評估及疊代 AutoML 圖片分類模型

本頁面說明如何評估 AutoML Image 圖片分類模型，以便疊代模型。

Gemini Enterprise Agent Platform 提供模型評估指標，協助您判斷模型效能，例如精確度和召回率指標。Agent Platform 會使用測試集計算評估指標。

如何使用模型評估指標

模型評估指標會透過量化方式，評估模型在測試集的效能。如何解讀及使用這些指標，取決於您的業務需求，以及模型訓練要解決的問題。舉例來說，您對偽陽性的容忍度可能低於偽陰性，反之亦然。這類問題會影響您要著重的指標。

如要進一步瞭解如何疊代模型來提升效能，請參閱「疊代模型」。

Agent Platform 傳回的評估指標

Agent Platform 會傳回多項評估指標，例如精確度、喚回度和可信度門檻。Agent Platform 傳回的指標取決於模型目標。舉例來說，Agent Platform 為圖像分類模型和圖像物件偵測模型提供的評估指標就不同。

結構定義檔案可從 Cloud Storage 位置下載，用於決定 Agent Platform 為各項目標提供的評估指標。下列分頁提供結構定義檔案的連結，並說明各模型目標的評估指標。

您可以透過下列 Cloud Storage 位置查看及下載結構定義檔案：
gs://google-cloud-aiplatform/schema/modelevaluation/

• AuPRC：精確度與喚回度 (PR) 曲線下的面積，也稱為平均精確度。這個值的範圍介於 0 到 1 之間，值越大代表模型品質越高。
• 對數損失：模型推論與目標值之間的交叉熵。範圍從零到無限大，值越低代表模型品質越高。
• 可信度門檻：決定要傳回哪些推論的可信度分數。模型會傳回等於或高於這個值的推論。可信度門檻越高，精確度就越高，但召回率會下降。Agent Platform 會傳回不同門檻值的信心指標，顯示門檻對精確度和召回率的影響。
• 召回率：模型正確預測出這個類別的推論比例。也稱為真陽率。
• 精確度：模型產生的分類推論中，正確推論所占的比例。
• 混淆矩陣：混淆矩陣會顯示模型正確預測結果的頻率。對於預測錯誤的結果，矩陣會顯示模型預測的結果。混淆矩陣可協助您瞭解模型在哪些地方「混淆」了兩個結果。

取得評估指標

您可以取得模型的匯總評估指標，以及特定目標、類別或標籤的評估指標。特定類別或標籤的評估指標也稱為「評估切片」。以下內容說明如何使用 Google Cloud 控制台或 API，取得匯總評估指標和評估切片。

Google Cloud 控制台

1. 前往 Google Cloud 控制台「Agent Platform」部分的「Models」(模型) 頁面。

   前往「Models」(模型) 頁面

2. 在「Region」(區域) 下拉式選單中，選取模型所在的區域。

3. 在模型清單中點選模型，開啟模型的「Evaluate」(評估) 分頁。

   在「Evaluate」(評估) 分頁中，您可以查看模型的匯總評估指標，例如「Average precision」(平均精確度) 和「Recall」(喚回度)。

   如果模型目標有評估切片，控制台會顯示標籤清單。您可以點選標籤，查看該標籤的評估指標，如下例所示：

   

   (預先發布版) 針對每個標籤，您可以查看分組為真陽性、偽陽性和偽陰性的預測圖片。每張圖片都有 L2 平方距離指標，用於衡量圖片與訓練圖片的相似程度。標有「離群值」徽章的圖片與所有訓練圖片之間的差距相當大。如要提升模型效能，建議您新增更多與離群值相似的訓練圖片。

API

無論資料類型和目標為何，取得評估指標的 API 要求都相同，但輸出內容不同。以下範例顯示相同要求，但回應不同。

取得匯總模型評估指標

匯總模型評估指標提供模型整體的相關資訊。如要查看特定切片的資訊，請列出模型評估切片。

如要查看匯總模型評估指標，請使用 projects.locations.models.evaluations.get 方法。

Agent Platform 會傳回一組信賴度指標。每個元素都會顯示不同 confidenceThreshold 值的評估指標 (從 0 開始，最高為 1)。查看不同的門檻值，即可瞭解門檻對精確度和喚回度等其他指標的影響。

選取與您語言或環境對應的分頁：

REST

使用任何要求資料之前，請先修改下列項目的值：

• LOCATION：模型儲存區域。
• PROJECT：您的專案 ID。。
• MODEL_ID：模型資源的 ID。
• PROJECT_NUMBER：系統自動產生的專案編號。
• EVALUATION_ID：模型評估 ID (顯示在回覆中)。

HTTP 方法和網址：

GET https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT/locations/LOCATION/models/MODEL_ID/evaluations

