v8.4.196
v8.4.196 - `ultralytics 8.4.196` Improve ONNX INT8 accuracy with per-channel weight quantization (#26978)
Changed 2
- ONNX INT8 exports for opset 13 and newer now use per-channel weight quantization, while older opsets retain per-tensor quantization
- DGX Spark guidance now recommends TensorRT FP16 over INT8 for YOLO26 at batch size 1
Fixed 7
- Speed estimation in video solutions now uses the source video frame rate by default, while an explicitly set FPS still takes precedence behavior changed · inferred, marked in the notes below
- Analytics output video now uses the chart's actual dimensions, avoiding empty output videos in some configurations
- Resume now selects the most recently modified checkpoint, including on Windows
- DOTA readers and annotation visualization now accept UTF-8 BOM label files
- Semantic training preserves small positive class-weight totals when normalizing cross-entropy loss
- Re-exporting to RKNN, Ascend, ExecuTorch, or IMX replaces old files in the output directory
- Screen and webcam prediction examples no longer never return, now using streaming for continuous sources
2 changes to check, marked in the notes below
From ultralytics
🌟 Summary
Release v8.4.196 improves ONNX INT8 model accuracy and makes several practical fixes to training, video analytics, exports, and documentation.
📊 Key Changes
- More accurate ONNX INT8 exports: The main update applies per-channel weight quantization for ONNX opset 13 and newer, helping preserve smaller weight values that can be lost with a single scale. Older opsets retain per-tensor quantization. On the reported COCO evaluation, detection mAP rose from 0.3557 to 0.3814, while segmentation box mAP rose from 0.3213 to 0.3595 and mask mAP from 0.2862 to 0.3046. Some classes may still score lower. The change was tested with all 167 existing ONNX tests passing, and 51 INT8 artifacts loading and running on ONNX Runtime 1.20.0. See PR #26978. behavior changed · inferred
- More reliable video solutions: Speed estimation now uses the source video’s frame rate by default, while an explicitly set FPS still takes precedence. Analytics output also uses the chart’s actual dimensions, avoiding empty output videos in some configurations. behavior changed · inferred
- Safer training and data handling: Resume now selects the checkpoint most recently modified, including on Windows. DOTA readers and annotation visualization now accept UTF-8 BOM label files, and semantic training preserves small positive class-weight totals when normalizing cross-entropy loss.
- Cleaner model exports: Re-exporting to RKNN, Ascend, ExecuTorch, or IMX replaces old files in the output directory, reducing the risk of accidentally loading a stale model.
- Updated guidance: The screen-prediction example now uses streaming, so it can handle a continuous source without collecting results indefinitely. DGX Spark guidance now recommends TensorRT FP16 over INT8 for YOLO26 at batch size 1, based on measurements showing INT8 was slower for smaller models and less accurate across the tested models. Documentation links, wording, and several technical details were also corrected.
🎯 Purpose & Impact
- 📈 Better quality from quantized ONNX models: Per-channel scales can reduce accuracy loss in INT8 exports, including the near-constant depth-map issue reported for YOLO26 depth exports.
- ⚙️ More dependable workflows: Checkpoint selection, DOTA label reading, video analytics, and model re-exports behave more predictably in common edge cases.
- 🧭 Clearer usage and deployment advice: Updated examples and hardware-specific precision recommendations help users avoid misleading results and choose a suitable export option.
What's Changed
- Polish wording on SAM 3, Solutions, Sony IMX500, Kaggle, IBM watsonx and FastSAM docs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/26980
- Weave reference links into docs prose and correct dataset, model and integration details by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/26981
- Select the most recently saved checkpoint when resuming by @Vaishnavi220506 in https://github.com/ultralytics/ultralytics/pull/26973
- Accept UTF-8 BOM labels in DOTA splitting and annotation visualization by @Sonica-B in https://github.com/ultralytics/ultralytics/pull/26882
- Use the source video fps for yolo solutions speed and size analytics output to the chart by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/26966
- Preserve weighted semantic cross-entropy normalization by @aswanth-07 in https://github.com/ultralytics/ultralytics/pull/26969
- Fix the screen and webcam examples that never return by @Echoslayer in https://github.com/ultralytics/ultralytics/pull/26979
- Replace the RKNN, Ascend, ExecuTorch and IMX export directories on re-export by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/26941
- Recommend TensorRT FP16 for YOLO26 on DGX Spark by @atakann in https://github.com/ultralytics/ultralytics/pull/26968
ultralytics 8.4.196Improve ONNX INT8 accuracy with per-channel weight quantization by @amanharshx in https://github.com/ultralytics/ultralytics/pull/26978
New Contributors
- @Sonica-B made their first contribution in https://github.com/ultralytics/ultralytics/pull/26882
- @atakann made their first contribution in https://github.com/ultralytics/ultralytics/pull/26968
Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.195...v8.4.196