共用方式為


Microsoft.MachineLearningServices 工作區/作業 2024-07-01-preview

Bicep 資源定義

工作區/作業資源類型可以使用目標作業來部署:

如需每個 API 版本中已變更屬性的清單,請參閱 變更記錄檔。

資源格式

若要建立 Microsoft.MachineLearningServices/workspaces/jobs 資源,請將下列 Bicep 新增至範本。

resource symbolicname 'Microsoft.MachineLearningServices/workspaces/jobs@2024-07-01-preview' = {
  parent: resourceSymbolicName
  name: 'string'
  properties: {
    componentId: 'string'
    computeId: 'string'
    description: 'string'
    displayName: 'string'
    experimentName: 'string'
    identity: {
      identityType: 'string'
      // For remaining properties, see IdentityConfiguration objects
    }
    isArchived: bool
    notificationSetting: {
      emailOn: [
        'string'
      ]
      emails: [
        'string'
      ]
      webhooks: {
        {customized property}: {
          eventType: 'string'
          webhookType: 'string'
          // For remaining properties, see Webhook objects
        }
      }
    }
    properties: {
      {customized property}: 'string'
    }
    services: {
      {customized property}: {
        endpoint: 'string'
        jobServiceType: 'string'
        nodes: {
          nodesValueType: 'string'
          // For remaining properties, see Nodes objects
        }
        port: int
        properties: {
          {customized property}: 'string'
        }
      }
    }
    tags: {
      {customized property}: 'string'
    }
    jobType: 'string'
    // For remaining properties, see JobBaseProperties objects
  }
}

FineTuningVertical 物件

設定 modelProvider 屬性以指定物件的類型。

針對 AzureOpenAI,請使用:

{
  hyperParameters: {
    batchSize: int
    learningRateMultiplier: int
    nEpochs: int
  }
  modelProvider: 'AzureOpenAI'
}

針對 [自定義],請使用:

{
  hyperParameters: {
    {customized property}: 'string'
  }
  modelProvider: 'Custom'
}

TargetRollingWindowSize 物件

設定 mode 屬性以指定物件的類型。

針對 [自動],請使用:

{
  mode: 'Auto'
}

針對 [自定義],請使用:

{
  mode: 'Custom'
  value: int
}

JobInput 物件

設定 jobInputType 屬性以指定物件的類型。

針對 custom_model,請使用:

{
  jobInputType: 'custom_model'
  mode: 'string'
  uri: 'string'
}

針對 常值,請使用:

{
  jobInputType: 'literal'
  value: 'string'
}

針對 mlflow_model,請使用:

{
  jobInputType: 'mlflow_model'
  mode: 'string'
  uri: 'string'
}

針對 mltable,請使用:

{
  jobInputType: 'mltable'
  mode: 'string'
  uri: 'string'
}

針對 triton_model,請使用:

{
  jobInputType: 'triton_model'
  mode: 'string'
  uri: 'string'
}

針對 uri_file,請使用:

{
  jobInputType: 'uri_file'
  mode: 'string'
  uri: 'string'
}

針對 uri_folder,請使用:

{
  jobInputType: 'uri_folder'
  mode: 'string'
  uri: 'string'
}

JobBaseProperties 物件

設定 jobType 屬性以指定物件的類型。

針對 AutoML,請使用:

{
  environmentId: 'string'
  environmentVariables: {
    {customized property}: 'string'
  }
  jobType: 'AutoML'
  outputs: {
    {customized property}: {
      description: 'string'
      jobOutputType: 'string'
      // For remaining properties, see JobOutput objects
    }
  }
  queueSettings: {
    jobTier: 'string'
  }
  resources: {
    dockerArgs: 'string'
    instanceCount: int
    instanceType: 'string'
    properties: {
      {customized property}: any(...)
    }
    shmSize: 'string'
  }
  taskDetails: {
    logVerbosity: 'string'
    targetColumnName: 'string'
    trainingData: {
      description: 'string'
      jobInputType: 'string'
      mode: 'string'
      uri: 'string'
    }
    taskType: 'string'
    // For remaining properties, see AutoMLVertical objects
  }
}

