登录

geti

open-edge-platformApache-2.0活跃维护

本地完成计算机视觉模型从标注、训练到优化部署的全流程，支持增量学习与模型压缩，针对 Intel XPU 加速并兼容 CUDA。

Build, train, optimize, and run computer vision models locally, from raw images to live inference. Open source, optimized for Intel® XPU (CUDA also supported).

app/v3.3.0发布于 2026-10-26 15:34 UTC1,370star +1 / 7 天489fork同步于 2026-11-10 11:48 UTC
计算机视觉增量学习模型压缩本地训练OpenVINO
对比 GitHub 官网automlcomputer-visiondeep-learninggetiimage-classification

怎么上手

由 README 提炼，执行前请确认
原文 
uv pip install getitune # CPU-only by default
uv pip install "getitune[xpu]" --extra-index-url https://download.pytorch.org/whl/xpu
硬件要求按 README 说明

README说明优化支持Intel CPU、GPU、NPU，同时支持NVIDIA CUDA；getitune可纯CPU安装运行，具体显存要求README未说明。

纯 CPU 可运行支持 Apple Silicon需要 CUDA
下载：源码 zip·GitHub Releases（安装包）本站不镜像安装包

项目说明

由 README 自动翻译提炼
没人工点评？写下你的经验

端到端计算机视觉模型平台，从标注、微调到优化和本地推理一站式完成

本地完成计算机视觉模型从标注、训练到优化部署的全流程，支持增量学习与模型压缩，针对 Intel XPU 加速并兼容 CUDA。

核心能力

  • 覆盖标注、微调、优化、部署、推理全流程
  • 内置1600+检测分类分割模型变体
  • 用OpenVINO导出并量化INT8以适配Intel硬件
  • 提供Windows应用与Docker镜像的无代码界面
  • getitune Python库支持代码化训练与导出

容易踩的坑

Ultralytics YOLO模型默认不包含在包内，需额外指定ultralytics extra安装；CUDA或XPU加速需按对应extra并加PyTorch额外索引源安装；Windows应用与Docker安装方式不同，需按安装指南选择。

内容由模型根据 README 生成，可能不完整或不准确，以官方文档为准。看原文 README

最近发版

5 个版本app/v3.3.0
全部 release
  1. app/v3.3.0Geti v3.3.0最新正式版2026-10-26 · 2026-10-26 15:34 UTC

    🚀 Geti™ v3.3.0

    Geti™ v3.3.0 focuses on its core machine-learning foundations: it integrates several models from the Hugging Face ecosystem, enables dataset view-based workflows, allows concurrent predictions from multiple models and optimizes memory usage during training.

    ✨ Highlights

    • New SOTA models – including ConvNeXt V2, RT-DETRv2, EoMT, Mask2Former and DINOv3-based variants of several models.
    • Ultralytics YOLO models – are now available for everyone in pre-built Docker images and MSIX binaries images.
    • Dataset views for training and export – selected dataset views can now be used as input for training and export.
    • Multi-lingual support – UI translations are now available for more than 10 languages.
    • Improved memory efficiency – training workflows were optimized to use less memory compared to earlier versions, without sacrificing performance. The improvements are noticeable to all users, especially those running Geti on edge devices.

