ultralytics 8.3.211

v8.3.211 - `ultralytics 8.3.211` Skip non-finite gradients in `optimizer_step()` (#22350)

on Python PyPI

🌟 Summary

More robust training and downloads. This release skips optimizer updates on non‑finite gradients (NaN/Inf) to prevent crashes and centralizes asset URLs for more reliable tests and dataset downloads. 🛡️🌐

📊 Key Changes

• Training stability
  • Trainer now checks for non‑finite loss before backward; on NaN/Inf it logs a warning and skips the backward pass. ⚠️
  • New safety in optimizer_step(): gradient clipping with non‑finite detection and safe skip of the optimizer update instead of crashing.
  • Keeps scaler updates and gradient zeroing consistent to maintain training flow.
• Asset and download reliability
  • Introduced a single ASSETS_URL constant and replaced hardcoded asset links across tests, data utils, benchmarks, and converters.
  • downloads.is_url() is now safer and faster:
    • Returns False (rather than asserting) on invalid URLs.
    • Uses a quick HEAD request with timeout for existence checks.
  • safe_download() now aliases ASSETS_URL to the public assets host automatically, simplifying host switching.
• Logging improvements
  • Trainer logs now include the actual loss value when skipping non‑finite batches for easier debugging.
• Version
  • Bumped to 8.3.211.

Key PRs:

• Primary: Skip non‑finite gradients in optimizer_step() by @glenn-jocher — see PR ultralytics 8.3.211
• Supporting: Centralize ASSETS_URL and improve download robustness — see PR Use ASSETS_URL in tests

🎯 Purpose & Impact

• Fewer training interruptions: Prevents crashes from NaN/Inf gradients and lets long runs finish more reliably with YOLO11/YOLO26 and other Ultralytics models. 🚦
• Clearer debugging: Warnings explicitly report non‑finite loss, helping pinpoint unstable hyperparameters, data issues, or mixed‑precision edge cases.
• More reliable CI and user setups: Consistent asset hosting via ASSETS_URL reduces flaky tests and download issues across environments.
• Backward‑compatible: No user code changes required; behavior is safer by default.

Quick start or upgrade:

• pip install -U ultralytics

Tip: If you see warnings about non‑finite loss or gradients, try:

• Lowering the learning rate, reducing augmentation intensity, or disabling mixed precision (amp=False).

What's Changed

• Use ASSETS_URL in tests by @glenn-jocher in #22355
• ultralytics 8.3.211 Skip non-finite gradients in optimizer_step() by @glenn-jocher in #22350

Full Changelog: v8.3.210...v8.3.211