針對 [命令],請使用:

{
  codeId: 'string'
  command: 'string'
  distribution: {
    distributionType: 'string'
    // For remaining properties, see DistributionConfiguration objects
  }
  environmentId: 'string'
  environmentVariables: {
    {customized property}: 'string'
  }
  inputs: {
    {customized property}: {
      description: 'string'
      jobInputType: 'string'
      // For remaining properties, see JobInput objects
    }
  }
  jobType: 'Command'
  limits: {
    jobLimitsType: 'string'
    timeout: 'string'
  }
  outputs: {
    {customized property}: {
      description: 'string'
      jobOutputType: 'string'
      // For remaining properties, see JobOutput objects
    }
  }
  queueSettings: {
    jobTier: 'string'
  }
  resources: {
    dockerArgs: 'string'
    instanceCount: int
    instanceType: 'string'
    properties: {
      {customized property}: any(...)
    }
    shmSize: 'string'
  }
}

針對 FineTuning,請使用:

{
  fineTuningDetails: {
    model: {
      description: 'string'
      jobInputType: 'string'
      mode: 'string'
      uri: 'string'
    }
    taskType: 'string'
    trainingData: {
      description: 'string'
      jobInputType: 'string'
      // For remaining properties, see JobInput objects
    }
    validationData: {
      description: 'string'
      jobInputType: 'string'
      // For remaining properties, see JobInput objects
    }
    modelProvider: 'string'
    // For remaining properties, see FineTuningVertical objects
  }
  jobType: 'FineTuning'
  outputs: {
    {customized property}: {
      description: 'string'
      jobOutputType: 'string'
      // For remaining properties, see JobOutput objects
    }
  }
  queueSettings: {
    jobTier: 'string'
  }
  resources: {
    instanceTypes: [
      'string'
    ]
  }
}

針對 管線,請使用:

{
  inputs: {
    {customized property}: {
      description: 'string'
      jobInputType: 'string'
      // For remaining properties, see JobInput objects
    }
  }
  jobs: {
    {customized property}: any(...)
  }
  jobType: 'Pipeline'
  outputs: {
    {customized property}: {
      description: 'string'
      jobOutputType: 'string'
      // For remaining properties, see JobOutput objects
    }
  }
  settings: any(...)
  sourceJobId: 'string'
}

針對 Spark,請使用:

{
  archives: [
    'string'
  ]
  args: 'string'
  codeId: 'string'
  conf: {
    {customized property}: 'string'
  }
  entry: {
    sparkJobEntryType: 'string'
    // For remaining properties, see SparkJobEntry objects
  }
  environmentId: 'string'
  environmentVariables: {
    {customized property}: 'string'
  }
  files: [
    'string'
  ]
  inputs: {
    {customized property}: {
      description: 'string'
      jobInputType: 'string'
      // For remaining properties, see JobInput objects
    }
  }
  jars: [
    'string'
  ]
  jobType: 'Spark'
  outputs: {
    {customized property}: {
      description: 'string'
      jobOutputType: 'string'
      // For remaining properties, see JobOutput objects
    }
  }
  pyFiles: [
    'string'
  ]
  queueSettings: {
    jobTier: 'string'
  }
  resources: {
    instanceType: 'string'
    runtimeVersion: 'string'
  }
}

針對 掃掠,請使用:

