訓練圖片物件偵測模型

本頁面說明如何使用 Google Cloud 控制台或 Vertex AI API,從圖片資料集訓練 AutoML 物件偵測模型。

訓練 AutoML 模型

Google Cloud 控制台

  1. 在 Google Cloud 控制台的 Vertex AI 專區中,前往「Datasets」頁面。

    前往「資料集」頁面

  2. 按一下要用來訓練模型的資料集名稱,開啟詳細資料頁面。

  3. 按一下「訓練新模型」

  4. 選取「AutoML」AutoML做為訓練方法。

  5. 在「選擇模型的使用位置」部分,選擇模型主機位置:「雲端」「邊緣」「Vertex AI Vision」

  6. 按一下「繼續」

  7. 輸入模型的名稱。

  8. 如要手動設定訓練資料的分割方式,請展開「進階選項」,然後選取資料分割選項。瞭解詳情

  9. 按一下「開始訓練」

    模型訓練可能需要數小時,視資料大小和複雜度而定,以及您是否指定訓練預算。您可以關閉這個分頁,稍後再返回查看。模型訓練完成後,您會收到電子郵件通知。

API

選取下方分頁,查看適用於您語言或環境的說明:

REST

使用任何要求資料之前,請先替換以下項目:

  • LOCATION:資料集所在區域,也是建立模型的區域。例如:us-central1
  • PROJECT:您的專案 ID
  • TRAININGPIPELINE_DISPLAYNAME:必填。trainingPipeline 的顯示名稱。
  • DATASET_ID:用於訓練的資料集 ID 號碼。
  • fractionSplit:選用。這是資料的其中一種機器學習用途分割選項。fractionSplit 的值總和必須為 1。例如:
    • {"trainingFraction": "0.7","validationFraction": "0.15","testFraction": "0.15"}
  • MODEL_DISPLAYNAME*:上傳 (建立) 的模型顯示名稱 TrainingPipeline。
  • MODEL_DESCRIPTION*:模型的說明。
  • modelToUpload.labels*:任何一組鍵/值配對,用於整理模型。例如:
    • "env": "prod"
    • "tier": "backend"
  • MODELTYPE:要訓練的雲端託管模型類型。選項包括:
    • CLOUD_1 - 最適合在 Google Cloud 中使用的模型,無法匯出。相較於上述 CLOUD_HIGH_ACCURACY_1 和 CLOUD_LOW_LATENCY_1 模型,預期這項模型會有更高的預測品質和更低的延遲。
    • CLOUD_HIGH_ACCURACY_1 - 最適合在 Google Cloud 中使用的模型,無法匯出。這個模型的延遲時間預計會較長,但預測品質也應高於其他雲端模型。
    • CLOUD_LOW_LATENCY_1 - 最適合在 Google Cloud 中使用的模型,無法匯出。這個模型預期延遲時間較短,但預測品質可能不如其他雲端模型。
    如需其他模型類型選項,請參閱參考說明文件
  • NODE_HOUR_BUDGET:實際訓練費用會等於或低於這個值。如果是雲端模型,預算必須介於 20,000 至 900,000 毫節點小時 (含)。 預設值為 216,000,假設使用 9 個節點,這代表一個日曆天。
  • PROJECT_NUMBER:系統自動為專案產生的專案編號

HTTP 方法和網址:

POST https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT/locations/LOCATION/trainingPipelines

JSON 要求主體:

{
  "displayName": "TRAININGPIPELINE_DISPLAYNAME",
  "inputDataConfig": {
    "datasetId": "DATASET_ID",
    "fractionSplit": {
      "trainingFraction": "DECIMAL",
      "validationFraction": "DECIMAL",
      "testFraction": "DECIMAL"
    }
  },
  "modelToUpload": {
    "displayName": "MODEL_DISPLAYNAME",
    "description": "MODEL_DESCRIPTION",
    "labels": {
      "KEY": "VALUE"
    }
  },
  "trainingTaskDefinition": "gs://google-cloud-aiplatform/schema/trainingjob/definition/automl_image_object_detection_1.0.0.yaml",
  "trainingTaskInputs": {
    "modelType": ["MODELTYPE"],
    "budgetMilliNodeHours": NODE_HOUR_BUDGET
  }
}

