ultralytics 8.3.205

v8.3.205 - `ultralytics 8.3.205` Reset checkpoint arguments after training (#22286)

on Python PyPI

🌟 Summary

Cleaner post-training behavior and clearer visuals: v8.3.205 refines how training configs are restored from checkpoints, improves fitness plots with smarter outlier filtering, and updates docs for more predictable inference and easy-start notebooks. ✅

📊 Key Changes

• Reset checkpoint overrides after training (@Y-T-G) 🔁
  • Trainer now restores overrides via a new helper (_reset_ckpt_args) instead of using raw model.args.
  • Fixes a community-reported issue where results could be saved to unintended run directories.
• Smarter Tune scatterplots (@glenn-jocher) 📈
  • Applies iterative 3-sigma rejection on low outliers to produce clearer, more reliable fitness plots and best-run selection.
• CI reliability boost (@glenn-jocher) 🔄
  • Pytest wrapped in a retryable GitHub Action with a single automatic retry to reduce flaky failures.
• Inference docs: rect padding clarified (@Y-T-G) 🧠
  • Added note explaining default minimal padding behavior during predict, and how batch size and image sizes affect padding.
• New one-click training for Construction-PPE dataset (@RizwanMunawar) 🚀
  • Added a Colab badge to quickly launch a ready-to-run notebook.
• SAM docs link updated (@onuralpszr) 🔗
  • Now points to the official Segment Anything GitHub for accurate, up-to-date references.

Useful links:

• Current PR: Reset checkpoint arguments after training
• Plotting improvements: 3-sigma outlier rejection for Tune plots
• CI stability: Retry slow CI tests once
• Inference behavior note: Predict mode docs update
• Construction-PPE notebook: Open the Colab tutorial
• SAM reference: Segment Anything GitHub

🎯 Purpose & Impact

• More reliable training workflows 🛠️
  • Prevents stale or unintended args from leaking from checkpoints—reducing surprises when resuming or finalizing training and fixing incorrect results directory issues.
• Clearer insights during tuning/evolution 📊
  • Outlier filtering makes plots easier to read and improves the stability of best-run identification.
• Predictable inference behavior 🧩
  • Better understanding of minimal vs. square padding helps you plan for memory, speed, and consistent outputs when running predict.
• Faster onboarding and experimentation 🚀
  • One-click Colab for the Construction-PPE dataset lowers the barrier for training YOLO models without local setup.
• Stable development pipeline ✅
  • CI retries reduce flaky failures, improving contributor experience and build reliability.

What's Changed

• Docs: Add rect behavior note to predict.md by @Y-T-G in #22266
• Add Construction-PPE notebook in docs by @RizwanMunawar in #22269
• docs: 📝 Update link to Segment Anything GitHub in SAM documentation by @onuralpszr in #22268
• Apply 3-sigma iterative rejection to Tune scatterplots by @glenn-jocher in #22290
• Retry Slow CI tests once by @glenn-jocher in #22292
• ultralytics 8.3.205 Reset checkpoint arguments after training by @Y-T-G in #22286

Full Changelog: v8.3.204...v8.3.205