{
  earlyTermination: {
    delayEvaluation: int
    evaluationInterval: int
    policyType: 'string'
    // For remaining properties, see EarlyTerminationPolicy objects
  }
  inputs: {
    {customized property}: {
      description: 'string'
      jobInputType: 'string'
      // For remaining properties, see JobInput objects
    }
  }
  jobType: 'Sweep'
  limits: {
    jobLimitsType: 'string'
    maxConcurrentTrials: int
    maxTotalTrials: int
    timeout: 'string'
    trialTimeout: 'string'
  }
  objective: {
    goal: 'string'
    primaryMetric: 'string'
  }
  outputs: {
    {customized property}: {
      description: 'string'
      jobOutputType: 'string'
      // For remaining properties, see JobOutput objects
    }
  }
  queueSettings: {
    jobTier: 'string'
  }
  samplingAlgorithm: {
    samplingAlgorithmType: 'string'
    // For remaining properties, see SamplingAlgorithm objects
  }
  searchSpace: any(...)
  trial: {
    codeId: 'string'
    command: 'string'
    distribution: {
      distributionType: 'string'
      // For remaining properties, see DistributionConfiguration objects
    }
    environmentId: 'string'
    environmentVariables: {
      {customized property}: 'string'
    }
    resources: {
      dockerArgs: 'string'
      instanceCount: int
      instanceType: 'string'
      properties: {
        {customized property}: any(...)
      }
      shmSize: 'string'
    }
  }
}

IdentityConfiguration 物件

設定 identityType 屬性以指定物件的類型。

針對 AMLToken,請使用:

{
  identityType: 'AMLToken'
}

針對 [受控],請使用:

{
  clientId: 'string'
  identityType: 'Managed'
  objectId: 'string'
  resourceId: 'string'
}

針對 UserIdentity,請使用:

{
  identityType: 'UserIdentity'
}

ForecastHorizon 物件

設定 mode 屬性以指定物件的類型。

針對 [自動],請使用:

{
  mode: 'Auto'
}

針對 [自定義],請使用:

{
  mode: 'Custom'
  value: int
}

EarlyTerminationPolicy 物件

設定 policyType 屬性以指定物件的類型。

針對 Bandit,請使用:

{
  policyType: 'Bandit'
  slackAmount: int
  slackFactor: int
}

針對 MedianStopping,請使用:

{
  policyType: 'MedianStopping'
}

針對 [截斷][選取],請使用:

{
  policyType: 'TruncationSelection'
  truncationPercentage: int
}

季節性物件

設定 mode 屬性以指定物件的類型。

針對 [自動],請使用:

{
  mode: 'Auto'
}

針對 [自定義],請使用:

{
  mode: 'Custom'
  value: int
}

JobOutput 物件

設定 jobOutputType 屬性以指定對象的類型。

針對 custom_model,請使用:

{
  jobOutputType: 'custom_model'
  mode: 'string'
  uri: 'string'
}

針對 mlflow_model,請使用:

{
  jobOutputType: 'mlflow_model'
  mode: 'string'
  uri: 'string'
}

針對 mltable,請使用:

{
  jobOutputType: 'mltable'
  mode: 'string'
  uri: 'string'
}

針對 triton_model,請使用:

{
  jobOutputType: 'triton_model'
  mode: 'string'
  uri: 'string'
}

針對 uri_file,請使用:

{
  jobOutputType: 'uri_file'
  mode: 'string'
  uri: 'string'
}

針對 uri_folder,請使用:

{
  jobOutputType: 'uri_folder'
  mode: 'string'
  uri: 'string'
}

TargetLags 物件

設定 mode 屬性以指定物件的類型。

針對 [自動],請使用:

{
  mode: 'Auto'
}

針對 [自定義],請使用:

{
  mode: 'Custom'
  values: [
    int
  ]
}

Webhook 物件

設定 webhookType 屬性以指定物件的類型。

針對 AzureDevOps,請使用:

{
  webhookType: 'AzureDevOps'
}

SamplingAlgorithm 物件

設定 samplingAlgorithmType 屬性以指定物件的類型。

針對 貝氏,請使用:

{
  samplingAlgorithmType: 'Bayesian'
}

針對 Grid,請使用:

{
  samplingAlgorithmType: 'Grid'
}

針對 [隨機],請使用:

{
  rule: 'string'
  samplingAlgorithmType: 'Random'
  seed: int
}

SparkJobEntry 物件

設定 sparkJobEntryType 屬性以指定物件的類型。

針對 SparkJobPythonEntry,請使用:

{
  file: 'string'
  sparkJobEntryType: 'SparkJobPythonEntry'
}

針對 SparkJobScalaEntry,請使用:

{
  className: 'string'
  sparkJobEntryType: 'SparkJobScalaEntry'
}

DistributionConfiguration 物件

設定 distributionType 屬性以指定物件的類型。

針對 Mpi,請使用:

{
  distributionType: 'Mpi'
  processCountPerInstance: int
}

針對 PyTorch,請使用:

{
  distributionType: 'PyTorch'
  processCountPerInstance: int
}

針對 TensorFlow,請使用:

{
  distributionType: 'TensorFlow'
  parameterServerCount: int
  workerCount: int
}

NCrossValidations 物件

設定 mode 屬性以指定物件的類型。

針對 [自動],請使用:

{
  mode: 'Auto'
}

針對 [自定義],請使用:

{
  mode: 'Custom'
  value: int
}

Nodes 物件

設定 nodesValueType 屬性以指定對象的類型。

針對 [全部],請使用:

{
  nodesValueType: 'All'
}

AutoMLVertical 物件

設定 taskType 屬性以指定物件的類型。

針對 [分類],請使用:

{
  cvSplitColumnNames: [
    'string'
  ]
  featurizationSettings: {
    blockedTransformers: [
      'string'
    ]
    columnNameAndTypes: {
      {customized property}: 'string'
    }
    datasetLanguage: 'string'
    enableDnnFeaturization: bool
    mode: 'string'
    transformerParams: {
      {customized property}: [
        {
          fields: [
            'string'
          ]
          parameters: any(...)
        }
      ]
    }
  }
  limitSettings: {
    enableEarlyTermination: bool
    exitScore: int
    maxConcurrentTrials: int
    maxCoresPerTrial: int
    maxTrials: int
    timeout: 'string'
    trialTimeout: 'string'
  }
  nCrossValidations: {
    mode: 'string'
    // For remaining properties, see NCrossValidations objects
  }
  positiveLabel: 'string'
  primaryMetric: 'string'
  taskType: 'Classification'
  testData: {
    description: 'string'
    jobInputType: 'string'
    mode: 'string'
    uri: 'string'
  }
  testDataSize: int
  trainingSettings: {
    allowedTrainingAlgorithms: [
      'string'
    ]
    blockedTrainingAlgorithms: [
      'string'
    ]
    enableDnnTraining: bool
    enableModelExplainability: bool
    enableOnnxCompatibleModels: bool
    enableStackEnsemble: bool
    enableVoteEnsemble: bool
    ensembleModelDownloadTimeout: 'string'
    stackEnsembleSettings: {
      stackMetaLearnerKWargs: any(...)
      stackMetaLearnerTrainPercentage: int
      stackMetaLearnerType: 'string'
    }
  }
  validationData: {
    description: 'string'
    jobInputType: 'string'
    mode: 'string'
    uri: 'string'
  }
  validationDataSize: int
  weightColumnName: 'string'
}

針對 預測,請使用:

{
  cvSplitColumnNames: [
    'string'
  ]
  featurizationSettings: {
    blockedTransformers: [
      'string'
    ]
    columnNameAndTypes: {
      {customized property}: 'string'
    }
    datasetLanguage: 'string'
    enableDnnFeaturization: bool
    mode: 'string'
    transformerParams: {
      {customized property}: [
        {
          fields: [
            'string'
          ]
          parameters: any(...)
        }
      ]
    }
  }
  forecastingSettings: {
    countryOrRegionForHolidays: 'string'
    cvStepSize: int
    featureLags: 'string'
    forecastHorizon: {
      mode: 'string'
      // For remaining properties, see ForecastHorizon objects
    }
    frequency: 'string'
    seasonality: {
      mode: 'string'
      // For remaining properties, see Seasonality objects
    }
    shortSeriesHandlingConfig: 'string'
    targetAggregateFunction: 'string'
    targetLags: {
      mode: 'string'
      // For remaining properties, see TargetLags objects
    }
    targetRollingWindowSize: {
      mode: 'string'
      // For remaining properties, see TargetRollingWindowSize objects
    }
    timeColumnName: 'string'
    timeSeriesIdColumnNames: [
      'string'
    ]
    useStl: 'string'
  }
  limitSettings: {
    enableEarlyTermination: bool
    exitScore: int
    maxConcurrentTrials: int
    maxCoresPerTrial: int
    maxTrials: int
    timeout: 'string'
    trialTimeout: 'string'
  }
  nCrossValidations: {
    mode: 'string'
    // For remaining properties, see NCrossValidations objects
  }
  primaryMetric: 'string'
  taskType: 'Forecasting'
  testData: {
    description: 'string'
    jobInputType: 'string'
    mode: 'string'
    uri: 'string'
  }
  testDataSize: int
  trainingSettings: {
    allowedTrainingAlgorithms: [
      'string'
    ]
    blockedTrainingAlgorithms: [
      'string'
    ]
    enableDnnTraining: bool
    enableModelExplainability: bool
    enableOnnxCompatibleModels: bool
    enableStackEnsemble: bool
    enableVoteEnsemble: bool
    ensembleModelDownloadTimeout: 'string'
    stackEnsembleSettings: {
      stackMetaLearnerKWargs: any(...)
      stackMetaLearnerTrainPercentage: int
      stackMetaLearnerType: 'string'
    }
  }
  validationData: {
    description: 'string'
    jobInputType: 'string'
    mode: 'string'
    uri: 'string'
  }
  validationDataSize: int
  weightColumnName: 'string'
}

針對 ImageClassification,請使用:

{
  limitSettings: {
    maxConcurrentTrials: int
    maxTrials: int
    timeout: 'string'
  }
  modelSettings: {
    advancedSettings: 'string'
    amsGradient: bool
    augmentations: 'string'
    beta1: int
    beta2: int
    checkpointFrequency: int
    checkpointModel: {
      description: 'string'
      jobInputType: 'string'
      mode: 'string'
      uri: 'string'
    }
    checkpointRunId: 'string'
    distributed: bool
    earlyStopping: bool
    earlyStoppingDelay: int
    earlyStoppingPatience: int
    enableOnnxNormalization: bool
    evaluationFrequency: int
    gradientAccumulationStep: int
    layersToFreeze: int
    learningRate: int
    learningRateScheduler: 'string'
    modelName: 'string'
    momentum: int
    nesterov: bool
    numberOfEpochs: int
    numberOfWorkers: int
    optimizer: 'string'
    randomSeed: int
    stepLRGamma: int
    stepLRStepSize: int
    trainingBatchSize: int
    trainingCropSize: int
    validationBatchSize: int
    validationCropSize: int
    validationResizeSize: int
    warmupCosineLRCycles: int
    warmupCosineLRWarmupEpochs: int
    weightDecay: int
    weightedLoss: int
  }
  primaryMetric: 'string'
  searchSpace: [
    {
      amsGradient: 'string'
      augmentations: 'string'
      beta1: 'string'
      beta2: 'string'
      distributed: 'string'
      earlyStopping: 'string'
      earlyStoppingDelay: 'string'
      earlyStoppingPatience: 'string'
      enableOnnxNormalization: 'string'
      evaluationFrequency: 'string'
      gradientAccumulationStep: 'string'
      layersToFreeze: 'string'
      learningRate: 'string'
      learningRateScheduler: 'string'
      modelName: 'string'
      momentum: 'string'
      nesterov: 'string'
      numberOfEpochs: 'string'
      numberOfWorkers: 'string'
      optimizer: 'string'
      randomSeed: 'string'
      stepLRGamma: 'string'
      stepLRStepSize: 'string'
      trainingBatchSize: 'string'
      trainingCropSize: 'string'
      validationBatchSize: 'string'
      validationCropSize: 'string'
      validationResizeSize: 'string'
      warmupCosineLRCycles: 'string'
      warmupCosineLRWarmupEpochs: 'string'
      weightDecay: 'string'
      weightedLoss: 'string'
    }
  ]
  sweepSettings: {
    earlyTermination: {
      delayEvaluation: int
      evaluationInterval: int
      policyType: 'string'
      // For remaining properties, see EarlyTerminationPolicy objects
    }
    samplingAlgorithm: 'string'
  }
  taskType: 'ImageClassification'
  validationData: {
    description: 'string'
    jobInputType: 'string'
    mode: 'string'
    uri: 'string'
  }
  validationDataSize: int
}