    Models and training

    • New transformer-based models – some of the best models available in HuggingFace Transformers are now fully supported in Geti™ for training, inference and optimization. All of them are compatible and validated on Intel® XPU.
      • Classification: ConvNeXt V2 (Atto, Base, Large), DINOv3 ConvNeXt (Tiny, Small, Base, Large), DINOv3 ViT (Small, Base, Large)
      • Detection: RT-DETRv2 (R18, R34, R50, R101)
      • Instance segmentation: EoMT (Large), EoMT-DINOv3 (Large), Mask2Former-Swin (Small, Base, Large)
    • Built-in Ultralytics YOLO models – thanks to an agreement with Ultralytics, all YOLO models (YOLO11/12/26) are now included in all Geti's official pre-built Docker images and Geti Windows application (MSIX). This means that users no longer need to build Geti from source to enable and use these models.
    • Train from dataset views – it is now possible to select a dataset view as input for training, as an alternative to the existing options of training on the entire dataset or on a previous dataset revision. This unlocks the ability to fine-tune models on a custom subset of your data, which is useful to conduct narrow-scoped experiments or if you want to train a small model specialized on a certain subset of a very large heterogeneous dataset.
    • Configurable number of workers – the training configuration was extended with a new parameter to control the number of worker processes for data loading. Users with powerful machines can benefit from more workers to speed up data loading and augmentation, while users with constrained devices may use fewer workers to avoid filling up the memory.
    • Consistent subset assignment for video frames – when feasible, all the frames of one video are assigned to the same subset (instead of randomly distributed across subsets) to ensure the three subsets stay as independent as possible.
    • Downloadable metric charts – after training, it is now possible to download individual model metric charts as images.
    • Better diagnostics for failed jobs – the full traceback is now shown in the logs after a training or quantization job fails.

    Projects, datasets and annotation

    • Export dataset views – it is finally possible to export dataset views, including their media and annotations, in any of the supported dataset formats (Geti, COCO, YOLO, VOC). The exported zip archive can be shared and/or later imported to another project, just like any exported dataset.
    • Monitor jobs – a new "Jobs" panel allows you to track the status and progress of training and quantization jobs. It is accessible from any page (top-right corner) and it allows to optionally filter the jobs by project, view logs and cancel individual jobs.
    • Monitor uploads – when uploading media or datasets, the UI now shows a prog
  2. app/v3.2.0Geti v3.2.02026-09-15 · 2026-09-15 14:47 UTC
    查看更新说明

    🚀 Geti™ v3.2.0

    Geti 3.2 introduces many new classification models and various enhancements for dataset management and inference, as well as stability and performance improvements.

    To upgrade from an existing Geti™ 3.x deployment to 3.2, follow the upgrade guide.

    ✨ Highlights

    • Pytorch Image Models (timm) – 1600+ classification models from timm are now available in Geti™ and ready for fine-tuning, optimization and inference.
    • Dataset views – organize and filter your dataset flexibly by creating dataset views, custom subsets where you have full control over the assignment of dataset items.
    • Configurable confidence threshold – adjust the confidence threshold when requesting predictions in the annotator and when configuring a model in the inference pipeline.
    • Quick ‘Annotation vs prediction’ comparison – view the predictions generated by a trained model while remaining in the ‘Models’ page (‘Training datasets’ tab) and compare them with the respective annotations in one click.
    • More reliable ‘Auto-segmentation’ – this annotation tool is now more stable and responsive thanks to several optimizations and fixes.

    🗂️ Projects, datasets and annotation

    • Dataset views – the ‘Dataset’ page now offers the possibility to create and manage dataset views. Think of a view as a way to define a custom subset within your dataset, grouping media that have a specific property in common (e.g. photos taken at night). You can freely assign (and unassign) media to one or more dataset views, rename them, and of course filter the dataset by view. In a future version, Geti™ will also allow you to use dataset views as input for various workflows such as model fine-tuning and dataset export.
    • Configurable confidence threshold (annotator) – in the ‘Prediction’ tab you can now adjust the threshold value used to filter predictions based on their confidence score. Lowering the value allows you to see more predictions, while increasing it may help sometimes to filter out false positives. This option is available for all detection and segmentation models.
    • Improved project list layout – the ‘Project’ screen was optimized to display more projects in the available area.
    • Inference settings icon - in the ‘Prediction’ tab, all inference-related parameters have been grouped under a cogwheel icon.
    • Bulk selection of media - in the ‘Dataset’ page, use "CTRL + Left click" to add an item to the selection, and "Shift + Left click" to select a range of items.
    • More annotator hotkeys – Use “Left” and “Right” arrows to navigate between media while in the annotator, “Enter” to submit annotations and “Tab” to cycle through different shapes.

