ultralytics 8.3.210

v8.3.210 - `ultralytics 8.3.210` Skip non-finite training forward passes (#22353)

on Python PyPI

🌟 Summary

More stable training by skipping non-finite loss batches, smoother RT-DETR device/precision moves, and better ONNX profiling for multi-input/dynamic-shape models. 🚀

📊 Key Changes

• Training stability

  • Skip non-finite loss batches (NaN/Inf) during forward pass with a clear warning: “Non-finite forward pass, skipping batch...” 🛡️ (See PR: Skip non-finite training forward passes)
  • Clean initialization of running training loss (tloss) for more consistent logging 📉
  • Version bump to 8.3.210 🧩

• RT-DETR device/precision consistency

  • Added a custom _apply to RTDETRDetectionModel so anchors and masks follow .to(), .cuda(), .cpu(), .half(), etc. 🔁 (See PR: Add _apply to RTDETR)

• ONNX benchmarking improvements

  • Profiling now supports multi-input models and dynamic shapes with sensible defaults and clear validation errors ✅
  • Builds input dictionaries for all inputs and preserves correct dtypes during warmup and timed runs ⏱️ (See PR: Support profiling for multi-input ONNX models)

🎯 Purpose & Impact

• More reliable training out of the box

  • Prevents crashes and bad updates from invalid losses—especially helpful with mixed precision (FP16/BF16) and aggressive hyperparameters ⚖️
  • You may notice occasional warnings and slightly fewer processed batches per epoch; overall convergence should be steadier 📈

• Fewer device/dtype gotchas for RT-DETR

  • Seamless moves across CPU/GPU and precision changes without manual fixes, reducing edge-case bugs in training, inference, export, and deployment 🔧

• Robust and accurate ONNX profiling

  • Benchmark complex ONNX models (multiple inputs, dynamic shapes) without tedious setup, enabling clearer performance comparisons and faster iteration 📊

Quick update tip:

• Upgrade to the latest version: pip install -U ultralytics ✅

What's Changed

• Add _apply method to RTDETR for better device consistency by @Y-T-G in #22341
• Support profiling for ONNX models with multiple inputs by @lmycross in #22262
• ultralytics 8.3.210 Skip non-finite training forward passes by @glenn-jocher in #22353

Full Changelog: v8.3.209...v8.3.210