如要傳送要求,請選擇以下其中一個選項:

curl

將要求主體儲存在名為 request.json 的檔案中,然後執行下列指令:

curl -X POST \
-H "Authorization: Bearer $(gcloud auth print-access-token)" \
-H "Content-Type: application/json; charset=utf-8" \
-d @request.json \
"https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT/locations/LOCATION/trainingPipelines"

PowerShell

將要求主體儲存在名為 request.json 的檔案中,然後執行下列指令:

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

Invoke-WebRequest `
-Method POST `
-Headers $headers `
-ContentType: "application/json; charset=utf-8" `
-InFile request.json `
-Uri "https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT/locations/LOCATION/trainingPipelines" | Select-Object -Expand Content

回覆內容會包含規格資訊和 TRAININGPIPELINE_ID

Java

在試用這個範例之前,請先按照Java使用用戶端程式庫的 Vertex AI 快速入門中的操作說明進行設定。 詳情請參閱 Vertex AI Java API 參考說明文件

如要向 Vertex AI 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」。


import com.google.cloud.aiplatform.util.ValueConverter;
import com.google.cloud.aiplatform.v1.DeployedModelRef;
import com.google.cloud.aiplatform.v1.EnvVar;
import com.google.cloud.aiplatform.v1.FilterSplit;
import com.google.cloud.aiplatform.v1.FractionSplit;
import com.google.cloud.aiplatform.v1.InputDataConfig;
import com.google.cloud.aiplatform.v1.LocationName;
import com.google.cloud.aiplatform.v1.Model;
import com.google.cloud.aiplatform.v1.Model.ExportFormat;
import com.google.cloud.aiplatform.v1.ModelContainerSpec;
import com.google.cloud.aiplatform.v1.PipelineServiceClient;
import com.google.cloud.aiplatform.v1.PipelineServiceSettings;
import com.google.cloud.aiplatform.v1.Port;
import com.google.cloud.aiplatform.v1.PredefinedSplit;
import com.google.cloud.aiplatform.v1.PredictSchemata;
import com.google.cloud.aiplatform.v1.TimestampSplit;
import com.google.cloud.aiplatform.v1.TrainingPipeline;
import com.google.cloud.aiplatform.v1.schema.trainingjob.definition.AutoMlImageObjectDetectionInputs;
import com.google.cloud.aiplatform.v1.schema.trainingjob.definition.AutoMlImageObjectDetectionInputs.ModelType;
import com.google.rpc.Status;
import java.io.IOException;

public class CreateTrainingPipelineImageObjectDetectionSample {

  public static void main(String[] args) throws IOException {
    // TODO(developer): Replace these variables before running the sample.
    String trainingPipelineDisplayName = "YOUR_TRAINING_PIPELINE_DISPLAY_NAME";
    String project = "YOUR_PROJECT_ID";
    String datasetId = "YOUR_DATASET_ID";
    String modelDisplayName = "YOUR_MODEL_DISPLAY_NAME";
    createTrainingPipelineImageObjectDetectionSample(
        project, trainingPipelineDisplayName, datasetId, modelDisplayName);
  }

  static void createTrainingPipelineImageObjectDetectionSample(
      String project, String trainingPipelineDisplayName, String datasetId, String modelDisplayName)
      throws IOException {
    PipelineServiceSettings pipelineServiceSettings =
        PipelineServiceSettings.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 (PipelineServiceClient pipelineServiceClient =
        PipelineServiceClient.create(pipelineServiceSettings)) {
      String location = "us-central1";
      String trainingTaskDefinition =
          "gs://google-cloud-aiplatform/schema/trainingjob/definition/"
              + "automl_image_object_detection_1.0.0.yaml";
      LocationName locationName = LocationName.of(project, location);

      AutoMlImageObjectDetectionInputs autoMlImageObjectDetectionInputs =
          AutoMlImageObjectDetectionInputs.newBuilder()
              .setModelType(ModelType.CLOUD_HIGH_ACCURACY_1)
              .setBudgetMilliNodeHours(20000)
              .setDisableEarlyStopping(false)
              .build();

