ultralytics/ultralytics v8.3.203

v8.3.203 - `ultralytics 8.3.203` Restore `GradScaler` on resuming training (#22189)

on GitHub

🌟 Summary

Safer, smoother training resumes with AMP plus sturdier exports and better version checks. This release improves mixed‑precision resume stability, future‑proofs ONNX opset selection, tightens export requirements (OpenVINO/CoreML), and polishes compatibility across models and PyTorch versions. 🚀🛡️

📊 Key Changes

• Training resume reliability (priority)
  • Restore AMP GradScaler when resuming training to keep loss scaling stable PR #22189 (fixes #21913, #17282) ⚖️
• ONNX export robustness
  • Smart opset selection via best_onnx_opset() and IR version cap for better ONNX Runtime compatibility PR #22142 🔧
  • Clear Torch gating for opset logic (Torch ≥ 1.13 uses ONNX_MAX_OPSET-1; older Torch uses a map) PR #22155 🧭
• Export requirements updates
  • OpenVINO export now requires PyTorch ≥ 2.1 PR #22156 ⚠️
  • CoreML export now requires PyTorch ≥ 1.11 PR #22162 🍎
• Model and training flow improvements
  • Disable compile=True for YOLO-World and YOLOE with clear assertions; keep compile-optimized path when enabled elsewhere PR #22148 🧠
  • RTDETR requires PyTorch ≥ 1.11 with runtime asserts and aligned tests PR #22161 🎯
• Data and inference fixes
  • Correctly detect image URLs with query strings (no infinite stream loop) PR #22122 🌐✅
• Compatibility cleanups
  • Unified meshgrid version flags to avoid errors on older Torch versions PR #22158, PR #22163 🧩
  • FastSAM workaround for PyTorch 1.9 mask summation bug PR #22169 🩹
• Reliability and UX
  • Warn if multiple Ultralytics installations are detected to prevent conflicts PR #22131 🔍
  • Safer tuner cleanup by checking best_save_dir exists PR #22137 🧹
• Docs and tooling
  • Added reference docs for best_onnx_opset and multiple-install checks; improved slug generation for content tabs; corrected CI badge; enhanced Model YAML Guide and nav PRs #22142, #22157, #22184, #22154 📚
  • Pin dev tooling: uv set to 0.8.19 for consistent environments PR #22186 🧰

🎯 Purpose & Impact

• More reliable mixed-precision resumes 🎛️
  • Restoring the GradScaler avoids instability and prolonged warm-up when continuing training with AMP.
• Fewer export surprises across environments 📦
  • ONNX opset selection and IR capping reduce runtime/load errors.
  • Clear, upfront version requirements for OpenVINO (Torch ≥ 2.1) and CoreML (Torch ≥ 1.11) prevent confusing failures.
• Better user guidance and DX 🧭
  • Early warnings for multiple installs help resolve version/path conflicts quickly.
  • Compile behavior is explicit for YOLO-World and YOLOE, reducing trial-and-error.
• Broader compatibility and stability 🧱
  • Meshgrid and FastSAM fixes ensure older Torch users can run without errors.
• Smoother tuning and documentation experience ✨
  • Safer cleanup logic and improved docs navigation/references make workflows clearer and more predictable.

Tip: If you export to OpenVINO or CoreML, verify your PyTorch version first:

import torch
print(torch.__version__)

Upgrade instructions: see the official PyTorch install guide on the PyTorch website.

What's Changed

• Fix Tuner missing best_save_dir by @glenn-jocher in #22137
• Add check to warn about multiple Ultralytics installation by @Y-T-G in #22131
• Disable compile for WorldModel and YOLOEModel by @Y-T-G in #22148
• Fix url image inference by @Burhan-Q in #22122
• ultralytics 8.3.203 Future-proof ONNX opset selection by @glenn-jocher in #22142
• Use opset table for torch<1.13 by @glenn-jocher in #22155
• Require torch>=2.1 for OpenVINO exports by @glenn-jocher in #22156
• Fix meshgrid with indexing bug on torch==1.10 by @glenn-jocher in #22158
• Require torch>=1.13 for YOLOE models by @glenn-jocher in #22160
• Enable slug generation for content tab anchor links in MkDocs by @Y-T-G in #22157
• Require torch>=1.11 for RTDETR models by @glenn-jocher in #22161
• Fix transformer.py meshgrid for torch<1.11 by @glenn-jocher in #22163
• Require torch>=1.11 for CoreML exports by @glenn-jocher in #22162
• PyTorch 1.9 FastSAM sum(0).sum() predictor fix by @Y-T-G in #22169
• Enable comments and add cloning instruction to Model YAML configuration guide by @Y-T-G in #22154
• Add Actions CI badge to Docs by @glenn-jocher in #22184
• Add uv required-version = "==0.8.19" by @glenn-jocher in #22186
• ultralytics 8.3.203 Restore GradScaler on resuming training by @Y-T-G in #22189

Full Changelog: v8.3.202...v8.3.203