如要傳送要求，請選擇以下其中一個選項：

curl

注意：下列指令假設您已執行 gcloud init 或 gcloud auth login 透過使用者帳戶登入 gcloud CLI，或是使用 Cloud Shell 自動登入 gcloud CLI。您可以執行 gcloud auth list，查看目前使用的帳戶。

執行下列指令：

curl -X GET \
     -H "Authorization: Bearer $(gcloud auth print-access-token)" \
     "https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT/locations/LOCATION/models/MODEL_ID/evaluations"

PowerShell

注意：下列指令假設您已執行 gcloud init 或 gcloud auth login，透過使用者帳戶登入 gcloud CLI。您可以執行 gcloud auth list，查看目前使用的帳戶。

執行下列指令：

$cred = gcloud auth print-access-token
$headers = @{ "Authorization" = "Bearer $cred" }

Invoke-WebRequest `
    -Method GET `
    -Headers $headers `
    -Uri "https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT/locations/LOCATION/models/MODEL_ID/evaluations" | Select-Object -Expand Content

您應該會收到如下的 JSON 回覆：

回應

{
  "modelEvaluations": [
    {
      "name": "projects/PROJECT_NUMBER/locations/LOCATION/models/MODEL_ID/evaluations/EVALUATION_ID",
      "metricsSchemaUri": "gs://google-cloud-aiplatform/schema/modelevaluation/classification_metrics_1.0.0.yaml",
      "metrics": {
        "auPrc": 0.9946683,
        "logLoss": 0.043421753,
        "confidenceMetrics": [
          {
            "recall": 1,
            "precision": 0.2
          },
          {
            "confidenceThreshold": 0.05,
            "recall": 0.98637605,
            "precision": 0.8960396
          },
          {
            "confidenceThreshold": 0.1,
            "recall": 0.98637605,
            "precision": 0.92346936
          },
          {
            "confidenceThreshold": 0.15,
            "recall": 0.9782016,
            "precision": 0.93979055
          },
          ...
        ],
        "confusionMatrix": {
          "annotationSpecs": [
            {
              "id": "216348703974227968",
              "displayName": "tulips"
            },
            {
              "id": "1369270208581074944",
              "displayName": "dandelion"
            },
            {
              "id": "2522191713187921920",
              "displayName": "sunflowers"
            },
            {
              "id": "4828034722401615872",
              "displayName": "daisy"
            },
            {
              "id": "7133877731615309824",
              "displayName": "roses"
            }
          ],
          "rows": [
            [
              78,
              0,
              0,
              0,
              2
            ],
            [
              0,
              89,
              0,
              1,
              0
            ],
            [
              1,
              0,
              67,
              1,
              1
            ],
            [
              0,
              1,
              0,
              61,
              1
            ],
            [
              6,
              0,
              0,
              0,
              58
            ]
          ]
        }
      },
      "createTime": "2020-10-08T23:35:54.559702Z",
      "sliceDimensions": [
        "annotationSpec"
      ]
    }
  ]
}

Java

在試用這個範例之前，請先按照「使用用戶端程式庫的 Agent Platform 快速入門導覽課程」中的 Java 設定說明操作。

如要向 Agent Platform 進行驗證，請設定應用程式預設憑證。詳情請參閱「為本機開發環境設定驗證機制」。

import com.google.cloud.aiplatform.v1.ModelEvaluation;
import com.google.cloud.aiplatform.v1.ModelEvaluationName;
import com.google.cloud.aiplatform.v1.ModelServiceClient;
import com.google.cloud.aiplatform.v1.ModelServiceSettings;
import java.io.IOException;

public class GetModelEvaluationImageClassificationSample {

  public static void main(String[] args) throws IOException {
    // TODO(developer): Replace these variables before running the sample.
    // To obtain evaluationId run the code block below after setting modelServiceSettings.
    //
    // try (ModelServiceClient modelServiceClient = ModelServiceClient.create(modelServiceSettings))
    // {
    //   String location = "us-central1";
    //   ModelName modelFullId = ModelName.of(project, location, modelId);
    //   ListModelEvaluationsRequest modelEvaluationsrequest =
    //   ListModelEvaluationsRequest.newBuilder().setParent(modelFullId.toString()).build();
    //   for (ModelEvaluation modelEvaluation :
    //     modelServiceClient.listModelEvaluations(modelEvaluationsrequest).iterateAll()) {
    //       System.out.format("Model Evaluation Name: %s%n", modelEvaluation.getName());
    //   }
    // }
    String project = "YOUR_PROJECT_ID";
    String modelId = "YOUR_MODEL_ID";
    String evaluationId = "YOUR_EVALUATION_ID";
    getModelEvaluationImageClassificationSample(project, modelId, evaluationId);
  }

  static void getModelEvaluationImageClassificationSample(
      String project, String modelId, String evaluationId) throws IOException {
    ModelServiceSettings modelServiceSettings =
        ModelServiceSettings.newBuilder()
            .setEndpoint("us-central1-aiplatform.googleapis.com:443")
            .build();

