🌟 Summary
Ultralytics 8.4.197 makes AMP checks work without downloading model weights and removes PyYAML from the base dependencies. 🚀
📊 Key Changes
- AMP checks no longer download a checkpoint. The check builds an untrained YOLO26n from its YAML config and compares FP32 and AMP results across convolution and attention layers.
- AMP checks are more independent of training setup. They check the model’s device without relying on pretrained weights, internet access, or resetting callbacks.
- Ultralytics now includes its own minimal YAML reader and writer. It supports the YAML structures commonly used in Ultralytics configs, datasets, and export metadata. PyYAML is no longer a required base dependency.
- Documentation examples now use
ultralytics.utils.YAMLto save YAML files.
🎯 Purpose & Impact
- 📴 Smoother offline training: AMP checks can run without downloading weights, avoiding unexpected network requests and skipped checks.
- 🛡️ More reliable AMP validation: Comparing individual layers helps identify precision issues before AMP is used for training.
- 📦 Simpler installation: Removing PyYAML from the base requirements reduces a dependency for standard installs. The built-in parser covers Ultralytics’ common YAML needs, but it is not intended to support every YAML feature.
- 📄 More consistent YAML handling: Users can use the same built-in utility to load and save supported Ultralytics YAML files.
What's Changed
- Replace PyYAML with a minimal built-in YAML parser by @glenn-jocher in #26983
ultralytics 8.4.197Check AMP on an untrained YOLO26n without downloading weights by @glenn-jocher in #26967
Full Changelog: v8.4.196...v8.4.197