ultralytics 8.4.113

v8.4.113 - Add Dragonwing IQ-8275 QNN export target (#25558)

on Python PyPI

🌟 Summary

🚀 Ultralytics 8.4.113 expands Qualcomm QNN deployment to Dragonwing IQ-8275 devices, improves model visualization and export reliability, and standardizes task support documentation across the project.

📊 Key Changes

• Qualcomm Dragonwing IQ-8275 QNN export

  • Adds iq-8275 and qcs8275 as QNN export targets using Qualcomm SoC model 82.
  • Centralizes the mapping between supported QNN targets and their HTP or SoC provider options.
  • Adds documentation for supported Snapdragon and Dragonwing targets, including the unsupported IQ-615.
  • Enables QNN export testing on Linux x86-64 with onnxruntime-qnn==2.4.0.
  • Example usage: model.export(format="qnn", name="iq-8275").
  • Requires a compatible Qualcomm/Yocto BSP and target-side QNN, FastRPC, DSP firmware, and driver components.

• New class activation heatmaps for prediction 🔥

  • Replaces the previous visualize=True feature-map dump with a more useful LayerCAM-style heatmap.
  • Saves one heatmap image per input, showing which image regions influenced the predicted class scores.
  • Respects confidence and class filters and is available for PyTorch models.
  • Reduces output clutter compared with the former multi-file feature-map visualization.

• More accurate attention FLOPs reporting 📊

  • THOP-based profiling now counts functional attention matrix multiplications in YOLO12 attention blocks.
  • Models containing area attention are measured at the requested image size rather than extrapolated from a small stride-sized input.
  • Reported GFLOPs should better reflect the real computational cost of attention-heavy models.

• Improved RKNN and quantized export support

  • Normalizes detection and pose coordinates during INT8 RKNN export to preserve class-score precision.
  • Restores coordinates at runtime for RKNN inference.
  • Adds compatibility handling for current rknn-toolkit2 dependencies, including setuptools<82.
  • Improves ONNX compatibility and cleans up temporary normalized graphs after export.

• Export and runtime reliability fixes 🛠️

  • Resets detection shape caches at TorchScript, ONNX, and OpenVINO export boundaries.
  • Avoids redundant TorchScript retracing.
  • Replaces ONNX advanced indexing with gather where needed.
  • Fixes YOLOE-26 prompt-free RKNN export for detection and segmentation.
  • Corrects pose loss selection for end-to-end models using non-Pose26 heads.
  • Prevents C2PSA failures when its channel count is below 64.
  • Fixes MuSGD handling of custom model head locations and higher-rank parameters.

• Performance and stability improvements ⚡

  • Vectorizes Deep OC-SORT global motion compensation operations.
  • Reduces temporary memory usage and plotting time for segmentation masks by processing them in row bands.
  • Keeps accumulated heatmap overlays visible when a tracking frame temporarily contains no detections.
  • Prevents grayscale video stream failures from killing readers or crashing inference.
  • Makes mixed-text ordering deterministic across processes and distributed workers.
  • Keeps profiler samples when runtime variance reaches zero.

• Documentation consistency and accuracy 📚

  • Establishes one canonical task and mode order across code, documentation, tables, and the Ultralytics Platform.
  • Consolidates supported-task tables into a shared macro and reports model-family-specific support more accurately.
  • Corrects multiple docstrings, doctest examples, export claims, task descriptions, and API references.
  • Updates Hailo, QNN, GraphDef, RKNN, and other integration pages with clearer model and hardware limitations.

🎯 Purpose & Impact

• For Qualcomm users: IQ-8275 and QCS8275 devices can now be targeted directly during QNN export, simplifying deployment to Dragonwing hardware. Hardware-side BSP compatibility must still be verified.
• For model developers: Heatmaps provide a clearer way to understand model decisions, while corrected FLOPs estimates make performance comparisons more trustworthy.
• For edge deployment: RKNN and QNN improvements increase export compatibility and improve quantized inference reliability on specialized accelerators.
• For training workflows: MuSGD, pose-loss, C2PSA, and higher-rank tensor fixes reduce silent training errors and improve support for custom architectures.
• For tracking and streaming applications: Video recovery, heatmap persistence, and Deep OC-SORT updates improve robustness in real-world, imperfect inputs.
• For documentation users: Shared and empirically validated compatibility tables make it easier to determine which model family, task, and deployment target is supported. 🚀

What's Changed

• Fix the dangling chi2inv95 reference in the Kalman gating distance docstring by @raimbekovm in #25516
• Set imgsz in SAM pre_transform docstring examples by @raimbekovm in #25512
• Consolidate export task tables into a shared macro and correct unproven claims by @glenn-jocher in #25523
• docs: fix missing 'to' in Probs description by @rudrakumar07 in #25522
• keep the mixed text order stable across processes by @raimbekovm in #25540
• Vectorize the Deep OC-SORT GMC warp and drop a redundant mean copy by @raimbekovm in #25534
• Fix MuSGD head grouping for custom backbones by @fcakyon in #25532
• Fix Docker CI failures and export warnings by @glenn-jocher in #25525
• Keep profiler samples when the run time variance is zero by @raimbekovm in #25513
• Fix MuSGD batching for higher-rank parameters by @songjiahao-wq in #25317
• Fix PoseModel using PoseLoss26 for non-Pose26 heads when end2end=True by @Zenka737 in #25261
• Replace pytube/pafy with yt-dlp for YouTube streams by @ambitious-octopus in #16336
• Keep stream reader alive when grayscale frame read fails by @raimbekovm in #24709
• Revert "Replace pytube/pafy with yt-dlp for YouTube streams" by @glenn-jocher in #25546
• Count the attention matmuls YOLO12 area-attention runs functionally by @glenn-jocher in #25545
• Clarify IoA wording in Copy-Paste augmentation documentation by @gizembm in #25528
• fix: prevent C2PSA division by zero when c < 64 by @rudrakumar07 in #25517
• Fix supported-tasks table rendering on docs.ultralytics.com by @raimbekovm in #25544
• correct the declared output in fourteen docstring examples by @raimbekovm in #25543
• Remove the unmatchable printed output from the Kalman filter docstring examples by @raimbekovm in #25533
• Describe what emojis() does to the HUBModelError message by @raimbekovm in #25538
• Restore the supported-tasks macro and drop the dead model_name sets by @glenn-jocher in #25550
• Render supported tasks per model family by @glenn-jocher in #25551
• Use one canonical task and mode order everywhere by @glenn-jocher in #25552
• Replace predict feature-map dumps with class activation heatmaps by @Y-T-G in #25548
• Reduce Annotator.masks in row bands to cut mask plotting time and memory by @JESUSROYETH in #25554
• Fix Heatmap overlay disappearing on frames without tracks by @JESUSROYETH in #25556
• Fix YOLOE-26 prompt-free RKNN export crash by @zgh2022 in #25536
• Fix RKNN export by pinning package versions by @lakshanthad in #25529
• Fix RKNN INT8 exports returning all-zero class scores by @JESUSROYETH in #25524
• Fix ECC global motion compensation warping against the first frame by @JESUSROYETH in #25555
• replace the set_logging stream handler instead of accumulating one per call by @raimbekovm in #25541
• Add Dragonwing IQ-8275 QNN export target by @glenn-jocher in #25558

New Contributors

• @songjiahao-wq made their first contribution in #25317
• @zgh2022 made their first contribution in #25536

Full Changelog: v8.4.112...v8.4.113