はじめる

Ultralytics YOLOv5

概要#

YOLOv5uは、物体検出向けに更新されたUltralytics YOLOv5です。Ultralyticsが開発したYOLOv5モデルのアーキテクチャを基盤とし、YOLOv5uはアンカーフリーかつobjectnessフリーの分離ヘッドを統合しています。このヘッドは、最初にYOLOv8モデルで導入されました。この変更により、精度と速度のトレードオフが改善され、研究と実用の両方に適した効率的な選択肢となっています。

YOLOv5物体検出モデルのアーキテクチャとパフォーマンス

オリジナルのYOLOv5リポジトリでトレーニングしたモデルは、Ultralyticsライブラリと互換性がありません

Ultralyticsは、YOLOv5モデルのアンカーフリーバリアントを提供しています。ultralytics/yolov5リポジトリで学習したモデルは、ultralytics/ultralyticsライブラリでは読み込めません。ここでYOLOv5を使用するには、Ultralytics YOLOv5uのチェックポイントから新しいモデルを学習してください(例:yolov5su.pt)。

Ultralytics Platformで試す

Ultralytics PlatformでYOLOv5モデルを直接確認し、実行できます。

Input
Upload an image or select an example to test yolov5nu. Predict Docs
Live Demo

Drop an image here

or click to browse

Supports JPEG, PNG, WebP, AVIF, HEIC, JP2, TIFF, BMP & more (max 10MB)

Confidence
0.25
IoU
0.50
Image Size
640
Prediction results will appear here.

主な機能#

  • アンカーフリーのUltralytics分離ヘッド: 従来の物体検出モデルは、物体の位置を予測するために事前定義されたアンカーボックスに依存します。一方、YOLOv5uはアンカーフリーのUltralytics分離ヘッドを使用し、より柔軟な検出メカニズムを実現して、多様なシナリオでの性能を向上させます。

  • 最適化された精度と速度のトレードオフ: YOLOv5uは速度と精度のバランスを取り、競争力のある精度でリアルタイム検出を実現します。これは、自動運転車、ロボティクス、リアルタイム動画分析など、迅速な応答が求められるアプリケーションで有用です。

  • 多様な事前学習済みモデル: YOLOv5uでは、入力解像度640の5種類のサイズ(n、s、m、l、x)のCOCO事前学習済みモデルに加え、入力解像度1280のP6バリアントを提供しています。ハードウェアに適した速度と精度のバランスを選び、独自のデータでファインチューニングできます。

サポート対象のタスクとモード#

さまざまな事前学習済み重みを持つYOLOv5uモデルは、物体検出と、開発からデプロイまでに必要なすべてのモードに対応しています。

モデルタイプ事前学習済み重みタスクトレーニング検証推論エクスポート
YOLOv5uyolov5nu, yolov5su, yolov5mu, yolov5lu, yolov5xu, yolov5n6u, yolov5s6u, yolov5m6u, yolov5l6u, yolov5x6u物体検出✅✅✅✅

この表では、YOLOv5uのモデルバリエーションと、学習、検証、推論、エクスポートの各モードへの対応状況をまとめています。

性能指標#

性能

これらのモデルをCOCOでトレーニングした使用例については検出ドキュメントをご覧ください。これらのモデルには80個の事前学習済みクラスが含まれています。COCOについてもご確認ください。

モデルYAMLサイズ
(ピクセル)
mAPval50-95速度
CPU ONNX
(ms)
速度
A100 TensorRT
(ms)
パラメーター
(M)
FLOPs
(B)
yolov5nu.ptyolov5n.yaml64034.373.61.062.67.7
yolov5su.ptyolov5s.yaml64043.0120.71.279.124.0
yolov5mu.ptyolov5m.yaml64049.0233.91.8625.164.2
yolov5lu.ptyolov5l.yaml64052.2408.42.5053.2135.0
yolov5xu.ptyolov5x.yaml64053.2763.23.8197.2246.4
yolov5n6u.ptyolov5n6.yaml128042.1211.01.834.331.3
yolov5s6u.ptyolov5s6.yaml128048.6422.62.3415.397.6
yolov5m6u.ptyolov5m6.yaml128053.6810.94.3641.2261.3
yolov5l6u.ptyolov5l6.yaml128055.71470.95.4786.1549.4
yolov5x6u.ptyolov5x6.yaml128056.82436.58.98155.51002.8

