如何获取GCP Vertex AI AutoML目标检测不同阈值下的逐类别评估指标

1. AutoML API 位置不匹配报错 #

你使用eu-automl.googleapis.com端点时，传入的位置参数不能是具体的区域europe-west4，必须替换为统一的区域标识eu，修改model_full_id构造代码即可：

# 把原来的europe-west4替换为eu
model_full_id = client.model_path(project_id, "eu", model_id)

注：更推荐使用Vertex AI统一SDK完成指标查询，无需混用旧版AutoML API，避免多套接口的适配问题

2. Vertex AI ModelService 无效evaluation_id报错 #

evaluation_id不能随意填写，需要先调用列表接口拉取当前模型下的所有评估记录ID，再传入查询接口：

from google.cloud import aiplatform

def list_model_evaluations(project: str, model_id: str, location: str = "europe-west4"):
    api_endpoint = f"{location}-aiplatform.googleapis.com"
    client_options = {"api_endpoint": api_endpoint}
    client = aiplatform.gapic.ModelServiceClient(client_options=client_options)
    parent = client.model_path(project, location, model_id)
    # 拉取所有评估记录
    evaluations = client.list_model_evaluations(parent=parent)
    eval_ids = []
    for eval in evaluations:
        # 从评估资源名中提取evaluation_id
        eval_id = eval.name.split("/")[-1]
        eval_ids.append(eval_id)
        print(f"找到评估ID: {eval_id}, 评估时间: {eval.create_time}")
    return eval_ids

用上述接口拿到可用的evaluation_id后，再传入你原来的get_model_evaluation_image_object_detection_sample方法即可避免400报错。

成功拉取到ModelEvaluation响应后，目标检测的阈值对应指标存放在metrics字段的boundingBoxMetrics数组中：

• 数组每个元素对应一组IoU阈值+置信度阈值的组合
• 每个组合下的annotationMetrics数组为对应每个类别的指标，包含precision、recall、f1Score等字段
• 每个类别的annotationSpecId对应你在数据集中定义的标签ID，可提前通过模型的标签列表映射为可读的标签名

完整的指标查询示例代码：

def get_per_class_threshold_metrics(project: str, model_id: str, evaluation_id: str, location: str = "europe-west4"):
    api_endpoint = f"{location}-aiplatform.googleapis.com"
    client_options = {"api_endpoint": api_endpoint}
    client = aiplatform.gapic.ModelServiceClient(client_options=client_options)
    eval_name = client.model_evaluation_path(project, location, model_id, evaluation_id)
    response = client.get_model_evaluation(name=eval_name)
    
    bbox_metrics = response.metrics.get("boundingBoxMetrics", [])
    for metric_entry in bbox_metrics:
        iou_thresh = metric_entry["iouThreshold"]
        conf_thresh = metric_entry["confidenceThreshold"]
        print(f"\n=== IoU阈值: {iou_thresh}, 置信度阈值: {conf_thresh} ===")
        # 遍历每个类别的指标
        for class_metric in metric_entry["annotationMetrics"]:
            label_id = class_metric["annotationSpecId"]
            precision = class_metric.get("precision", 0)
            recall = class_metric.get("recall", 0)
            f1 = class_metric.get("f1Score", 0)
            print(f"类别ID {label_id}: 精确率={precision:.4f}, 召回率={recall:.4f}, F1={f1:.4f}")

内容的提问来源于stack exchange，提问作者TwinPenguins