    🧠 Models and training

    • Pytorch Image Models (timm) – the classification task now offers more than 1600 models thanks to the integration with ‘timm’, a large collection of SOTA models and weights across 80+ architecture families. You can find these models listed under the “Other models (timm)” card in the training dialog for single-label and multi-label classification, along with a special selector to choose the backbone and variant. All models can be fine-tuned, optimized, used for inference and exported. Note that Geti™ only exposes ‘timm’ models with a source code license that is compatible with Apache 2.0; at the same time, license of the respective pretrained weights is displayed in the UI so that you can pick a suitable model for your use case.
    • Annotation vs prediction – in the ‘Models’ page, when opening any image from the ‘Training datasets’ tab, you can now switch between the ‘Annotation’ and
  3. app/v3.1.0Geti v3.1.02026-08-13 · 2026-08-13 14:40 UTC
    查看更新说明

    🚀 Geti™ v3.1.0 - new models, improved pipeline reliability and much more!

    Geti™ 3.1 builds on the same lightweight edge-native architecture of Geti™ 3.0, extending it with several new SOTA models while also improving dataset navigation, inference pipeline monitoring, quantization and other core features.

    This release also extends the documentation and it updates some key dependencies (including PyTorch and OpenVINO™) for better performance, stability and security.

    To upgrade from an existing Geti™ 3.0 deployment to 3.1, follow the upgrade guide.

    🔷 Highlights

    • Broader model zoo - enabled the new EdgeCrafter (ECDet) architecture and more RF-DETR variants, giving users more flexibility to choose the best model for their accuracy and speed requirements.
    • Filter datasets by subset - the Dataset page and the annotator now offer the possibility to filter media by a specific subset (training, validation or testing).
    • Client-provided video files as pipeline sources - upload a video straight from your local filesystem when configuring a source for the inference pipeline, without needing the file to be available in the server beforehand.
    • Pipeline observability – it is now easier to monitor the status of the inference pipeline, by means of a health indicator (with fine-grained status info reported for each component) and a one-click 'Test connection' button for sources and sinks.
    • Increased UI responsiveness – thanks to HTTP/2, concurrent API requests are handled more efficiently, improving the general responsiveness of the application.

    🧠 Models and training

    • EdgeCrafter for object detection - the new ECDet (a.k.a. EdgeCrafter) model is now available and fully supported for training, quantization and inference in Geti™. ECDet is available in different size variants: S, M, L, and X.
    • New size variants of RF-DETR were added for object detection (N) and instance segmentation (N, 2XL), in addition to the ones already present since 3.0.
    • Tiling - All detection and instance segmentation models now support tiling for training and inference, closing the gap for some models that didn't already support it (e.g. D-FINE, RF-DETR, ...)
    • Iteration limit for accuracy-aware quantization - a new parameter is available to optionally cap the number of iterations during accuracy-aware quantization, preventing it from running indefinitely.

    🗂️ Projects, datasets and annotation

    • Project filtering - projects can be searched by name and filtered by task type.
    • Pinned active project - any project with a running inference pipeline is now pinned to the top of the list and tagged as "Active" badge for better visibility.
    • Annotator media filtering - all media filters are now also available inside the annotator, making it much faster to navigate between images and videos without leaving the annotator.
    • Subset filtering - a new filter is available to select media assigned to a specific subset (training, validation, testing or unassigned).

    🚀 Deployment and inference pipeline

    • Video upload for source creation - it is now possible to create a new source of type 'Video file' by uploading a video from your local filesystem. Previously, the video needed to be already available in the server beforehand, which is not a practical setup for many users.
    • Test connection - sources and sinks can be validated before use through the new 'Test connection', with an inline connection-status badge reflecting the result.
    • Pipeline health monitor - a new indicator in the 'Pipeline' page aggregates status information about the source, model and sink
  4. app/v3.0.0Geti v3.0.02026-06-18 · 2026-06-18 13:01 UTC
    查看更新说明

    🚀 Geti™ v3.0.0 — AI development moves to the edge

    Geti 3.0 marks a fundamental achievement in the evolution of vision AI development. Geti is now a lightweight application that runs directly on edge devices and AI PC hardware, while still remaining compatible with traditional cloud and cluster‑based infrastructure. With Geti v3 you can fine‑tune models and run inference directly on edge and client hardware — including Intel® Panther Lake and Arc™ Battlemage (B‑series) GPUs — keeping your data, training, and deployment on the same machine.