      InputDataConfig trainingInputDataConfig =
          InputDataConfig.newBuilder().setDatasetId(datasetId).build();
      Model model = Model.newBuilder().setDisplayName(modelDisplayName).build();
      TrainingPipeline trainingPipeline =
          TrainingPipeline.newBuilder()
              .setDisplayName(trainingPipelineDisplayName)
              .setTrainingTaskDefinition(trainingTaskDefinition)
              .setTrainingTaskInputs(ValueConverter.toValue(autoMlImageObjectDetectionInputs))
              .setInputDataConfig(trainingInputDataConfig)
              .setModelToUpload(model)
              .build();

      TrainingPipeline trainingPipelineResponse =
          pipelineServiceClient.createTrainingPipeline(locationName, trainingPipeline);

      System.out.println("Create Training Pipeline Image Object Detection Response");
      System.out.format("Name: %s\n", trainingPipelineResponse.getName());
      System.out.format("Display Name: %s\n", trainingPipelineResponse.getDisplayName());

      System.out.format(
          "Training Task Definition %s\n", trainingPipelineResponse.getTrainingTaskDefinition());
      System.out.format(
          "Training Task Inputs: %s\n", trainingPipelineResponse.getTrainingTaskInputs());
      System.out.format(
          "Training Task Metadata: %s\n", trainingPipelineResponse.getTrainingTaskMetadata());
      System.out.format("State: %s\n", trainingPipelineResponse.getState());

      System.out.format("Create Time: %s\n", trainingPipelineResponse.getCreateTime());
      System.out.format("StartTime %s\n", trainingPipelineResponse.getStartTime());
      System.out.format("End Time: %s\n", trainingPipelineResponse.getEndTime());
      System.out.format("Update Time: %s\n", trainingPipelineResponse.getUpdateTime());
      System.out.format("Labels: %s\n", trainingPipelineResponse.getLabelsMap());

      InputDataConfig inputDataConfig = trainingPipelineResponse.getInputDataConfig();
      System.out.println("Input Data Config");
      System.out.format("Dataset Id: %s", inputDataConfig.getDatasetId());
      System.out.format("Annotations Filter: %s\n", inputDataConfig.getAnnotationsFilter());

      FractionSplit fractionSplit = inputDataConfig.getFractionSplit();
      System.out.println("Fraction Split");
      System.out.format("Training Fraction: %s\n", fractionSplit.getTrainingFraction());
      System.out.format("Validation Fraction: %s\n", fractionSplit.getValidationFraction());
      System.out.format("Test Fraction: %s\n", fractionSplit.getTestFraction());

      FilterSplit filterSplit = inputDataConfig.getFilterSplit();
      System.out.println("Filter Split");
      System.out.format("Training Filter: %s\n", filterSplit.getTrainingFilter());
      System.out.format("Validation Filter: %s\n", filterSplit.getValidationFilter());
      System.out.format("Test Filter: %s\n", filterSplit.getTestFilter());

      PredefinedSplit predefinedSplit = inputDataConfig.getPredefinedSplit();
      System.out.println("Predefined Split");
      System.out.format("Key: %s\n", predefinedSplit.getKey());

      TimestampSplit timestampSplit = inputDataConfig.getTimestampSplit();
      System.out.println("Timestamp Split");
      System.out.format("Training Fraction: %s\n", timestampSplit.getTrainingFraction());
      System.out.format("Validation Fraction: %s\n", timestampSplit.getValidationFraction());
      System.out.format("Test Fraction: %s\n", timestampSplit.getTestFraction());
      System.out.format("Key: %s\n", timestampSplit.getKey());

      Model modelResponse = trainingPipelineResponse.getModelToUpload();
      System.out.println("Model To Upload");
      System.out.format("Name: %s\n", modelResponse.getName());
      System.out.format("Display Name: %s\n", modelResponse.getDisplayName());
      System.out.format("Description: %s\n", modelResponse.getDescription());

      System.out.format("Metadata Schema Uri: %s\n", modelResponse.getMetadataSchemaUri());
      System.out.format("Metadata: %s\n", modelResponse.getMetadata());
      System.out.format("Training Pipeline: %s\n", modelResponse.getTrainingPipeline());
      System.out.format("Artifact Uri: %s\n", modelResponse.getArtifactUri());

