ultralytics 8.2.27 replace onnxsim with onnxslim

[Kayzwer]

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Simplified ONNX export and improved Docker image configuration for arm64 architecture.

📊 Key Changes

• Dockerfile Update: Removed unnecessary build-essential package installation, streamlining Docker setup for arm64.
• Export Documentation: Clarified simplify option usage in model export settings.
• Export Testing Enhancements: Integrated simplify parameter in ONNX export test matrices and utilized onnxslim for model simplification in tests.
• Software Dependencies: Shifted from onnxsim to onnxslim to avoid complex dependencies like cmake.
• Version Bump: Updated project version to 8.2.27.

🎯 Purpose & Impact

• Efficiency: Removing build-essential from Dockerfile reduces image size and build time 💾.
• Clarity: Updated documentation and variable naming improve understanding of export options 📚.
• Testing Robustness: Expanded export test coverage increases confidence in model exports 🧪.
• Simplification: Replacing onnxsim with onnxslim avoids complications related to dependencies, simplifying user experience and installation process 🛠️.
• Project Evolution: Increment in project version marks progress and enhancement, readying for new features and improvements 🚀.

[Kayzwer]

@inisis I did not change the doc this time, I have no idea what conflict happened there.

[codecov]

Codecov Report

Attention: Patch coverage is 10.00000% with 9 lines in your changes are missing coverage. Please review.

Project coverage is 37.22%. Comparing base (dd13707) to head (01a59f9).

Files	Patch %	Lines
ultralytics/engine/exporter.py	0.00%	6 Missing ⚠️
ultralytics/utils/benchmarks.py	0.00%	3 Missing ⚠️

Additional details and impacted files

@@             Coverage Diff             @@
##             main   #12989       +/-   ##
===========================================
- Coverage   70.26%   37.22%   -33.04%     
===========================================
  Files         124      124               
  Lines       15681    15681               
===========================================
- Hits        11018     5837     -5181     
- Misses       4663     9844     +5181     

Flag	Coverage Δ
Benchmarks	?
GPU	37.22% <10.00%> (ø)
Tests	?

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.

[inisis]

@Kayzwer should we modify here to make onnxslim true. https://github.com/ultralytics/ultralytics/blob/main/ultralytics/engine/exporter.py#L673

[Kayzwer]

@Kayzwer should we modify here to make onnxslim true. https://github.com/ultralytics/ultralytics/blob/main/ultralytics/engine/exporter.py#L673

I set slim=True by default in the config yaml file

[inisis]

@Kayzwer should we modify here to make onnxslim true. https://github.com/ultralytics/ultralytics/blob/main/ultralytics/engine/exporter.py#L673

I set slim=True by default in the config yaml file

In order to test onnxslim only, I dont't think this is the right way, because it will call both.

[Kayzwer]

@Kayzwer should we modify here to make onnxslim true. https://github.com/ultralytics/ultralytics/blob/main/ultralytics/engine/exporter.py#L673

I set slim=True by default in the config yaml file

In order to test onnxslim only, I dont't think this is the right way, because it will call both.

no, I mean I did not set slim=True in code, I set it in config yaml file only

[inisis]

@Kayzwer should we modify here to make onnxslim true. https://github.com/ultralytics/ultralytics/blob/main/ultralytics/engine/exporter.py#L673

I set slim=True by default in the config yaml file

In order to test onnxslim only, I dont't think this is the right way, because it will call both.

no, I mean I did not set slim=True in code, I set it in config yaml file only

but there are hard-coded part, and will not be affected by config yaml, as I mentioned before

[Burhan-Q]

@Kayzwer since the onnx export is used in a few other export formats, it would be good to update the unit test matrix functions (test_export_*_matrix) in tests/test_export.py and tests/test_cuda.py that would check the addition of the new argument; see example. Additionally, with the addition of a new export argument, the docs tables (I think there are ~11) would all need to be updated to include this new argument.

[inisis]

@Kayzwer Can you add me as a collaborator in your repo, this code needs to be changed.

[Kayzwer]

@Kayzwer Can you add me as a collaborator in your repo, this code needs to be changed.

invited

[inisis]

@Kayzwer I don't have permission to push

[Kayzwer]

holy shit github, I have no idea where to click to make an collaborator have write access

[Kayzwer]

@Kayzwer I don't have permission to push

sry man, I cant do that, cause microsoft is piece of shit

[Kayzwer]

@Kayzwer I don't have permission to push

I think you just take my code and make a pull request on ultralytics repo

[inisis]

Hi, @glenn-jocher @Burhan-Q @Kayzwer I have tested on my machine, and it works fines, this is pretty much the final version of code, please help review it ❤️❤️❤️,

[glenn-jocher]

Hi there! Thanks for sharing your work and the test results. 🚀 It looks like you've put a lot of effort into this. I'll review the code and get back to you with any feedback or suggestions. Keep up the great work!

[Kayzwer]

no prob

[glenn-jocher]

@Kayzwer @inisis I'm satisfied with the benchmarks here. It's unclear if there are super performance improvements, but I really like the package simplicity vs onnxsim that's not python native, requires cmake dependency for building install wheels etc, onnxslim is much simpler in that sense and should help with more robust installs across all the environments users run ultralytics in.

I'm replacing onnxsim with onnxslim now in this PR under the same simplify flag.

[inisis]

@Kayzwer @inisis I'm satisfied with the benchmarks here. It's unclear if there are super performance improvements, but I really like the package simplicity vs onnxsim that's not python native, requires cmake dependency for building install wheels etc, onnxslim is much simpler in that sense and should help with more robust installs across all the environments users run ultralytics in.

I'm replacing onnxsim with onnxslim now in this PR under the same simplify flag.

Hi @glenn-jocher personally I don't want to do comparison, it's up to you to make choice🤔. I'd rather keep both🫣

[glenn-jocher]

Hi @inisis, thanks for sharing your thoughts! We appreciate your input and understand the desire to maintain flexibility. The decision to streamline to onnxslim under the simplify flag is aimed at enhancing the installation process and ensuring consistency across various environments. However, we'll keep monitoring feedback and usage to ensure we meet the community's needs effectively. Your feedback is invaluable in this ongoing process! 🌟

[glenn-jocher]

@inisis @Kayzwer PR merged! Thank you for your contributions!

[lakshanthad]

Verified on RPi CI also and all ✅ :)
#13281

[inisis]

Hi @glenn-jocher should we apply this to other yolo repos like https://github.com/ultralytics/yolov5

[glenn-jocher]

Hi @inisis, great suggestion! It would be beneficial to maintain consistency across our repositories. Let's evaluate the integration on the YOLOv5 repo as well. I'll start by reviewing the current setup and testing to ensure a smooth transition. Thanks for pointing this out! 🚀

[inisis]

@glenn-jocher May I add you on linkedin, and this is me , I already sent an invitation

[glenn-jocher]

Hi @inisis,

Thank you for reaching out and for your interest in connecting on LinkedIn! I appreciate the invitation and will be sure to check it out. 😊

If you have any questions or need further assistance with Ultralytics projects, feel free to ask here. We're always happy to help and engage with the community!