    The result is a faster, simpler, and edge-native way to go from data to a deployed model: on a laptop, a workstation, or an edge device.

    🔷 Major features and improvements

    A whole new installation experience:

    • Native Windows app — run Geti directly on Windows via a simple MSIX installer.

    • Single Docker image — run Geti from one container on Linux or Windows, while maintaining compatibility with Kubernetes deployments.

    State‑of‑the‑art model architectures

    • RF‑DETR S / M / L (Roboflow) — a real‑time DETR transformer on a DINOv2 backbone, SOTA accuracy/latency on COCO, for detection and instance segmentation.

    • DINOv3 DETR S / M / L — Meta's DINOv3 self‑supervised backbone combined with DEIMv2 + DETR for high‑quality detection with less labeled data.

    A new Ultralytics integration introduces YOLO26 — the latest NMS‑free, edge‑optimized models (Nano / Small / Medium) for object detection and instance segmentation. The integration covers the full model lifecycle: training, inference, quantization, and OpenVINO™ model export.

    Dramatically lower hardware requirements: Geti v3 no longer depends on Kubernetes as the default architecture, the overall footprint is significantly reduced requiring fewer CPU resources and less memory (the minimum recommended configuration is 8 CPU threads, 16 GB RAM, and 40 GB of free disk space)[VJ1.1][YV1.2]. Smaller models can be trained on CPU with modest memory usage, while a GPU is recommended for larger models.

    Apache 2.0 license: Geti 3.0 code is now under the permissive Apache 2.0 license, simplifying adoption, integration, and redistribution.

    More features:

    • Dataset versioning - track how datasets evolve as images and annotations are added, with the ability to explore revisions, link models to a specific dataset state, and train on any prior revision.

    • Model versioning / lineage - track model lineage with links to parent revisions and weights, view exact training hyperparameters, and fine-tune from any previous model version.

    • Configurable training devices - select from available hardware (CPU, GPU), including specific GPUs in multi-device setups.

    • 16‑bit media support - train and infer on 16-bit single-channel images and video with configurable intensity range preprocessing (not supported for YOLOX‑S/M/L).

    • Decoupled edge-ready pipelines - build flexible inference pipelines where data sources (e.g. cameras, files, streams) and output targets can be swapped or reconfigured without retraining models, enabling easier integration and faster iteration across different deployment scenarios.

    • Pipeline export for edge deployment - package complete pipelines, including model, preprocessing, and configuration, into OpenVINO™‑optimized bundles ready for deployment on edge and Intel®‑based hardware.

    • Integrated deployment & inference — a new Inference page builds custom pipelines (source → model → sink) to deploy models inside Geti and monitor predictions in real time. Sources include USB/IP cameras and video files; optional sinks include MQTT, webhook, and folder.

    • Improved evaluation metrics - expanded training-time and evaluation metrics, including mAP at multiple thresholds for detection and segmentation tasks, providing deeper insight i

  5. 2.6.02025-10-13 · 2025-10-13 08:30 UTC
    查看更新说明

    What's Changed

用户点评

登录用户的真实使用记录，可按版本和硬件筛选
共 0 条

    还没有人点评。第一个说出真实体验的人最有价值。

    README 原文

    未翻译
    在 GitHub 打开

    Container build Codecov OpenSSF Scorecard License PyPI version PyPI downloads

    Application: Install • Upgrade • Docs • License Library (getitune): Install • Docs

    Geti™ is an open-source, end-to-end platform for building and continuously improving your own computer vision models. It covers annotation, fine-tuning, optimization, deployment, and inference — with 1,600+ model variants for detection, classification, and segmentation, optimized by OpenVINO™ for Intel® CPUs, GPUs, and NPUs.