使用例#

この例では、YOLOv5のシンプルな学習と推論の例を紹介します。

例

PyTorchの事前学習済み*.ptモデルと構成*.yamlファイルをYOLO()クラスに渡すことで、Pythonでモデルインスタンスを作成できます。

from ultralytics import YOLO

# COCOで事前トレーニングされたYOLOv5nuモデルを読み込みます
model = YOLO("yolov5nu.pt")

# モデル情報を表示(任意)
model.info()

# COCO8サンプルデータセットでモデルを100エポック学習する
results = model.train(data="coco8.yaml", epochs=100, imgsz=640)

# 'bus.jpg'画像に対してYOLOv5nuモデルで推論を実行します
results = model("path/to/bus.jpg")

引用と謝辞#

Ultralytics YOLOv5の論文

モデルが急速に進化しているため、UltralyticsはYOLOv5に関する正式な研究論文を発表していません。静的なドキュメントの作成よりも、技術の進歩と使いやすさの向上に注力しています。YOLOアーキテクチャ、機能、使用方法に関する最新情報については、GitHubリポジトリおよびドキュメントをご覧ください。

研究でYOLOv5またはYOLOv5uを使用する場合は、次のようにUltralytics YOLOv5リポジトリを引用してください。

引用
@software{yolov5,
  title = {Ultralytics YOLOv5},
  author = {Glenn Jocher},
  year = {2020},
  version = {7.0},
  license = {AGPL-3.0},
  url = {https://github.com/ultralytics/yolov5},
  doi = {10.5281/zenodo.3908559},
  orcid = {0000-0001-5950-6979}
}

YOLOv5モデルは、AGPL-3.0ライセンスおよびEnterpriseライセンスで提供されていることにご注意ください。

よくある質問#

  • Ultralytics YOLOv5uは、YOLOv5の更新版です。アンカーフリーかつオブジェクトネスフリーの分割型ヘッドを統合し、リアルタイム物体検出タスクにおける精度と速度のトレードオフを改善しています。従来のYOLOv5とは異なり、YOLOv5uはアンカーフリーの検出メカニズムを採用しているため、さまざまなシナリオに柔軟に適応できます。

  • YOLOv5uのアンカーフリーのUltralyticsヘッドは、事前定義されたアンカーボックスへの依存をなくすことで、物体検出パフォーマンスを向上させます。これにより、さまざまなサイズや形状のオブジェクトをより効率的に処理できる、柔軟で適応性の高い検出メカニズムを実現します。この機能強化は精度と速度のバランスに直接貢献し、YOLOv5uをリアルタイムアプリケーションに適したモデルにしています。アーキテクチャの詳細については、主な機能をご覧ください。

  • はい。事前学習済みYOLOv5uモデルは、トレーニング、検証、推論、エクスポートの各モードで物体検出に対応しています。

  • YOLOv5uモデルの性能指標は、使用するプラットフォームとハードウェアによって異なります。たとえば、YOLOv5nuモデルはCOCOデータセットで34.3 mAPを達成し、CPU(ONNX)では73.6 ms、A100 TensorRTでは1.06 msの速度です。CPU ONNXとA100 TensorRTの速度を含む、すべてのYOLOv5uモデルの詳細な指標については、性能指標のセクションをご覧ください。

  • 事前トレーニング済みモデルを読み込み、データセットを指定してトレーニングコマンドを実行すると、YOLOv5uモデルをトレーニングできます。簡単な例を紹介します。

