pypi ultralytics 8.3.238
v8.3.238 - `ultralytics 8.3.238` Refactor SAM3 forward convolutions (#22942)
on Python PyPI

latest release: 8.3.239
10 hours ago

🌟 Summary

Ultralytics v8.3.238 is a refinement-heavy release that makes SAM3 concept/video segmentation more robust, faster to work with, and easier to install, while also stabilizing model export (especially TFLite/ONNX) and polishing docs & CI workflows. 🧠🧩

📊 Key Changes

  • SAM3 architecture refactors (core of this release) 🧱

    • New shared helper for SAM3 neck feature generation (sam_forward_feature_levels) reduces duplicated code and unifies how multi-scale features + positional encodings are produced.
    • VL combiner and neck now reuse this helper for both SAM3 and optional SAM2 paths, simplifying the forward logic and making behavior more consistent.
    • SAM3 image/video predictors now share backbone feature preparation with SAM2 (_prepare_backbone_features), centralizing multi-prompt batching and reducing redundant computation.

  • SAM3 stability & correctness fixes 🔧

    • Batching bug fix: SAM3SemanticPredictor no longer corrupts its cached backbone features when you query the same image with different numbers of text prompts—fixes shape mismatch crashes in “encode once, query many times” workflows.
    • MPS (Apple Silicon) fix: SAM3 attention now enforces contiguous tensors and simplifies dtype conversions, improving reliability on Metal (MPS) backends.
    • Tokenizer deps auto-install: When using SAM3 semantic/text features, missing Python packages (ftfy, regex, iopath) are auto-checked and prompted for install instead of crashing with ModuleNotFoundError.
    • Cleaned up SAM/SAM2/SAM3 docstrings and docs so predictors are correctly described and imports align with the public API (from ultralytics.models.sam import ...).

  • SAM3 video API & docs improvements 🎥

    • Official docs now use the simplified import path from ultralytics.models.sam import SAM3VideoSemanticPredictor.
    • Clearer SAM3 docs: updated model size (3.4 GB), revised YOLO11 vs SAM3 efficiency comparison, and marked SAM3 as fully available from v8.3.237.

  • Export & deployment reliability 🛫

    • TFLite / ONNX export guards:
      • FP16 conversion for ONNX is now only run when --half and --format=onnx and device=cpu are all true—prevents ONNX-specific code running for other formats and causing avoidable errors.
      • Fixed TFLite export dtype mismatches when combining half=True with nms=True.
    • MNN export safety:
      • MNN slow tests are skipped if torch<1.10, and the exporter now asserts torch>=1.10 for MNN to avoid runtime segmentation faults.
    • ONNX version pin clarified: kept onnx>=1.12.0,<=1.19.1 with comments that this remains until onnx_graphsurgeon supports newer ONNX.
    • Small TensorFlow export patch to better handle helper function availability.

  • CI, tooling & infra 🏗️

    • CI Summary job now depends on SlowTests, and Slack alerts will also trigger if SlowTests fail—making flaky or long-running issues more visible.
    • Docker runner image now runs apt-get update, installs libicu-dev, and clears apt lists to reduce image bloat.
    • GitHub Actions actions/upload-artifact bumped from v5 → v6 (Node 24 runtime).

  • Examples & ecosystem tweaks 🌍

    • .gitignore now allows examples/**/requirements.txt; new example-specific requirements were added (e.g., for ONNXRuntime/OpenCV demos).
    • RT-DETR ONNXRuntime example: safer downloads, better input validation, optional NMS filter, and fixed COCO YAML URL.
    • Interactive tracking UI: safer defaults (e.g., save_video=False by default).

  • Docs & UX polish 📚

    • Large pass standardizing terminology (pretrained vs pre-trained), cleaning grammar, clarifying Explorer’s deprecated status and pointing users to Ultralytics HUB.
    • Refreshed guides for Raspberry Pi, Jetson, Vertex AI, tracking, auto-annotation, and project-planning docs for clearer language and more realistic expectations.
    • Swapped in new YOLO11 videos for VisDrone and MNN export to give users up-to-date tutorials.

🎯 Purpose & Impact

  • More reliable SAM3 workflows for research & production 🧠🚀

    • The SAM3 refactors and bug fixes make “encode once, query many times” safe and efficient, especially when changing text prompts or using multi-prompt batching.
    • Shared backbone feature logic across SAM2/SAM3 reduces implementation divergence, which lowers the risk of subtle bugs and speeds up future maintenance.

  • Smoother out-of-the-box experience for SAM3 users 💡

    • Auto-install prompts for tokenizer dependencies and robust MPS handling mean fewer cryptic crashes and less manual environment wrangling—especially on macOS laptops and new users trying semantic/text prompts.
    • Consistent, public-facing import paths (from ultralytics.models.sam import ...) make it easier to copy examples and reduce breakage if internals move.

  • Safer, more predictable model export across formats 📦

    • Guarding FP16 and TFLite paths prevents surprising export-time failures, especially in mixed environments (CPU-only, different export formats, or older Torch).
    • Clear version pins and runtime checks reduce the “works on my machine” problem when exporting to ONNX, MNN, or TensorFlow.

  • Higher CI signal quality & maintainability ⚙️

    • Including SlowTests in the CI Summary and Slack alerts helps catch real regressions in long-path code (exports, hardware-specific flows) before they reach users.
    • Docker and workflow updates keep the automation stack modern and lean.

  • Clearer documentation for a wide audience 📖✨

    • Standardized wording, updated videos, and more honest model specs (e.g., SAM3’s large size vs YOLO11) give users more realistic expectations about performance, hardware needs, and when to choose which model.
    • Better messaging around deprecated Explorer features and the role of Ultralytics HUB helps new users find the right tools faster.

What's Changed

New Contributors

Full Changelog: v8.3.237...v8.3.238

Check out latest releases or
releases around ultralytics 8.3.238

Don't miss a new ultralytics release

NewReleases is sending notifications on new releases.

Get notifications