--- layout: single title: "Ultralytics YOLOv8" date: 2023-03-24 08:00:00 +0800 categories: [框架与库, 容器与云原生] tags: [YOLO, Comet, ClearML, Ultralytics] --- ## Ultralytics * [Ultralytics YOLOv8](https://github.com/ultralytics/ultralytics) * [Ultralytics YOLOv8 Docs](https://docs.ultralytics.com) ## 构建环境 ### [Ultralytics 镜像](https://hub.docker.com/r/ultralytics/ultralytics) * GPU ```shell docker pull ultralytics/ultralytics:latest ``` * CPU ```shell docker pull ultralytics/ultralytics:latest-cpu ``` * Apple Silicon ```shell docker pull ultralytics/ultralytics:latest-arm64 ``` ### 本地安装 ```shell pip install ultralytics ``` ## 基于 COCO128 数据集的目标检测范例 ### 运行容器 ```shell git clone https://github.com/ultralytics/ultralytics.git docker run --runtime=nvidia -it --name ultralytics -v `pwd`/ultralytics:/usr/src/ultralytics ultralytics/ultralytics:latest ``` ### yolo 命令的使用参数 ``` yolo TASK MODE ARGS ``` * [yolo Usage](https://docs.ultralytics.com/usage/cfg/) ### 训练模型 ```shell yolo train data=coco128.yaml model=yolov8n.pt ``` ### 训练可视化（Comet） ```shell pip install comet_ml export COMET_API_KEY=xxx yolo train data=coco128.yaml model=yolov8n.pt project=coco128 ``` * project 指定项目名字，在 Comet 中会自动生成，默认使用的：yolov8 [![](https://camo.githubusercontent.com/84f0493939e0c4de4e6dbe113251b4bfb5353e57134ffd9fcab6b8714514d4d1/68747470733a2f2f636f6c61622e72657365617263682e676f6f676c652e636f6d2f6173736574732f636f6c61622d62616467652e737667)](https://colab.research.google.com/drive/1RG0WOQyxlDlo5Km8GogJpIEJlg_5lyYO?usp=sharing) ### ClearML * 安装 ```shell pip install clearml ``` * 配置 ```shell clearml-init ``` * 集成 ```py from clearml import Task task = Task.init(project_name="my project", task_name="my task") ``` ### 评估模型 ```shell yolo val data=coco128.yaml model=runs/detect/train/weights/best.pt project=coco128 ``` ### 预测 ```shell yolo predict data=coco128.yaml model=runs/detect/train/weights/best.pt source=project/images/val/ save=true ``` * save=true 保存预测结果（runs/detect/predict） * show=true 展示预测结果（UI Window） 对检测出来的结果裁剪分类保存 ```shell yolo predict data=project/data.yaml model=runs/detect/train/weights/best.pt source=project/images/val/ save_crop=true ``` * classes=0 或 classes=[0,2,5] 过滤指定的类别 ### 导出模型 ```shell yolo export model=yolov8s.pt format=onnx ``` 导出的模型格式 | Format | format Argument | Model | Metadata | | ------------- | --------------- | ----------------------- | :------: | | PyTorch | | yolov8n.pt | ✅ | | TorchScript | torchscript | yolov8n.torchscript | ✅ | | ONNX | onnx | yolov8n.onnx | ✅ | | OpenVINO | openvino | yolov8n_openvino_model/ | ✅ | | TensorRT | engine | yolov8n.engine | ✅ | | CoreML | coreml | yolov8n.mlmodel | ✅ | | TF SavedModel | saved_model | yolov8n_saved_model/ | ✅ | | TF GraphDef | pb | yolov8n.pb | ❌ | | TF Lite | tflite | yolov8n.tflite | ✅ | | TF Edge TPU | edgetpu | yolov8n_edgetpu.tflite | ✅ | | TF.js | tfjs | yolov8n_web_model/ | ✅ | | PaddlePaddle | paddle | yolov8n_paddle_model/ | ✅ | ## 基于 mnist160 数据集的分类 ```shell yolo classify train data=mnist160 model=yolov8n-cls.pt epochs=100 imgsz=64 yolo classify val