    // Initialize client that will be used to send requests. This client only needs to be created
    // once, and can be reused for multiple requests. After completing all of your requests, call
    // the "close" method on the client to safely clean up any remaining background resources.
    try (ModelServiceClient modelServiceClient = ModelServiceClient.create(modelServiceSettings)) {
      String location = "us-central1";
      ModelEvaluationName modelEvaluationName =
          ModelEvaluationName.of(project, location, modelId, evaluationId);

      ModelEvaluation modelEvaluation = modelServiceClient.getModelEvaluation(modelEvaluationName);

      System.out.println("Get Model Evaluation Image Classification Response");
      System.out.format("Model Name: %s\n", modelEvaluation.getName());
      System.out.format("Metrics Schema Uri: %s\n", modelEvaluation.getMetricsSchemaUri());
      System.out.format("Metrics: %s\n", modelEvaluation.getMetrics());
      System.out.format("Create Time: %s\n", modelEvaluation.getCreateTime());
      System.out.format("Slice Dimensions: %s\n", modelEvaluation.getSliceDimensionsList());
    }
  }
}

Node.js

在試用這個範例之前，請先按照「使用用戶端程式庫的 Agent Platform 快速入門導覽課程」中的 Node.js 設定說明操作。

如要向 Agent Platform 進行驗證，請設定應用程式預設憑證。詳情請參閱「為本機開發環境設定驗證機制」。

/**
 * TODO(developer): Uncomment these variables before running the sample
 * (not necessary if passing values as arguments). To obtain evaluationId,
 * instantiate the client and run the following the commands.
 */
// const parentName = `projects/${project}/locations/${location}/models/${modelId}`;
// const evalRequest = {
//   parent: parentName
// };
// const [evalResponse] = await modelServiceClient.listModelEvaluations(evalRequest);
// console.log(evalResponse);

// const modelId = 'YOUR_MODEL_ID';
// const evaluationId = 'YOUR_EVALUATION_ID';
// const project = 'YOUR_PROJECT_ID';
// const location = 'YOUR_PROJECT_LOCATION';

// Imports the Google Cloud Model Service Client library
const {ModelServiceClient} = require('@google-cloud/aiplatform');

// Specifies the location of the api endpoint
const clientOptions = {
  apiEndpoint: 'us-central1-aiplatform.googleapis.com',
};

// Instantiates a client
const modelServiceClient = new ModelServiceClient(clientOptions);

async function getModelEvaluationImageClassification() {
  // Configure the name resources
  const name = `projects/${project}/locations/${location}/models/${modelId}/evaluations/${evaluationId}`;
  const request = {
    name,
  };

  // Create get model evaluation request
  const [response] = await modelServiceClient.getModelEvaluation(request);

  console.log('Get model evaluation image classification response');
  console.log(`\tName : ${response.name}`);
  console.log(`\tMetrics schema uri : ${response.metricsSchemaUri}`);
  console.log(`\tMetrics : ${JSON.stringify(response.metrics)}`);
  console.log(`\tCreate time : ${JSON.stringify(response.createTime)}`);
  console.log(`\tSlice dimensions : ${response.sliceDimensions}`);

  const modelExplanation = response.modelExplanation;
  if (modelExplanation === null) {
    console.log(`\tModel explanation: ${JSON.stringify(modelExplanation)}`);
  } else {
    const meanAttributions = modelExplanation.meanAttributions;
    for (const meanAttribution of meanAttributions) {
      console.log('\t\tMean attribution');
      console.log(
        `\t\t\tBaseline output value : \
           ${meanAttribution.baselineOutputValue}`
      );
      console.log(
        `\t\t\tInstance output value : \
           ${meanAttribution.instanceOutputValue}`
      );
      console.log(
        `\t\t\tFeature attributions : \
           ${JSON.stringify(meanAttribution.featureAttributions)}`
      );
      console.log(`\t\t\tOutput index : ${meanAttribution.outputIndex}`);
      console.log(
        `\t\t\tOutput display name : \
           ${meanAttribution.outputDisplayName}`
      );
      console.log(
        `\t\t\tApproximation error : \
           ${meanAttribution.approximationError}`
      );
    }
  }
}
getModelEvaluationImageClassification();