      System.out.format(
          "Supported Deployment Resources Types: %s\n",
          modelResponse.getSupportedDeploymentResourcesTypesList());
      System.out.format(
          "Supported Input Storage Formats: %s\n",
          modelResponse.getSupportedInputStorageFormatsList());
      System.out.format(
          "Supported Output Storage Formats: %s\n",
          modelResponse.getSupportedOutputStorageFormatsList());

      System.out.format("Create Time: %s\n", modelResponse.getCreateTime());
      System.out.format("Update Time: %s\n", modelResponse.getUpdateTime());
      System.out.format("Labels: %sn\n", modelResponse.getLabelsMap());

      PredictSchemata predictSchemata = modelResponse.getPredictSchemata();
      System.out.println("Predict Schemata");
      System.out.format("Instance Schema Uri: %s\n", predictSchemata.getInstanceSchemaUri());
      System.out.format("Parameters Schema Uri: %s\n", predictSchemata.getParametersSchemaUri());
      System.out.format("Prediction Schema Uri: %s\n", predictSchemata.getPredictionSchemaUri());

      for (ExportFormat exportFormat : modelResponse.getSupportedExportFormatsList()) {
        System.out.println("Supported Export Format");
        System.out.format("Id: %s\n", exportFormat.getId());
      }

      ModelContainerSpec modelContainerSpec = modelResponse.getContainerSpec();
      System.out.println("Container Spec");
      System.out.format("Image Uri: %s\n", modelContainerSpec.getImageUri());
      System.out.format("Command: %s\n", modelContainerSpec.getCommandList());
      System.out.format("Args: %s\n", modelContainerSpec.getArgsList());
      System.out.format("Predict Route: %s\n", modelContainerSpec.getPredictRoute());
      System.out.format("Health Route: %s\n", modelContainerSpec.getHealthRoute());

      for (EnvVar envVar : modelContainerSpec.getEnvList()) {
        System.out.println("Env");
        System.out.format("Name: %s\n", envVar.getName());
        System.out.format("Value: %s\n", envVar.getValue());
      }

      for (Port port : modelContainerSpec.getPortsList()) {
        System.out.println("Port");
        System.out.format("Container Port: %s\n", port.getContainerPort());
      }

      for (DeployedModelRef deployedModelRef : modelResponse.getDeployedModelsList()) {
        System.out.println("Deployed Model");
        System.out.format("Endpoint: %s\n", deployedModelRef.getEndpoint());
        System.out.format("Deployed Model Id: %s\n", deployedModelRef.getDeployedModelId());
      }

      Status status = trainingPipelineResponse.getError();
      System.out.println("Error");
      System.out.format("Code: %s\n", status.getCode());
      System.out.format("Message: %s\n", status.getMessage());
    }
  }
}

Node.js

在試用這個範例之前,請先按照Node.js使用用戶端程式庫的 Vertex AI 快速入門中的操作說明進行設定。 詳情請參閱 Vertex AI Node.js API 參考說明文件

如要向 Vertex AI 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」。

/**
 * TODO(developer): Uncomment these variables before running the sample.\
 * (Not necessary if passing values as arguments)
 */

// const datasetId = 'YOUR_DATASET_ID';
// const modelDisplayName = 'YOUR_MODEL_DISPLAY_NAME';
// const trainingPipelineDisplayName = 'YOUR_TRAINING_PIPELINE_DISPLAY_NAME';
// const project = 'YOUR_PROJECT_ID';
// const location = 'YOUR_PROJECT_LOCATION';

const aiplatform = require('@google-cloud/aiplatform');
const {definition} =
  aiplatform.protos.google.cloud.aiplatform.v1.schema.trainingjob;
const ModelType = definition.AutoMlImageObjectDetectionInputs.ModelType;

// Imports the Google Cloud Pipeline Service Client library
const {PipelineServiceClient} = aiplatform.v1;