data=mnist160 model=runs/classify/train/weights/best.pt ``` ## [Benchmark](https://docs.ultralytics.com/modes/benchmark/) ### GPU ```shell yolo benchmark model=yolov8n.pt imgsz=640 half=False device=0 ``` ``` Ultralytics YOLOv8.0.58 🚀 Python-3.10.9 torch-2.0.0 CUDA:0 (Tesla T4, 15110MiB) Server ✅ (64 CPUs, 251.6 GB RAM, 381.2/548.6 GB disk) Benchmarks complete for yolov8n.pt on coco128.yaml at imgsz=640 (964.68s) Format Status❔ Size (MB) metrics/mAP50-95(B) Inference time (ms/im) 0 PyTorch ✅ 6.2 0.4478 13.39 1 TorchScript ✅ 12.4 0.4525 5.33 2 ONNX ✅ 12.2 0.4525 10.97 3 OpenVINO ❌ 0.0 NaN NaN 4 TensorRT ✅ 16.1 0.4525 4.80 5 CoreML ❌ 0.0 NaN NaN 6 TensorFlow SavedModel ✅ 30.6 0.4525 90.53 7 TensorFlow GraphDef ✅ 12.3 0.4525 80.57 8 TensorFlow Lite ❌ 0.0 NaN NaN 9 TensorFlow Edge TPU ❌ 0.0 NaN NaN 10 TensorFlow.js ❌ 0.0 NaN NaN 11 PaddlePaddle ✅ 24.4 0.4525 334.54 ``` ### CPU ```shell yolo benchmark model=yolov8n.pt imgsz=640 half=False device=cpu ``` ``` Ultralytics YOLOv8.0.58 🚀 Python-3.10.9 torch-2.0.0 CPU Setup complete ✅ (64 CPUs, 251.6 GB RAM, 383.1/548.6 GB disk) Benchmarks complete for yolov8n.pt on coco128.yaml at imgsz=640 (647.94s) Format Status❔ Size (MB) metrics/mAP50-95(B) Inference time (ms/im) 0 PyTorch ✅ 6.2 0.4478 186.26 1 TorchScript ✅ 12.4 0.4525 231.86 2 ONNX ✅ 12.2 0.4525 61.08 3 OpenVINO ✅ 12.3 0.4525 21.58 4 TensorRT ❌ 0.0 NaN NaN 5 CoreML ❎ 12.1 NaN NaN 6 TensorFlow SavedModel ✅ 30.6 0.4525 86.86 7 TensorFlow GraphDef ✅ 12.3 0.4525 80.34 8 TensorFlow Lite ✅ 12.2 0.4525 190.41 9 TensorFlow Edge TPU ❌ 40.9 NaN NaN 10 TensorFlow.js ❌ 12.3 NaN NaN 11 PaddlePaddle ❌ 0.0 NaN NaN ``` ## 工具 ### 标注数据 * [labelImg](https://github.com/heartexlabs/labelImg) * [Label Studio](https://github.com/heartexlabs/label-studio) ## 参考资料 * [Train Custom Data](https://github.com/ultralytics/yolov5/wiki/Train-Custom-Data) * [How do I save the images of the yolov8 training predictions?](https://github.com/ultralytics/ultralytics/issues/375) * [YOLO: Real-Time Object Detection](https://pjreddie.com/darknet/yolo/) * [yolov5](https://github.com/ultralytics/yolov5) * [Ultralytics YOLOv8](https://github.com/ultralytics/ultralytics) * [Change 'Save' location for prediction #1805](https://github.com/ultralytics/ultralytics/issues/1805) * [Metal](https://developer.apple.com/metal/) * [ONNX Runtime](https://onnxruntime.ai/) * [YOLOv8 深度详解！一文看懂，快速上手](https://zhuanlan.zhihu.com/p/598566644) * [Train YOLOv8 on Custom Dataset – A Complete Tutorial](https://learnopencv.com/train-yolov8-on-custom-dataset/) * [ultralytics/JSON2YOLO](https://github.com/ultralytics/JSON2YOLO) * [YOLOv8 Ultralytics: State-of-the-Art YOLO Models](https://learnopencv.com/ultralytics-yolov8/) * [Python Design Patterns - Singleton](https://www.tutorialspoint.com/python_design_patterns/python_design_patterns_singleton.htm) * [How to implement a subscriptable class in Python (subscriptable class, not subscriptable object)?](https://stackoverflow.com/questions/11469025/how-to-implement-a-subscriptable-class-in-python-subscriptable-class-not-subsc)