    The Geti™ application is powered by getitune, its engine for model training and optimization, which is also available standalone as a Python library. Choose the application for a no-code experience, or the getitune Python API for code-first workflows — both paths lead to production-ready models. Both are developed in this repository, under application/ and libraries/getitune/.

    Full lifecycle with an intuitive SME interface

    The Geti™ application runs as a native Windows app or a pre-built Docker image for Linux.

    For complete, step-by-step instructions - including prerequisites, GPU/accelerator support, container and source builds, the install script, and troubleshooting - see the Installation guide. To update an existing Geti™ installation to a newer version, follow the Upgrade guide.

    [!TIP] The documentation is a valuable resource to learn more about Geti™ and its capabilities.
    New users are encouraged to read the step-by-step tutorial "Train your first model".

    Full lifecycle with a simple Python API

    Geti's training engine is also available on PyPI as getitune, an open-source Python library for training, optimizing, and deploying models.

    uv pip install "getitune[xpu]" --extra-index-url https://download.pytorch.org/whl/xpu    # for Intel® XPU acceleration
uv pip install "getitune[cuda]" --extra-index-url https://download.pytorch.org/whl/cu128    # for NVIDIA® CUDA acceleration
uv pip install getitune # CPU-only by default

    [!IMPORTANT]
    Ultralytics YOLO models are NOT included in the package by default. To install these models, which are distributed under the AGPL-3.0 license, you must specify the ultralytics extra as explained in the getitune documentation.

    Pick a model, then train, export, and quantize it:

    from getitune.engine import create_engine
from getitune.utils import list_models

# Explore available models to pick the one you need
all_models = list_models()                        # All model names
detection_models = list_models(task="DETECTION")  # Filter by task
recipes = list_models(return_recipes=True)        # Full recipe YAML paths

# Choose a model and point it at your dataset
engine = create_engine(
    model="edgecrafter_m",
    data="./led_counting",
    work_dir="./my_workspace",  # checkpoints, exports, and logs
    device="auto",  # "auto", "cpu", "gpu", or "xpu"
)

# Train the model on your own Intel CPU or GPU
engine.train()

# Export the trained model to OpenVINO format (.xml + .bin)
exported_model_path = engine.export()

# Load the exported model into the OpenVINO engine
openvino_engine = create_engine(model=exported_model_path, data=engine.datamodule)

# Quantize the model to INT8, optimized for Intel hardware
optimized_model_path = openvino_engine.optimize()

# Run inference with the optimized INT8 model
predictions = openvino_engine.predict(checkpoint=optimized_model_path)

    See the getitune documentation for the full list of recipes, advanced configuration, dataset support, backend-specific options, and deployment/optimization examples.

    Key Features

    Train and fine-tune modern architectures such as RF-DETR, DINOv3 DETR, YOLO26, YOLOX, D-FINE, and Mask R-CNN. Would you like to see a specific model added? Let us know by opening a GitHub issue!

    Geti™ enables users to start building deep-learning computer vision models with as few as 10-20 images and take them to production in one environment - annotate, train, optimize, run inference, and improve accuracy in a rapid train-predict-annotate loop.

    Every model is automatically exported with OpenVINO™ for deployment across the full Intel® XPU portfolio (Arc™ GPUs, Core™ Ultra processors); NVIDIA® CUDA and CPU-only execution are also supported. Fine-tune and run inference directly on edge and client hardware - including Intel® Panther Lake and Arc™ Battlemage (B-series) GPUs - with no Kubernetes cluster or data-center GPU required. Built-in accuracy-aware INT8 quantization further reduces model size and latency on resource-constrained edge devices with minimal impact on accuracy.

    Build custom pipelines (source → model → sink) to deploy models inside Geti and monitor real-time predictions on video streams. Sources include USB/IP cameras and video files; optional sinks include folder, MQTT, and webhook. Complete pipelines can be exported as OpenVINO™-optimized bundles for edge deployment.