    例
    from ultralytics import YOLO
    
    # COCOで事前トレーニングされたYOLOv5nuモデルを読み込みます
    model = YOLO("yolov5nu.pt")
    
    # モデル情報を表示(任意)
    model.info()
    
    # COCO8サンプルデータセットでモデルを100エポック学習する
    results = model.train(data="coco8.yaml", epochs=100, imgsz=640)

    詳しい手順については、使用例をご覧ください。

コメント30

質問したり、結果を共有したり、ほかの人をサポートしたりできます。

  1. PriyaRanjanReddyGitHub

    Hi can we use yolov5nu.pt for training on my own dataset and then use the trained model for commercial use where I can build an application over it

    1. Glenn JocherUltralytics

      @PriyaRanjanReddy yes, you can use YOLOv5nu.pt for training on your dataset. For commercial use, you'll need an Ultralytics Enterprise License, which allows proprietary deployment without AGPL-3.0's open-source requirements. Visit https://ultralytics.com/license for licensing details.

    2. Glenn JocherUltralytics(編集済み)

      @PriyaRanjanReddy Ultralytics models, including YOLOv5u, YOLOv8, YOLO11 and YOLO26, are released under AGPL-3.0. You can use them commercially if your application complies with AGPL-3.0 (including open-sourcing it); otherwise you need an Ultralytics Enterprise License. Pricing varies based on your project's needs, so please visit our licensing page to discuss details.

  2. donia444GitHub

    hey ,i am trying to train yolov5nu on a custom dataset that is specifically for drone detection ,every time i train using this model ,i notice that my model overfits and gives a very very high unreal accuracy , but when i train on a different model like yolov5n that problem seem to not be found ,any idea why is that happening

    1. Glenn JocherUltralytics

      @donia444 it seems that your YOLOv5nu model might be overfitting due to its more complex architecture compared to YOLOv5n, which is simpler and might generalize better on your dataset. Consider using techniques like data augmentation, dropout, or early stopping to mitigate overfitting. You can find additional strategies in the Ultralytics documentation to help address this issue.

  3. MadhavKrishna1GitHub

    1.Does the kaggle dataset have a label file also? 2.How can i fine tune the model ,is there any built in library through which i could fine tune it.

    On Sun, 22 Sept 2024 at 13:12, Glenn Jocher @.***> wrote:

    @MadhavKrishna1 https://github.com/MadhavKrishna1 yOLOv5 models are indeed trained on the full COCO dataset, which includes 80 classes. For additional datasets similar to COCO, you might explore open-source repositories or platforms like Kaggle. Fine-tuning with your specific images should enhance performance on your target data.

    — Reply to this email directly, view it on GitHub https://github.com/orgs/ultralytics/discussions/10054#discussioncomment-10716632, or unsubscribe https://github.com/notifications/unsubscribe-auth/BCVBHHUR2WRPYUK46RYYFGTZXZYHRAVCNFSM6AAAAABGG3IMDCVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTANZRGY3DGMQ . You are receiving this because you were mentioned.Message ID: @.*** com>

    1. Glenn JocherUltralytics

      @MadhavKrishna1 1. Yes, Kaggle datasets often include label files, but it's best to check the specific dataset details on Kaggle. 2. You can fine-tune the model using the Ultralytics YOLO Python API, which provides built-in support for training and fine-tuning.

  4. MadhavKrishna1GitHub

    I would like to know about the training process of the YOLOv5 model. Specifically, is it trained on the entire COCO dataset? I am planning to fine-tune a downloaded YOLOv5 model using additional images that are similar to those in the COCO dataset. These images have 80 classes and similar sizes to those in COCO Could you provide guidance on finding this additional dataset?

    Thank you!

    1. Glenn JocherUltralytics

      @MadhavKrishna1 yOLOv5 models are indeed trained on the full COCO dataset, which includes 80 classes. For additional datasets similar to COCO, you might explore open-source repositories or platforms like Kaggle. Fine-tuning with your specific images should enhance performance on your target data.