Python

如要瞭解如何安裝或更新 Vertex AI SDK for Python，請參閱「安裝 Vertex AI SDK for Python」。詳情請參閱 Python API 參考文件。

from google.cloud import aiplatform


def get_model_evaluation_image_classification_sample(
    project: str,
    model_id: str,
    evaluation_id: str,
    location: str = "us-central1",
    api_endpoint: str = "us-central1-aiplatform.googleapis.com",
):
    """
    To obtain evaluation_id run the following commands where LOCATION
    is the region where the model is stored, PROJECT is the project ID,
    and MODEL_ID is the ID of your model.

    model_client = aiplatform.gapic.ModelServiceClient(
        client_options={
            'api_endpoint':'LOCATION-aiplatform.googleapis.com'
            }
        )
    evaluations = model_client.list_model_evaluations(parent='projects/PROJECT/locations/LOCATION/models/MODEL_ID')
    print("evaluations:", evaluations)
    """
    # The AI Platform services require regional API endpoints.
    client_options = {"api_endpoint": api_endpoint}
    # Initialize client that will be used to create and send requests.
    # This client only needs to be created once, and can be reused for multiple requests.
    client = aiplatform.gapic.ModelServiceClient(client_options=client_options)
    name = client.model_evaluation_path(
        project=project, location=location, model=model_id, evaluation=evaluation_id
    )
    response = client.get_model_evaluation(name=name)
    print("response:", response)

列出所有評估配量

projects.locations.models.evaluations.slices.list 方法會列出模型的所有評估切片。您必須取得模型的評估 ID，查看匯總評估指標時即可取得。

您可以使用模型評估切片，判斷模型在特定標籤上的成效。value 欄位會指出指標所屬的標籤。

Agent Platform 會傳回一組信賴度指標。每個元素都會顯示不同 confidenceThreshold 值的評估指標 (從 0 開始，最高為 1)。查看不同的門檻值，即可瞭解門檻對精確度和喚回度等其他指標的影響。

REST

使用任何要求資料之前，請先修改下列項目的值：

• LOCATION：模型所在區域。例如：us-central1。
• PROJECT：
• MODEL_ID：模型的 ID。
• EVALUATION_ID：模型評估 ID，其中包含要列出的評估切片。

HTTP 方法和網址：

GET https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT/locations/LOCATION/models/MODEL_ID/evaluations/EVALUATION_ID/slices

如要傳送要求，請選擇以下其中一個選項：

curl

注意：下列指令假設您已執行 gcloud init 或 gcloud auth login 透過使用者帳戶登入 gcloud CLI，或是使用 Cloud Shell 自動登入 gcloud CLI。您可以執行 gcloud auth list，查看目前使用的帳戶。

執行下列指令：

curl -X GET \
     -H "Authorization: Bearer $(gcloud auth print-access-token)" \
     "https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT/locations/LOCATION/models/MODEL_ID/evaluations/EVALUATION_ID/slices"

PowerShell

注意：下列指令假設您已執行 gcloud init 或 gcloud auth login，透過使用者帳戶登入 gcloud CLI。您可以執行 gcloud auth list，查看目前使用的帳戶。

執行下列指令：

$cred = gcloud auth print-access-token
$headers = @{ "Authorization" = "Bearer $cred" }

Invoke-WebRequest `
    -Method GET `
    -Headers $headers `
    -Uri "https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT/locations/LOCATION/models/MODEL_ID/evaluations/EVALUATION_ID/slices" | Select-Object -Expand Content

您應該會收到如下的 JSON 回覆：

回應

{
  "modelEvaluationSlices": [
    {
      "name": "projects/123456789012/locations/us-central1/models/1232584574060974976/evaluations/1234057709820346880/slices/123455414369100579",
      "slice": {
        "dimension": "annotationSpec",
        "value": "daisy"
      },
      "metricsSchemaUri": "gs://google-cloud-aiplatform/schema/modelevaluation/classification_metrics_1.0.0.yaml",
      "metrics": {
        "auPrc": 0.99713016,
        "auRoc": 0.99942565,
        "logLoss": 0.025896452,
        "confidenceMetrics": [
          {
            "recall": 1,
            "precision": 0.17166212
          },
          {
            "confidenceThreshold": 0.05,
            "recall": 1,
            "precision": 0.9402985
          },
          {
            "confidenceThreshold": 0.1,
            "recall": 1,
            "precision": 0.95454544
          },
          ...
      },
      "createTime": "2020-10-08T23:35:54.770876Z"
    },
    {
      "name": "projects/123456789012/locations/us-central1/models/1232584574060974976/evaluations/1234057709820346880/slices/123455414369101230",
      "slice": {
        "dimension": "annotationSpec",
        "value": "sunflowers"
      },
      "metricsSchemaUri": "gs://google-cloud-aiplatform/schema/modelevaluation/classification_metrics_1.0.0.yaml",
      "metrics": {
        "auPrc": 0.99369764,
        "auRoc": 0.9978355,
        "logLoss": 0.049820196,
        "confidenceMetrics": [
          {
            "recall": 1,
            "precision": 0.1907357
          },
          {
            "confidenceThreshold": 0.05,
            "recall": 0.95714283,
            "precision": 0.9852941
          },
          {
            "confidenceThreshold": 0.1,
            "recall": 0.95714283,
            "precision": 0.9852941
          },
          ...
      },
      "createTime": "2020-10-08T23:35:54.770876Z"
    }
  ]
}