    Geti™ supports multiple computer vision tasks that are commonly employed across various use cases - image classification, object detection and instance segmentation from the no-code web interface, with even more tasks available through the getitune library.

    Smart annotations in Geti™ enable users to easily create bounding boxes and polygons. These smart annotation features coupled with the AI-assisted annotations and state-of-the-art AI models such as the Segment Anything Model keep human experts in the loop while massively reducing the total annotation efforts needed by a human.

    Track how datasets and models evolve, link models to a specific dataset revision, view exact training hyperparameters, and fine-tune from any previous version. Import and export in COCO, Pascal VOC, YOLO, and a Geti-optimized native format, with label filtering to selectively include or exclude labels on import/export.

    Ecosystem

    Explore other open-source AI projects by Intel®:

    • Anomalib - An anomaly detection suite comprising state-of-the-art algorithms and features such as experiment management, hyper-parameter optimization and edge inference.
    • Instant Learn - A framework for developing, benchmarking, and deploying zero-shot visual prompting algorithms on the edge.
    • Datumaro - Dataset Management Framework, a Python library and a CLI tool to build, analyze and manage Computer Vision datasets.
    • OpenVINO™ - Software toolkit for optimizing and deploying deep learning models.
    • OpenVINO™ Model Server - A scalable inference server for models optimized with OpenVINO™.
    • Model API - A set of wrapper classes for particular tasks and model architectures, simplifying data preprocessing and postprocessing as well as routine procedures.
    • Physical AI Studio - An end-to-end framework for teaching robots to perform tasks through imitation learning from human demonstrations.

    Who uses Geti™?

    Geti™ is used by research institutes, industrial partners, universities and AI enthusiasts. Applications range from robotics to medical analysis or industrial quality control. There are some interesting members of the Geti™ community:

    Contribute

    To report a bug or submit a feature request, please open a GitHub issue. If you have an open question, ask in GitHub Discussions.

    For developers who would like to contribute with a pull request, see the Contributing guide for details.

    License

    Geti™ source code is licensed under the Apache License Version 2.0. Docker images contain Geti™ components licensed under Apache 2.0 and may also contain third-party components governed by different license terms, listed separately in this file. The Windows MSIX App is licensed under the Intel Simplified Software License.

    Regarding models trained with Geti™, their license depends on the model type and is indicated within the application; most are licensed under Apache 2.0, while some fall under more restrictive licenses (e.g. Ultralytics AGPL3.0 models).

    It is the end user's responsibility to ensure that the specific use of Geti™ and the models complies with the respective licenses.

    相关项目

    同分类里 star 最高的项目
    对比全部 →
    ultralytics
    ultralytics·Python

    YOLO 系列官方实现库，一套 PyTorch 代码覆盖目标检测、实例与语义分割、图像分类、姿态估计和目标跟踪，训练到部署全流程打通。

    目标检测+4检测 / 分割
    63270+8712kv8.4.2082026-11-10 09:31 UTC
    yolov5
    ultralytics·Python

    基于 PyTorch 的目标检测全家桶，覆盖训练、推理、实例分割、图像分类和模型导出，开箱即用，适合快速落地视觉任务。

    目标检测+4检测 / 分割
    58258+517kv7.02026-11-07 16:59 UTC
    pytorch-image-models
    huggingface·Python

    收录最全的 PyTorch 图像骨干网络合集，涵盖 ResNet、ViT、ConvNeXt 等主流模型，自带训练、评估、推理和导出脚本及预训练权重，省去自己搭网络和找权重的麻烦。

    图像分类+4检测 / 分割
    37272+75.2kv1.0.312026-11-07 23:22 UTC
    detectron2
    facebookresearch·Python

    Detectron2 是 Facebook 开源的目标检测与分割平台，提供模块化架构和丰富预训练模型，方便快速搭建和训练视觉识别任务。

    目标检测+4检测 / 分割
    34980+57.9kv0.62026-10-27 09:21 UTC