zh/modes/val/

[giscus[bot]]

zh/modes/val/

YOLOv8 模型验证指南。通过Python 和CLI 示例，了解如何使用验证设置和指标来评估YOLO 模型的性能。

https://docs.ultralytics.com/zh/modes/val/

[Ameridru]

请问在模型验证中，计算Precision、Recall时默认采用的IOU阈值是0.5吗，能否打印出不同IOU阈值下的Precision和Recall？

[glenn-jocher]

在 YOLOv8 模型验证过程中，默认使用的 IOU 阈值确实是 0.5 (iou=0.5)，用于计算 Precision 和 Recall。如果您想查看不同 IOU 阈值下的 Precision 和 Recall，可以通过修改 iou 参数值来实现。不过，请注意，目前 YOLOv8 验证模式的输出直接在 CLI 或 Python API 中不提供不同 IOU 阈值的详细数据。您可能需要自定义处理结果数据或使用其他工具来实现这一需求。

在 Python 中，您可以这样设置不同的 IOU 阈值：

from ultralytics import YOLO

# Load the model
model = YOLO('yolov8n.pt')

# Validate the model with a custom IOU threshold
iou_values = [0.5, 0.6, 0.7]  # Example IOU values
for iou in iou_values:
    metrics = model.val(iou=iou)
    print(f"IOU: {iou}, mAP: {metrics.box.map50}")

请根据您的实际需要自由调整 IOU 值列表 iou_values。希望这可以帮助您！如果有更具体的需求或进一步的问题，请随时询问。🙂

[echo385]

[glenn-jocher]

@echo385 目前，YOLOv8 的 Val 模式并不直接支持与 sahi 的集成来在高分辨率图像上绘制评价曲线。但您可以在模型验证后独立使用 sahi 或其他可视化工具来处理和展示您的评价曲线。

例如，您可以在验证模型后，使用 Python 进行结果处理并利用 matplotlib 或其他库来绘图。这里是一个基本的代码示例：

from ultralytics import YOLO
import matplotlib.pyplot as plt

# Load and validate the model
model = YOLO('path/to/your/model.pt')
metrics = model.val()  # Make sure the model has training settings integrated

# Example to plot mAP (modify as per your metric requirements)
plt.figure(figsize=(10, 5))
plt.plot(metrics.history['mAP'])
plt.title('Model mAP Over Validation Dataset')
plt.xlabel('Epochs')
plt.ylabel('mAP')
plt.grid(True)
plt.show()

这种方法使您可以灵活地展示所需的任何评价指标。希望这能帮助到您！如果有更具体的集成方式或需求，期待您的进一步说明。

[echo385]

[pderrenger]

@echo385 您好陈香枝，

感谢您的反馈和详细说明！目前YOLOv8的Val模式确实默认要求输入尺寸为640进行验证。如果您希望在不分割图片的情况下直接使用高分辨率图片进行验证，您可能需要对原有代码进行一些自定义修改以适应大尺寸的输入。

一种可能的方法是修改数据加载和预处理步骤，让模型能够接受和处理原始分辨率的图像。这可能需要较深的技术背景和一定的代码修改工作。

如果您需要进一步的具体技术指导，请随时告知，我们乐意提供更多帮助！同时，期待您分享任何成功的适应方案或进一步的问题。

祝好，
Ultralytics 团队 😊

[echo385]

请问在验证过程中哪里可以看见目标被识别为背景的图片，我想知道漏检的图片到底是哪些？

[glenn-jocher]

To view images where objects were identified as background or missed detections during the validation process, you can enable the save_json and plots arguments during validation. This will save the results in a JSON file and generate plots that visually compare predictions with ground truths, helping you identify which images contain missed detections.

Here's how you can set it up:

yolo detect val model=yolov8n.pt data=coco128.yaml save_json=True plots=True

This command will save detailed detection results and visualizations, allowing you to review the images and understand the model's performance better.

[Stephen-Hao]

类似@Ameridru的问题，在模型验证中，计算Precision、Recall时我可以设定不同的置信度阈值吗，能否打印出不同置信度阈值下的Precision和Recall？

[glenn-jocher]

@Stephen-Hao hello,

Thank you for your question! Yes, you can set different confidence thresholds to calculate Precision and Recall during model validation in YOLOv8. This can be done by adjusting the conf parameter when running the validation. Here's how you can do it:

Using Python:

from ultralytics import YOLO

# Load your model
model = YOLO("yolov8n.pt")  # or your custom model path

# Validate the model with different confidence thresholds
for conf_threshold in [0.1, 0.25, 0.5, 0.75]:
    metrics = model.val(conf=conf_threshold)
    print(f"Confidence Threshold: {conf_threshold}")
    print(f"Precision: {metrics.box.precision}")
    print(f"Recall: {metrics.box.recall}")

Using CLI:

# Validate the model with different confidence thresholds
for conf in 0.1 0.25 0.5 0.75; do
    yolo val model=yolov8n.pt conf=$conf
done

These commands will allow you to see how Precision and Recall change with different confidence thresholds. If you need more detailed guidance, please refer to our Model Validation Documentation.

If you encounter any issues, please ensure you are using the latest version of the Ultralytics package. Additionally, providing a minimum reproducible example can greatly help in diagnosing any problems. You can find more information on creating a reproducible example here.

Feel free to reach out if you have any more questions! 😊

[Stephen-Hao]

Yes! I see, Thank you for your reply!

[glenn-jocher]

@Stephen-Hao you're welcome! 😊

To address your query about setting different confidence thresholds for calculating Precision and Recall during model validation in YOLOv8, you can indeed customize the confidence threshold. This allows you to observe how Precision and Recall metrics vary with different thresholds.

Using Python:

from ultralytics import YOLO

# Load your model
model = YOLO("yolov8n.pt")  # or your custom model path

# Validate the model with different confidence thresholds
for conf_threshold in [0.1, 0.25, 0.5, 0.75]:
    metrics = model.val(conf=conf_threshold)
    print(f"Confidence Threshold: {conf_threshold}")
    print(f"Precision: {metrics.box.precision}")
    print(f"Recall: {metrics.box.recall}")

Using CLI:

# Validate the model with different confidence thresholds
for conf in 0.1 0.25 0.5 0.75; do
    yolo val model=yolov8n.pt conf=$conf
done

These commands will help you see how Precision and Recall change with different confidence thresholds. For more detailed guidance, you can refer to our Model Validation Documentation.

If you encounter any issues, please ensure you are using the latest version of the Ultralytics package. Providing a minimum reproducible example can also greatly help in diagnosing any problems. You can find more information on creating a reproducible example here.

Feel free to reach out if you have any more questions! 😊