Java

在試用這個範例之前，請先按照「使用用戶端程式庫的 Agent Platform 快速入門導覽課程」中的 Java 設定說明操作。

如要向 Agent Platform 進行驗證，請設定應用程式預設憑證。詳情請參閱「為本機開發環境設定驗證機制」。

import com.google.cloud.aiplatform.v1.ModelEvaluationName;
import com.google.cloud.aiplatform.v1.ModelEvaluationSlice;
import com.google.cloud.aiplatform.v1.ModelEvaluationSlice.Slice;
import com.google.cloud.aiplatform.v1.ModelServiceClient;
import com.google.cloud.aiplatform.v1.ModelServiceSettings;
import java.io.IOException;

public class ListModelEvaluationSliceSample {

  public static void main(String[] args) throws IOException {
    // TODO(developer): Replace these variables before running the sample.
    // To obtain evaluationId run the code block below after setting modelServiceSettings.
    //
    // try (ModelServiceClient modelServiceClient = ModelServiceClient.create(modelServiceSettings))
    // {
    //   String location = "us-central1";
    //   ModelName modelFullId = ModelName.of(project, location, modelId);
    //   ListModelEvaluationsRequest modelEvaluationsrequest =
    //   ListModelEvaluationsRequest.newBuilder().setParent(modelFullId.toString()).build();
    //   for (ModelEvaluation modelEvaluation :
    //     modelServiceClient.listModelEvaluations(modelEvaluationsrequest).iterateAll()) {
    //       System.out.format("Model Evaluation Name: %s%n", modelEvaluation.getName());
    //   }
    // }
    String project = "YOUR_PROJECT_ID";
    String modelId = "YOUR_MODEL_ID";
    String evaluationId = "YOUR_EVALUATION_ID";
    listModelEvaluationSliceSample(project, modelId, evaluationId);
  }

  static void listModelEvaluationSliceSample(String project, String modelId, String evaluationId)
      throws IOException {
    ModelServiceSettings modelServiceSettings =
        ModelServiceSettings.newBuilder()
            .setEndpoint("us-central1-aiplatform.googleapis.com:443")
            .build();

    // Initialize client that will be used to send requests. This client only needs to be created
    // once, and can be reused for multiple requests. After completing all of your requests, call
    // the "close" method on the client to safely clean up any remaining background resources.
    try (ModelServiceClient modelServiceClient = ModelServiceClient.create(modelServiceSettings)) {
      String location = "us-central1";
      ModelEvaluationName modelEvaluationName =
          ModelEvaluationName.of(project, location, modelId, evaluationId);

      for (ModelEvaluationSlice modelEvaluationSlice :
          modelServiceClient.listModelEvaluationSlices(modelEvaluationName).iterateAll()) {
        System.out.format("Model Evaluation Slice Name: %s\n", modelEvaluationSlice.getName());
        System.out.format("Metrics Schema Uri: %s\n", modelEvaluationSlice.getMetricsSchemaUri());
        System.out.format("Metrics: %s\n", modelEvaluationSlice.getMetrics());
        System.out.format("Create Time: %s\n", modelEvaluationSlice.getCreateTime());

        Slice slice = modelEvaluationSlice.getSlice();
        System.out.format("Slice Dimensions: %s\n", slice.getDimension());
        System.out.format("Slice Value: %s\n\n", slice.getValue());
      }
    }
  }
}

Node.js

在試用這個範例之前，請先按照「使用用戶端程式庫的 Agent Platform 快速入門導覽課程」中的 Node.js 設定說明操作。

如要向 Agent Platform 進行驗證，請設定應用程式預設憑證。詳情請參閱「為本機開發環境設定驗證機制」。

/**
 * TODO(developer): Uncomment these variables before running the sample
 * (not necessary if passing values as arguments). To obtain evaluationId,
 * instantiate the client and run the following the commands.
 */
// const parentName = `projects/${project}/locations/${location}/models/${modelId}`;
// const evalRequest = {
//   parent: parentName
// };
// const [evalResponse] = await modelServiceClient.listModelEvaluations(evalRequest);
// console.log(evalResponse);

// const modelId = 'YOUR_MODEL_ID';
// const evaluationId = 'YOUR_EVALUATION_ID';
// const project = 'YOUR_PROJECT_ID';
// const location = 'YOUR_PROJECT_LOCATION';

