ultralytics 8.4.194

v8.4.194 - `ultralytics 8.4.194` Size AutoBatch from the train dataset for every task (#26956)

on Python PyPI

🌟 Summary

Ultralytics 8.4.194 makes automatic batch sizing more reliable across tasks, alongside practical fixes for model loading, results saving, datasets, and deployment workflows.

📊 Key Changes

• 🚀 Automatic batch sizing now uses the training dataset for every task. It accounts for dataset size and labels across classification, detection, and multi-dataset training. If profiling runs out of memory during backward computation, AutoBatch treats that as a limit rather than relying on an invalid measurement.
• 🛡️ Model YAML parsing is safer and more consistent. Activations must resolve to supported torch.nn modules, and model entries must be PyTorch module classes. Custom YAMLs using other expressions may need updating.
• 💾 Saving prediction results is more robust. Results.save() and plot(save=True) now produce sensible .jpg filenames for sources such as webcams, videos, or missing paths, and create parent directories when needed.
• 🔄 Similarity search uses the selected image directory on each startup. This avoids reusing stale indexes from another dataset; the trade-off is that embeddings are rebuilt each time.
• 🧩 Dataset and URL handling is more dependable. Non-PNG semantic masks are resolved before cache hashing, and URL schemes such as HTTPS://, GS://, and UL:// are now handled without case sensitivity.
• 📈 Benchmarking and evaluation guidance are clearer. TensorFlow GraphDef benchmarks now include OBB and pose, and the docs explain that classes absent from validation receive the overall mAP value rather than a measured per-class score.

🎯 Purpose & Impact

• ✅ AutoBatch is less likely to choose a batch size that exceeds a small dataset or available GPU memory, and multi-dataset World/YOLOE training avoids label-access failures.
• 🔒 Consistent, restricted YAML parsing improves safety, though nonstandard custom model configurations may no longer load.
• 🧰 More reliable result saving, mask-cache updates, and URL routing reduce common workflow errors.
• 🔍 Similarity search results better reflect the images currently in the chosen folder, while requiring index-building work at startup.
• 📚 Documentation updates clarify metric interpretation and integration behavior; no new model architecture is introduced in this release.

What's Changed

• Fix classification dataset URL scheme handling by @Vaishnavi220506 in #26879
• Benchmark TensorFlow GraphDef for OBB and pose by @glenn-jocher in #26952
• Give each Colab VM its own anonymous settings id by @glenn-jocher in #26959
• Fix settings help URL punctuation by @glenn-jocher in #26962
• Update the COCO JSON guide's cache note now that the hash includes mtime by @Echoslayer in #26960
• Explain how metrics.box.maps fills classes absent from the val set by @Echoslayer in #26955
• Build the similarity search index from the data directory on every start by @raimbekovm in #26965
• Resolve non-PNG semantic mask paths before hashing the label cache by @raimbekovm in #26954
• Handle URL schemes case-insensitively at remaining routing sites by @cainiao33 in #26964
• Fix Results.save and plot default filename when path is None or extensionless by @Rahulbiradar9 in #26958
• Fix stale framing, logger defaults and nav order on integration pages by @raimbekovm in #26957
• Use one model YAML parsing path in parse_model by @Y-T-G in #26963
• ultralytics 8.4.194 Size AutoBatch from the train dataset for every task by @raimbekovm in #26956

Full Changelog: v8.4.193...v8.4.194