDFT5AVCNFSM6AAAAABEBF5SCOVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTSOBUGM3TSNRWGM>. You are receiving this because you were mentioned.Message ID: ***@***.***>

[glenn-jocher]

@iconictg hey there! 😊 It sounds like you're encountering some size constraints with training YOLOv9-C on Colab due to its large size. For training larger models like YOLOv9-C, you might consider using a more powerful environment or reducing the model size if possible.

For YOLOv9-C, ensure you're using the correct model configuration and have enough resources allocated on Colab. Sometimes, adjusting batch sizes or using a simpler model variant can help fit within Colab's limitations.

If YOLOv9-C is too large, you might explore training with YOLOv8 models or other YOLOv9 variants that require less memory. Here's a quick example of how to adjust the batch size for training:

from ultralytics import YOLO

model = YOLO("yolov9c.pt")
result = model.train(data="data.yaml", batch_size=16)  # Adjust batch_size as needed

Remember, training large models on Colab can be challenging due to resource limits. Consider using local resources or cloud services with higher capacities for such tasks. Hope this helps, and happy coding!

[iconictg]

Hi Glenn, I have trained YoloV8 with no issues at all. Yolov9-c was the only issues yes I have change the batch size too. thank you, Timothy Goebel

…

________________________________ From: Glenn Jocher ***@***.***> Sent: Friday, March 8, 2024 12:11 AM To: ultralytics/ultralytics ***@***.***> Cc: Timothy Goebel ***@***.***>; Mention ***@***.***> Subject: Re: [ultralytics/ultralytics] `ultralytics 8.1.23` add YOLOv9-C and E models (PR #8571) @iconictg<https://github.com/iconictg> hey there! 😊 It sounds like you're encountering some size constraints with training YOLOv9-C on Colab due to its large size. For training larger models like YOLOv9-C, you might consider using a more powerful environment or reducing the model size if possible. For YOLOv9-C, ensure you're using the correct model configuration and have enough resources allocated on Colab. Sometimes, adjusting batch sizes or using a simpler model variant can help fit within Colab's limitations. If YOLOv9-C is too large, you might explore training with YOLOv8 models or other YOLOv9 variants that require less memory. Here's a quick example of how to adjust the batch size for training: from ultralytics import YOLO model = YOLO("yolov9c.pt") result = model.train(data="data.yaml", batch_size=16) # Adjust batch_size as needed Remember, training large models on Colab can be challenging due to resource limits. Consider using local resources or cloud services with higher capacities for such tasks. Hope this helps, and happy coding! — Reply to this email directly, view it on GitHub<#8571 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/BGQD36PECVGZBEAV6BXHV4TYXFJAJAVCNFSM6AAAAABEBF5SCOVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTSOBVGA4TMOJZGI>. You are receiving this because you were mentioned.Message ID: ***@***.***>

[glenn-jocher]

@iconictg

Great to hear you've had success with YOLOv8! 🎉 For YOLOv9-C, adjusting the batch size is a good move. If you're still facing issues, it might be worth trying to tweak other parameters like imgsz or using gradient accumulation if you're limited by memory. Here's a quick snippet on how to use gradient accumulation:

from ultralytics import YOLO

# Assuming 'yolov9c.yaml' is your model configuration file
model = YOLO('yolov9c.yaml')
model.train(data='data.yaml', batch_size=4, accumulate=4)  # Adjust 'accumulate' as needed

This approach allows you to effectively increase your batch size without increasing memory usage. If you're still running into obstacles, consider reaching out with specific error messages or behaviors you're observing. We're here to help!

[masharib77]

Does yolov9c or yolov9e supports obb ?

[glenn-jocher]

Hi there! 😊 YOLOv9c and YOLOv9e currently do not natively support oriented bounding boxes (OBB). These models are optimized for regular bounding boxes. However, you can modify the model architecture or post-processing steps to integrate OBB functionality if needed. If you have specific requirements or further questions, feel free to ask!