// Imports the Google Cloud Model Service Client library
const {ModelServiceClient} = require('@google-cloud/aiplatform');

// Specifies the location of the api endpoint
const clientOptions = {
  apiEndpoint: 'us-central1-aiplatform.googleapis.com',
};

// Instantiates a client
const modelServiceClient = new ModelServiceClient(clientOptions);

async function listModelEvaluationSlices() {
  // Configure the parent resources
  const parent = `projects/${project}/locations/${location}/models/${modelId}/evaluations/${evaluationId}`;
  const request = {
    parent,
  };

  // Get and print out a list of all the evaluation slices for this resource
  const [response] =
    await modelServiceClient.listModelEvaluationSlices(request);
  console.log('List model evaluation response', response);
  console.log(response);
}
listModelEvaluationSlices();

Python

如要瞭解如何安裝或更新 Vertex AI SDK for Python，請參閱「安裝 Vertex AI SDK for Python」。詳情請參閱 Python API 參考文件。

from google.cloud import aiplatform


def list_model_evaluation_slices_sample(
    project: str,
    model_id: str,
    evaluation_id: str,
    location: str = "us-central1",
    api_endpoint: str = "us-central1-aiplatform.googleapis.com",
):
    """
    To obtain evaluation_id run the following commands where LOCATION
    is the region where the model is stored, PROJECT is the project ID,
    and MODEL_ID is the ID of your model.

    model_client = aiplatform.gapic.ModelServiceClient(
        client_options={
            'api_endpoint':'LOCATION-aiplatform.googleapis.com'
            }
        )
    evaluations = model_client.list_model_evaluations(parent='projects/PROJECT/locations/LOCATION/models/MODEL_ID')
    print("evaluations:", evaluations)
    """
    # The AI Platform services require regional API endpoints.
    client_options = {"api_endpoint": api_endpoint}
    # Initialize client that will be used to create and send requests.
    # This client only needs to be created once, and can be reused for multiple requests.
    client = aiplatform.gapic.ModelServiceClient(client_options=client_options)
    parent = client.model_evaluation_path(
        project=project, location=location, model=model_id, evaluation=evaluation_id
    )
    response = client.list_model_evaluation_slices(parent=parent)
    for model_evaluation_slice in response:
        print("model_evaluation_slice:", model_evaluation_slice)

取得單一配量的指標

如要查看單一配量的評估指標，請使用 projects.locations.models.evaluations.slices.get 方法。您必須擁有配量 ID，系統會在您列出所有配量時提供。以下範例適用於所有資料類型和目標。

REST

使用任何要求資料之前，請先修改下列項目的值：

• LOCATION：模型所在區域。例如 us-central1。
• PROJECT：
• MODEL_ID：模型的 ID。
• EVALUATION_ID：模型評估 ID，其中包含要擷取的評估切片。
• SLICE_ID：要取得的評估切片 ID。
• PROJECT_NUMBER：系統自動產生的專案編號。
• EVALUATION_METRIC_SCHEMA_FILE_NAME：結構定義檔案的名稱，用於定義要傳回的評估指標，例如 classification_metrics_1.0.0。

HTTP 方法和網址：

GET https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT/locations/LOCATION/models/MODEL_ID/evaluations/EVALUATION_ID/slices/SLICE_ID

如要傳送要求，請選擇以下其中一個選項：

curl

注意：下列指令假設您已執行 gcloud init 或 gcloud auth login 透過使用者帳戶登入 gcloud CLI，或是使用 Cloud Shell 自動登入 gcloud CLI。您可以執行 gcloud auth list，查看目前使用的帳戶。

執行下列指令：

curl -X GET \
     -H "Authorization: Bearer $(gcloud auth print-access-token)" \
     "https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT/locations/LOCATION/models/MODEL_ID/evaluations/EVALUATION_ID/slices/SLICE_ID"

PowerShell

注意：下列指令假設您已執行 gcloud init 或 gcloud auth login，透過使用者帳戶登入 gcloud CLI。您可以執行 gcloud auth list，查看目前使用的帳戶。

執行下列指令：

$cred = gcloud auth print-access-token
$headers = @{ "Authorization" = "Bearer $cred" }

Invoke-WebRequest `
    -Method GET `
    -Headers $headers `
    -Uri "https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT/locations/LOCATION/models/MODEL_ID/evaluations/EVALUATION_ID/slices/SLICE_ID" | Select-Object -Expand Content