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

// Instantiates a client
const pipelineServiceClient = new PipelineServiceClient(clientOptions);

async function createTrainingPipelineImageObjectDetection() {
  // Configure the parent resource
  const parent = `projects/${project}/locations/${location}`;

  const trainingTaskInputsObj =
    new definition.AutoMlImageObjectDetectionInputs({
      disableEarlyStopping: false,
      modelType: ModelType.CLOUD_1,
      budgetMilliNodeHours: 20000,
    });

  const trainingTaskInputs = trainingTaskInputsObj.toValue();
  const modelToUpload = {displayName: modelDisplayName};
  const inputDataConfig = {datasetId: datasetId};
  const trainingPipeline = {
    displayName: trainingPipelineDisplayName,
    trainingTaskDefinition:
      'gs://google-cloud-aiplatform/schema/trainingjob/definition/automl_image_object_detection_1.0.0.yaml',
    trainingTaskInputs,
    inputDataConfig,
    modelToUpload,
  };
  const request = {
    parent,
    trainingPipeline,
  };

  // Create training pipeline request
  const [response] =
    await pipelineServiceClient.createTrainingPipeline(request);

  console.log('Create training pipeline image object detection response');
  console.log(`Name : ${response.name}`);
  console.log('Raw response:');
  console.log(JSON.stringify(response, null, 2));
}
createTrainingPipelineImageObjectDetection();

Python

如要瞭解如何安裝或更新 Python 適用的 Vertex AI SDK,請參閱「安裝 Python 適用的 Vertex AI SDK」。 詳情請參閱 Python API 參考說明文件

from google.cloud import aiplatform
from google.cloud.aiplatform.gapic.schema import trainingjob


def create_training_pipeline_image_object_detection_sample(
    project: str,
    display_name: str,
    dataset_id: str,
    model_display_name: str,
    location: str = "us-central1",
    api_endpoint: str = "us-central1-aiplatform.googleapis.com",
):
    # 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.PipelineServiceClient(client_options=client_options)
    training_task_inputs = trainingjob.definition.AutoMlImageObjectDetectionInputs(
        model_type="CLOUD_HIGH_ACCURACY_1",
        budget_milli_node_hours=20000,
        disable_early_stopping=False,
    ).to_value()

    training_pipeline = {
        "display_name": display_name,
        "training_task_definition": "gs://google-cloud-aiplatform/schema/trainingjob/definition/automl_image_object_detection_1.0.0.yaml",
        "training_task_inputs": training_task_inputs,
        "input_data_config": {"dataset_id": dataset_id},
        "model_to_upload": {"display_name": model_display_name},
    }
    parent = f"projects/{project}/locations/{location}"
    response = client.create_training_pipeline(
        parent=parent, training_pipeline=training_pipeline
    )
    print("response:", response)

使用 REST 控制資料分割

您可以控制訓練資料在訓練、驗證和測試集之間的分割方式。使用 Vertex AI API 時,請使用 Split 物件判斷資料分割。Split 物件可以做為其中一種物件類型,納入 InputConfig 物件中,每種物件類型都提供不同的訓練資料分割方式。你只能選取一種方法。

  • FractionSplit
    • TRAINING_FRACTION:用於訓練集的訓練資料比例。
    • VALIDATION_FRACTION:用於驗證集的訓練資料比例。不適用於影片資料。
    • TEST_FRACTION:用於測試集的訓練資料比例。

    如果指定了任何分數,就必須指定所有分數。分數加總必須為 1.0。分數的預設值會因資料類型而異。瞭解詳情

    "fractionSplit": {
      "trainingFraction": TRAINING_FRACTION,
      "validationFraction": VALIDATION_FRACTION,
      "testFraction": TEST_FRACTION
    },
    
  • FilterSplit
    • TRAINING_FILTER:符合這項篩選條件的資料項目會用於訓練集。
    • VALIDATION_FILTER:符合這項篩選條件的資料項目會用於驗證集。影片資料必須為「-」。
    • TEST_FILTER:符合這項篩選條件的資料項目會用於測試集。

    這些篩選器可與 ml_use 標籤搭配使用,也可以與您套用至資料的任何標籤搭配使用。進一步瞭解如何使用 ml-use 標籤和其他標籤篩選資料。

    以下範例說明如何搭配 ml_use 標籤使用 filterSplit 物件,並納入驗證集:

    "filterSplit": {
    "trainingFilter": "labels.aiplatform.googleapis.com/ml_use=training",
    "validationFilter": "labels.aiplatform.googleapis.com/ml_use=validation",
    "testFilter": "labels.aiplatform.googleapis.com/ml_use=test"
    }