    1. Glenn JocherUltralytics(編集済み)

      @Abhinay-Sadineni yes, yolov5su.pt is an Ultralytics YOLOv5u checkpoint and loads directly with the YOLO class from the ultralytics package, so ensure you have the latest version installed. Checkpoints trained with the original ultralytics/yolov5 repo are the ones that cannot be loaded this way. For more details, visit the YOLOv5 documentation.

    2. chef-SAbhinayGitHub

      is this yolov5su.pt(yolov5 model) compatible with latest ultralytics package ?

      my experiment setup doesn't allow the autoinstall 'ultralytics.yolo' (i mean external traffic is blocked)

      @glenn-jocher

      will the latest package also issue autoinstall warning?

    3. Glenn JocherUltralytics

      To use yolov5su.pt with the latest Ultralytics package, ensure you have the latest version installed. If your setup blocks external traffic, you can manually download the model and dependencies. The autoinstall warning may still appear if dependencies are missing. For a seamless experience, consider using the model with the Ultralytics YOLO class after ensuring all dependencies are locally available.

    4. kanagarajan0709GitHub

      I built a custom dataset from Ultralytics Hub platform using Yolov5su model type and downloaded the trained model. When I tried to load getting this error : "An error occurred: 'Detect' object has no attribute 'grid'. "

      This is how I load the model : yolov5path - points to latest yolov5 git repo. model path - refers to YOLOv5su model trained using Ultralytics Hub.

      yolo_model = torch.hub.load(yolov5path, 'custom', force_reload=True, path=model_path, source='local')

      YOLOv5 v7.0-338-gff063284 Python-3.12.4 torch-2.3.1+cpu CPU

      YOLOv5s summary (fused): 193 layers, 9,114,632 parameters, 0 gradients, 23.8 GFLOPs Adding AutoShape... An error occurred: 'Detect' object has no attribute 'grid'.

    5. Glenn JocherUltralytics(編集済み)

      @kanagarajan0709 yolov5su.pt is an Ultralytics YOLOv5u checkpoint and loads with the YOLO class from the latest ultralytics package. The "Detect object has no attribute 'grid'" error appears when a checkpoint from the original ultralytics/yolov5 repo is loaded with ultralytics, or when yolov5su.pt is loaded with the old yolov5 repo code. Load yolov5su.pt with from ultralytics import YOLO, and use the yolov5 repo only for checkpoints trained there. See the YOLOv5 documentation.

    6. kanagarajan0709GitHub(編集済み)

      To resolve this, please ensure you are using the correct version of the YOLOv5 repository that supports the yolov5su.pt model. If the issue persists, consider using the torch.hub.load method with the appropriate YOLOv5 version.

      Yes , that's what I would be interested to know about. Could you please let me know which Yolov5 git hub repo that supports loading of yolov5su model.

      I am currently on latest git version of this - https://github.com/ultralytics/yolov5. Last commit e8a30cf820b273afbbf1c73a3f1088471a84fade (HEAD -> master, origin/master, origin/HEAD) Author: Glenn Jocher glenn.jocher@ultralytics.com

      Also I did use torch.hub.load method to load the model as below , yolo_model = torch.hub.load(yolov5path, 'custom', force_reload=True, path=model_path, source='local')

    7. Glenn JocherUltralytics(編集済み)

      To load yolov5su.pt, use the YOLO class from the latest ultralytics package (YOLO("yolov5su.pt")). It is an Ultralytics YOLOv5u checkpoint and is not meant for the ultralytics/yolov5 repo, which loads checkpoints trained there.

  5. Halafz7GitHub

    Good Content! I want to ask something. Is it possible to use yolov5u for Classification purposes?

    1. Paula DerrengerUltralytics(編集済み)

      @Halafz7 thank you for your kind words! 😊

      YOLOv5u is designed for object detection. For classification, use a classification model such as YOLO26, which supports it directly.