您應該會收到如下的 JSON 回覆：

回應

{
  "name": "projects/PROJECT_NUMBER/locations/LOCATION/models/MODEL_ID/evaluations/EVALUATION_ID/slices/SLICE_ID",
  "slice": {
    "dimension": "annotationSpec",
    "value": "a particular class or label"
  },
  "metricsSchemaUri": "gs://google-cloud-aiplatform/schema/modelevaluation/EVALUATION_METRIC_SCHEMA_FILE_NAME.yaml",
  "metrics": {
    evaluation metrics for the slice
  },
  "createTime": "2020-10-08T23:35:54.770876Z"
}

Java

在試用這個範例之前，請先按照「使用用戶端程式庫的 Agent Platform 快速入門導覽課程」中的 Java 設定說明操作。

如要向 Agent Platform 進行驗證，請設定應用程式預設憑證。詳情請參閱「為本機開發環境設定驗證機制」。

import com.google.cloud.aiplatform.v1.ModelEvaluationSlice;
import com.google.cloud.aiplatform.v1.ModelEvaluationSlice.Slice;
import com.google.cloud.aiplatform.v1.ModelEvaluationSliceName;
import com.google.cloud.aiplatform.v1.ModelServiceClient;
import com.google.cloud.aiplatform.v1.ModelServiceSettings;
import java.io.IOException;

public class GetModelEvaluationSliceSample {

  public static void main(String[] args) throws IOException {
    // TODO(developer): Replace these variables before running the sample.
    // To obtain evaluationId run the code block below after setting modelServiceSettings.
    //
    // try (ModelServiceClient modelServiceClient = ModelServiceClient.create(modelServiceSettings))
    // {
    //   String location = "us-central1";
    //   ModelName modelFullId = ModelName.of(project, location, modelId);
    //   ListModelEvaluationsRequest modelEvaluationsrequest =
    //   ListModelEvaluationsRequest.newBuilder().setParent(modelFullId.toString()).build();
    //   for (ModelEvaluation modelEvaluation :
    //     modelServiceClient.listModelEvaluations(modelEvaluationsrequest).iterateAll()) {
    //       System.out.format("Model Evaluation Name: %s%n", modelEvaluation.getName());
    //   }
    // }
    String project = "YOUR_PROJECT_ID";
    String modelId = "YOUR_MODEL_ID";
    String evaluationId = "YOUR_EVALUATION_ID";
    String sliceId = "YOUR_SLICE_ID";
    getModelEvaluationSliceSample(project, modelId, evaluationId, sliceId);
  }

  static void getModelEvaluationSliceSample(
      String project, String modelId, String evaluationId, String sliceId) throws IOException {
    ModelServiceSettings modelServiceSettings =
        ModelServiceSettings.newBuilder()
            .setEndpoint("us-central1-aiplatform.googleapis.com:443")
            .build();

    // Initialize client that will be used to send requests. This client only needs to be created
    // once, and can be reused for multiple requests. After completing all of your requests, call
    // the "close" method on the client to safely clean up any remaining background resources.
    try (ModelServiceClient modelServiceClient = ModelServiceClient.create(modelServiceSettings)) {
      String location = "us-central1";
      ModelEvaluationSliceName modelEvaluationSliceName =
          ModelEvaluationSliceName.of(project, location, modelId, evaluationId, sliceId);

      ModelEvaluationSlice modelEvaluationSlice =
          modelServiceClient.getModelEvaluationSlice(modelEvaluationSliceName);

      System.out.println("Get Model Evaluation Slice Response");
      System.out.format("Model Evaluation Slice Name: %s\n", modelEvaluationSlice.getName());
      System.out.format("Metrics Schema Uri: %s\n", modelEvaluationSlice.getMetricsSchemaUri());
      System.out.format("Metrics: %s\n", modelEvaluationSlice.getMetrics());
      System.out.format("Create Time: %s\n", modelEvaluationSlice.getCreateTime());

      Slice slice = modelEvaluationSlice.getSlice();
      System.out.format("Slice Dimensions: %s\n", slice.getDimension());
      System.out.format("Slice Value: %s\n", slice.getValue());
    }
  }
}

Node.js

在試用這個範例之前，請先按照「使用用戶端程式庫的 Agent Platform 快速入門導覽課程」中的 Node.js 設定說明操作。

如要向 Agent Platform 進行驗證，請設定應用程式預設憑證。詳情請參閱「為本機開發環境設定驗證機制」。

/**
 * TODO(developer): Uncomment these variables before running the sample
 * (not necessary if passing values as arguments). To obtain evaluationId,
 * instantiate the client and run the following the commands.
 */
// const parentName = `projects/${project}/locations/${location}/models/${modelId}`;
// const evalRequest = {
//   parent: parentName
// };
// const [evalResponse] = await modelServiceClient.listModelEvaluations(evalRequest);
// console.log(evalResponse);

// const modelId = 'YOUR_MODEL_ID';
// const evaluationId = 'YOUR_EVALUATION_ID';
// const sliceId = 'YOUR_SLICE_ID';
// const project = 'YOUR_PROJECT_ID';
// const location = 'YOUR_PROJECT_LOCATION';

