🌟 Summary
Ultralytics 8.4.199 improves how training runs appear on Ultralytics Platform, makes several training and profiling workflows more reliable, and clarifies Platform and dataset features in the documentation. No model architectures changed in this release.
📊 Key Changes
- 🏷️ Preserves training run names on Ultralytics Platform: Local training now sends the original run name, including underscores and non-English characters, instead of stripping those characters before display.
- 🧠 Improves first-epoch out-of-memory recovery: When a training run runs out of memory, the retry now reduces the batch size actually used by the data loader. This can help training succeed when the requested batch is larger than the dataset.
- ⚡ Makes TensorRT profiling more consistent: Profiling rebuilds the TensorRT engine for each run and rounds image size to the model’s stride, helping ensure benchmark results reflect the requested settings.
- 📊 Corrects single-class segmentation plots: Pixel counts in
labels.jpgnow show the foreground class rather than the background for single-class semantic datasets. - 📚 Documents additional workflows and features: Dataset splits can now be described as accepting individual image files as well as directories and image lists. The docs also explain how to share a public model’s Predict tab, and cover public comments and ONNX INT8 quantization updates.
🎯 Purpose & Impact
- 📝 Run names are easier to recognize and match with local experiments, especially when they include underscores or non-ASCII characters.
- 🔄 OOM retries are more likely to make progress on small datasets or when the requested batch exceeds the data loader’s actual batch.
- 🔍 TensorRT benchmark results are less likely to rely on an outdated engine or a mismatched image size.
- ✅ Single-class segmentation charts provide more useful class statistics, making dataset checks clearer.
- 🌐 Users can find clearer guidance for sharing models, configuring datasets, and applying ONNX quantization.
What's Changed
- Platform docs: share a public model's Predict tab so others can test it by @glenn-jocher in #27002
- Count the class pixels in single-class semantic label plots by @raimbekovm in #26994
- Halve the loader batch on first-epoch OOM retry by @raimbekovm in #27001
- Rebuild the TensorRT engine and round imgsz to the stride in ProfileModels by @raimbekovm in #26995
- Sync docs with recent ONNX INT8, dataset split and Platform changes by @raimbekovm in #27000
ultralytics 8.4.199Send the raw run name to Ultralytics Platform by @laodouya in #26953
Full Changelog: v8.4.198...v8.4.199