ultralytics 8.3.181

v8.3.181 - `ultralytics 8.3.181` Support `half` inference for SAM models (#21735)

on Python PyPI

🌟 Summary

v8.3.181 focuses on safer, faster mixed-precision (FP16) support for SAM/SAM2 and YOLOE pipelines, plus reliability fixes across validation exports, source loading, and single-class training. 🚀

📊 Key Changes

• SAM FP16 support and dtype/device consistency (PR #21735, priority)

  • Enables half-precision inference for SAM models with consistent dtype handling across blocks, encoders, decoders, utils, and predict paths.
  • Avoids unnecessary float32 casts; ops now respect input tensor dtype (float16/bfloat16 where safe).
  • Predictor now normalizes before casting and sets model dtype based on args.half; unified self.torch_dtype for prompts/masks/buffers.

• SAM2 robustness without high_res_features (PR #21726)

  • Decoder gracefully falls back when high-res features are absent; accepts tensor or dict feature inputs.

• YOLOE device/half propagation and stability (PR #21670)

  • Predict now forwards device/half flags; prompt tensors follow model precision; softmax casting simplified for consistency.

• Validation export consistency across tasks (PR #21719)

  • New scale_preds unifies scaling to original image sizes for detect, OBB, pose, and segment before saving JSON/TXT.

• YOLOE visual prompt predictor switching fix (PR #21731)

  • Predictor instance now correctly switches after initialization when using visual prompts.

• Single-class training compatibility (PR #21725)

  • Restores classes with single_cls by safely constraining max class index to 0 (no label mutation).

• CSV source support for inference (PR #21729)

  • Dataloaders now accept .csv source lists with whitespace-safe parsing.

• Streamlit Live Inference improvements (PR #21553)

  • Accepts multiple export formats (.pt, .onnx, .torchscript, .mlpackage, .engine, OpenVINO) and respects full paths provided by users.

• YOLOE docs enhancements (PR #21728)

  • Clearer fine-tuning, linear probing, and new export examples; minor classify docs correction.

🎯 Purpose & Impact

• Faster inference on modern GPUs ⚡
  • FP16 support for SAM reduces memory use and can speed up inference on compatible hardware.
• Greater numerical stability and fewer dtype/device surprises 🛡️
  • Consistent dtype handling across SAM/SAM2 and YOLOE reduces precision mismatches and unintended casts.
• More robust segmentation and visual prompting workflows 🧩
  • SAM2 now works even without high-res features; YOLOE honors device/half flags and switches predictors reliably.
• Accurate and consistent validation exports across tasks 📏
  • JSON/TXT outputs now consistently match original image sizes for detect/OBB/pose/segment.
• Easier deployment and input management 🧰
  • CSV sources work out of the box; Streamlit Live Inference loads multiple model formats seamlessly.
• Smoother single-class training ✅
  • Prevents spurious “class exceeds count” errors, improving reliability for single-class projects.

What's Changed

• Fix save_txt coordinates scaling by @Y-T-G in #21719
• Add multiple export formats inference support in Live Inference solution by @RizwanMunawar in #21553
• Restore classes compatibility with single_cls by @Y-T-G in #21725
• YOLOE: Fix visual prompt predictor not switching after initialization by @Y-T-G in #21731
• Fix inference with CSV source by @Y-T-G in #21729
• docs: improved documentation for yoloe export by @picsalex in #21728
• YOLOE: Fix device selection and support half inference for visual prompt by @RizwanMunawar in #21670
• Support inference on SAM models without high_res_features by @Laughing-q in #21726
• ultralytics 8.3.181 Support half inference for SAM models by @Laughing-q in #21735

Full Changelog: v8.3.180...v8.3.181