// Imports the Google Cloud Model Service client library
const {ModelServiceClient} = require('@google-cloud/aiplatform');
// Specifies the location of the api endpoint
const clientOptions = {
  apiEndpoint: 'us-central1-aiplatform.googleapis.com',
};
// Specifies the location of the api endpoint
const modelServiceClient = new ModelServiceClient(clientOptions);

async function getModelEvaluationSlice() {
  // Configure the parent resource
  const name = `projects/${project}/locations/${location}/models/${modelId}/evaluations/${evaluationId}/slices/${sliceId}`;
  const request = {
    name,
  };

  // Get and print out a list of all the endpoints for this resource
  const [response] =
    await modelServiceClient.getModelEvaluationSlice(request);

  console.log('Get model evaluation slice');
  console.log(`\tName : ${response.name}`);
  console.log(`\tMetrics_Schema_Uri : ${response.metricsSchemaUri}`);
  console.log(`\tMetrics : ${JSON.stringify(response.metrics)}`);
  console.log(`\tCreate time : ${JSON.stringify(response.createTime)}`);

  console.log('Slice');
  const slice = response.slice;
  console.log(`\tDimension :${slice.dimension}`);
  console.log(`\tValue :${slice.value}`);
}
getModelEvaluationSlice();

Python

如要瞭解如何安裝或更新 Vertex AI SDK for Python，請參閱「安裝 Vertex AI SDK for Python」。詳情請參閱 Python API 參考文件。

from google.cloud import aiplatform


def get_model_evaluation_slice_sample(
    project: str,
    model_id: str,
    evaluation_id: str,
    slice_id: str,
    location: str = "us-central1",
    api_endpoint: str = "us-central1-aiplatform.googleapis.com",
):
    """
    To obtain evaluation_id run the following commands where LOCATION
    is the region where the model is stored, PROJECT is the project ID,
    and MODEL_ID is the ID of your model.

    model_client = aiplatform.gapic.ModelServiceClient(
        client_options={
            'api_endpoint':'LOCATION-aiplatform.googleapis.com'
            }
        )
    evaluations = model_client.list_model_evaluations(parent='projects/PROJECT/locations/LOCATION/models/MODEL_ID')
    print("evaluations:", evaluations)
    """
    # The AI Platform services require regional API endpoints.
    client_options = {"api_endpoint": api_endpoint}
    # Initialize client that will be used to create and send requests.
    # This client only needs to be created once, and can be reused for multiple requests.
    client = aiplatform.gapic.ModelServiceClient(client_options=client_options)
    name = client.model_evaluation_slice_path(
        project=project,
        location=location,
        model=model_id,
        evaluation=evaluation_id,
        slice=slice_id,
    )
    response = client.get_model_evaluation_slice(name=name)
    print("response:", response)

反覆訓練模型

如果模型不符合預期，您可以從模型評估指標著手，開始偵錯。舉例來說，如果精確度和喚回度分數偏低，可能代表模型需要更多訓練資料，或標籤不一致。精確度和喚回度完美，可能表示測試資料太容易預測，可能無法妥善一般化。

您可以疊代訓練資料並建立新模型。建立新模型後，您可以比較現有模型和新模型的評估指標。

以下建議有助於改善標示項目的模型，例如分類或偵測模型：

• 建議在訓練資料中加入更多或更多元的範例。舉例來說，如果是圖像分類模型，您可以加入廣角圖像、解析度較高或較低的圖像，或不同視角的圖像。如需更多指引，請參閱「準備資料」。
• 建議您移除樣本數不多的類別或標籤。如果樣本不足，模型就無法持續且準確地預測這些類別或標籤。
• 機器無法解讀類別或標籤的名稱，也無法理解兩者之間的細微差異，例如「door」和「door_with_knob」。您必須提供資料，協助機器辨識這類細微差異。
• 使用更多真陽性和真陰性範例擴充資料，尤其是接近決策邊界的範例，以減少模型混淆。
• 指定自己的資料分割 (訓練、驗證和測試)。Agent Platform 會隨機將項目指派給各個組合。因此，訓練集和驗證集可能會分配到近似重複項目，導致過度配適，進而降低測試集的效能。如要進一步瞭解如何自行設定資料分割，請參閱「關於 AutoML 模型的資料分割」。
• 如果模型的評估指標包含混淆矩陣，您可以查看模型是否混淆兩個標籤，其中模型預測特定標籤的次數明顯多於實際標籤。請檢查資料，確認樣本標籤正確無誤。
• 如果訓練時間較短 (節點時數上限較低)，可以延長訓練時間 (提高節點時數上限)，取得品質較高的模型。