      Python Example
      from ultralytics import YOLO
      
      # Load a pretrained YOLO26 classification model
      model = YOLO("yolo26n-cls.pt")
      
      # Run inference on an image
      results = model("path/to/image.jpg")
      
      # Display the results
      results[0].show()
      CLI Example
      # Run inference with a YOLO26 classification model
      yolo predict model=yolo26n-cls.pt source=path/to/image.jpg

      For more details, see the Classify docs.

      If you have any more questions or need further assistance, feel free to ask! 🚀

  6. MikolajBaranskiGitHub

    My questions (as of 12th April 2024) is whether there is any documentation regarding the Yolov5u model. The description on this page states that the Yolov5u incorporates the objectness-free split head into the Yolov5 architecture. Is there any further documentation on this model? The Yolov5 GitHub doesn't seem to present this model anywhere. While the page of v7.0 update only describes the introduction of segmentation. Is there any resource going into more detail into this specific release of the model?

    I see there is also documentation regarding the architecture of the Yolov5 v6.1/6.0 model (link below for reference). Is the Yolov5u an adaptations of that architecture? If so which part is changed?

    Github: https://github.com/ultralytics/yolov5 Architecture: https://docs.ultralytics.com/yolov5/tutorials/architecture_description/

    1. Paula DerrengerUltralytics

      Hey there! 😊 It looks like you're exploring the details of our YOLOv5u model. Indeed, YOLOv5u builds upon the YOLOv5 architecture, especially by incorporating an anchor-free, objectness-free split head, a feature inspired by advancements in YOLOv8. This enhancement is aimed at refining the model's ability to detect objects with a better accuracy-speed tradeoff.

      Although it seems there might not be detailed standalone documentation on the YOLOv5u model in the Ultralytics GitHub repo, the key updates and features you’re interested in are described in our Ultralytics YOLOv5 Overview. This page highlights the integration of the anchor-free split head into the model and its impact on performance and adaptability in various scenarios.

      Regarding your question on the architecture adaptation, yes, YOLOv5u adapts the YOLOv5 v6.1/6.0 architecture. The primary change lies in the detection head of the model, which has moved towards an anchor-free, objectness-free approach. This move enhances flexibility and detection efficiency, especially in complex scenarios where predefined anchor boxes might limit the detection capabilities.

      If you're looking for more technical specifics or examples, I'd recommend diving into the architecture description you mentioned. While it focuses on an earlier version, the core architectural principles remain relevant, with the YOLOv5u modifications centered around the detection head for improved performance.

      For direct experimentation or further exploration of the YOLOv5u's unique features, playing with the pre-trained models and reviewing their performance metrics might also offer valuable insights.

      I hope this clarifies your queries! If you have more questions or need further assistance, feel free to ask. Happy detecting! 🚀

    2. MikolajBaranskiGitHub

      Thank you that is a very helpful answer! Just to make a final clarification. I am using the YOLOv5u in my Masters Thesis and I would like to appropriately record the model I have used. For the YOLOv5xu.pt (the pretrained model), should I refer to it as the extra large version of the YOLOv5u v1.0, or is it the YOLOv5 v7.0, or is there some other versioning that would correctly identify this specific release of the model?

    3. Paula DerrengerUltralytics

      Hey, thanks for reaching out with your question! 😊

      For your Masters Thesis, when documenting the model, it's best to refer to it as YOLOv5u, specifying the version as YOLOv5 v7.0 if that's the version you're utilizing, along with the specific variant you're using, for instance, yolov5xu.pt for the extra-large model. It's key to mention both the overall version (v7.0, as an example) and the specific model variant (xu) to provide clear context and specificity about the architecture and features employed in your work.

      Here's how you could cite it:

      "I utilized the YOLOv5u extra-large variant (yolov5xu.pt) from the YOLOv5 v7.0 release for object detection tasks in my project."

      This approach ensures that your documentation is clear, precise, and tracks back to the correct version of the model relevant to your study.

      Hope this helps, and all the best with your thesis! 🚀 If there's anything more you need, feel free to ask.