🌟 Summary
The v8.3.27 release for Ultralytics primarily includes improvements in repository management and default training settings, alongside several enhancements to benchmarking and export functionality.
📊 Key Changes
- Default Training Epochs: Set a fallback of 100 epochs in
trainer.pyif no specific value is provided. - Author Information Update: Refreshed GitHub avatars and usernames for contributors in documentation.
- Removal of Jupyter Checks: Cleaned up unnecessary checks for Jupyter notebooks in
checks.py. - Benchmark Visualization: Introduced interactive benchmark graphs in
benchmark.mdallowing dynamic model comparison through selectable checkboxes. - Export Compatibility: Added checks to skip MNN export tests on Raspberry Pi and NVIDIA Jetson due to known unsupported issues.
🎯 Purpose & Impact
- 🔧 Enhanced Training Robustness: By setting a default epoch, training processes are safeguarded against accidental misconfiguration, ensuring a more reliable setup for users.
- 🏷️ Better Attribution: Updating author profiles ensures that contributors receive the correct recognition, improving engagement and transparency in the community.
- 🛠️ Clean Codebase: Removing obsolete Jupyter checks streamlines the code, making the library more efficient and easier to maintain.
- 📊 User-Friendly Benchmarking: The addition of visual tools for model comparison makes it easier for users to evaluate performance metrics, aiding in informed decision-making.
- 🛡️ Compatibility Safeguards: Preventing exports to MNN format on unsupported devices avoids potential errors and enhances user experience by providing clear usage boundaries.
What's Changed
- Add model comparison graphs in
benchmark.mdby @RizwanMunawar in #17212 - Skip MNN export for Raspberry Pi and NVIDIA Jetson by @lakshanthad in #17292
- Benchmark graph fix by @RizwanMunawar in #17296
ultralytics 8.3.27HUB timed training fix by @glenn-jocher in #17298
Full Changelog: v8.3.26...v8.3.27