針對 ImageClassificationMultilabel,請使用:

{
  limitSettings: {
    maxConcurrentTrials: int
    maxTrials: int
    timeout: 'string'
  }
  modelSettings: {
    advancedSettings: 'string'
    amsGradient: bool
    augmentations: 'string'
    beta1: int
    beta2: int
    checkpointFrequency: int
    checkpointModel: {
      description: 'string'
      jobInputType: 'string'
      mode: 'string'
      uri: 'string'
    }
    checkpointRunId: 'string'
    distributed: bool
    earlyStopping: bool
    earlyStoppingDelay: int
    earlyStoppingPatience: int
    enableOnnxNormalization: bool
    evaluationFrequency: int
    gradientAccumulationStep: int
    layersToFreeze: int
    learningRate: int
    learningRateScheduler: 'string'
    modelName: 'string'
    momentum: int
    nesterov: bool
    numberOfEpochs: int
    numberOfWorkers: int
    optimizer: 'string'
    randomSeed: int
    stepLRGamma: int
    stepLRStepSize: int
    trainingBatchSize: int
    trainingCropSize: int
    validationBatchSize: int
    validationCropSize: int
    validationResizeSize: int
    warmupCosineLRCycles: int
    warmupCosineLRWarmupEpochs: int
    weightDecay: int
    weightedLoss: int
  }
  primaryMetric: 'string'
  searchSpace: [
    {
      amsGradient: 'string'
      augmentations: 'string'
      beta1: 'string'
      beta2: 'string'
      distributed: 'string'
      earlyStopping: 'string'
      earlyStoppingDelay: 'string'
      earlyStoppingPatience: 'string'
      enableOnnxNormalization: 'string'
      evaluationFrequency: 'string'
      gradientAccumulationStep: 'string'
      layersToFreeze: 'string'
      learningRate: 'string'
      learningRateScheduler: 'string'
      modelName: 'string'
      momentum: 'string'
      nesterov: 'string'
      numberOfEpochs: 'string'
      numberOfWorkers: 'string'
      optimizer: 'string'
      randomSeed: 'string'
      stepLRGamma: 'string'
      stepLRStepSize: 'string'
      trainingBatchSize: 'string'
      trainingCropSize: 'string'
      validationBatchSize: 'string'
      validationCropSize: 'string'
      validationResizeSize: 'string'
      warmupCosineLRCycles: 'string'
      warmupCosineLRWarmupEpochs: 'string'
      weightDecay: 'string'
      weightedLoss: 'string'
    }
  ]
  sweepSettings: {
    earlyTermination: {
      delayEvaluation: int
      evaluationInterval: int
      policyType: 'string'
      // For remaining properties, see EarlyTerminationPolicy objects
    }
    samplingAlgorithm: 'string'
  }
  taskType: 'ImageClassificationMultilabel'
  validationData: {
    description: 'string'
    jobInputType: 'string'
    mode: 'string'
    uri: 'string'
  }
  validationDataSize: int
}

針對 ImageInstanceSegmentation,請使用:

{
  limitSettings: {
    maxConcurrentTrials: int
    maxTrials: int
    timeout: 'string'
  }
  modelSettings: {
    advancedSettings: 'string'
    amsGradient: bool
    augmentations: 'string'
    beta1: int
    beta2: int
    boxDetectionsPerImage: int
    boxScoreThreshold: int
    checkpointFrequency: int
    checkpointModel: {
      description: 'string'
      jobInputType: 'string'
      mode: 'string'
      uri: 'string'
    }
    checkpointRunId: 'string'
    distributed: bool
    earlyStopping: bool
    earlyStoppingDelay: int
    earlyStoppingPatience: int
    enableOnnxNormalization: bool
    evaluationFrequency: int
    gradientAccumulationStep: int
    imageSize: int
    layersToFreeze: int
    learningRate: int
    learningRateScheduler: 'string'
    maxSize: int
    minSize: int
    modelName: 'string'
    modelSize: 'string'
    momentum: int
    multiScale: bool
    nesterov: bool
    nmsIouThreshold: int
    numberOfEpochs: int
    numberOfWorkers: int
    optimizer: 'string'
    randomSeed: int
    stepLRGamma: int
    stepLRStepSize: int
    tileGridSize: 'string'
    tileOverlapRatio: int
    tilePredictionsNmsThreshold: int
    trainingBatchSize: int
    validationBatchSize: int
    validationIouThreshold: int
    validationMetricType: 'string'
    warmupCosineLRCycles: int
    warmupCosineLRWarmupEpochs: int
    weightDecay: int
  }
  primaryMetric: 'string'
  searchSpace: [
    {
      amsGradient: 'string'
      augmentations: 'string'
      beta1: 'string'
      beta2: 'string'
      boxDetectionsPerImage: 'string'
      boxScoreThreshold: 'string'
      distributed: 'string'
      earlyStopping: 'string'
      earlyStoppingDelay: 'string'
      earlyStoppingPatience: 'string'
      enableOnnxNormalization: 'string'
      evaluationFrequency: 'string'
      gradientAccumulationStep: 'string'
      imageSize: 'string'
      layersToFreeze: 'string'
      learningRate: 'string'
      learningRateScheduler: 'string'
      maxSize: 'string'
      minSize: 'string'
      modelName: 'string'
      modelSize: 'string'
      momentum: 'string'
      multiScale: 'string'
      nesterov: 'string'
      nmsIouThreshold: 'string'
      numberOfEpochs: 'string'
      numberOfWorkers: 'string'
      optimizer: 'string'
      randomSeed: 'string'
      stepLRGamma: 'string'
      stepLRStepSize: 'string'
      tileGridSize: 'string'
      tileOverlapRatio: 'string'
      tilePredictionsNmsThreshold: 'string'
      trainingBatchSize: 'string'
      validationBatchSize: 'string'
      validationIouThreshold: 'string'
      validationMetricType: 'string'
      warmupCosineLRCycles: 'string'
      warmupCosineLRWarmupEpochs: 'string'
      weightDecay: 'string'
    }
  ]
  sweepSettings: {
    earlyTermination: {
      delayEvaluation: int
      evaluationInterval: int
      policyType: 'string'
      // For remaining properties, see EarlyTerminationPolicy objects
    }
    samplingAlgorithm: 'string'
  }
  taskType: 'ImageInstanceSegmentation'
  validationData: {
    description: 'string'
    jobInputType: 'string'
    mode: 'string'
    uri: 'string'
  }
  validationDataSize: int
}

針對 ImageObjectDetection,請使用:

{
  limitSettings: {
    maxConcurrentTrials: int
    maxTrials: int
    timeout: 'string'
  }
  modelSettings: {
    advancedSettings: 'string'
    amsGradient: bool
    augmentations: 'string'
    beta1: int
    beta2: int
    boxDetectionsPerImage: int
    boxScoreThreshold: int
    checkpointFrequency: int
    checkpointModel: {
      description: 'string'
      jobInputType: 'string'
      mode: 'string'
      uri: 'string'
    }
    checkpointRunId: 'string'
    distributed: bool
    earlyStopping: bool
    earlyStoppingDelay: int
    earlyStoppingPatience: int
    enableOnnxNormalization: bool
    evaluationFrequency: int
    gradientAccumulationStep: int
    imageSize: int
    layersToFreeze: int
    learningRate: int
    learningRateScheduler: 'string'
    maxSize: int
    minSize: int
    modelName: 'string'
    modelSize: 'string'
    momentum: int
    multiScale: bool
    nesterov: bool
    nmsIouThreshold: int
    numberOfEpochs: int
    numberOfWorkers: int
    optimizer: 'string'
    randomSeed: int
    stepLRGamma: int
    stepLRStepSize: int
    tileGridSize: 'string'
    tileOverlapRatio: int
    tilePredictionsNmsThreshold: int
    trainingBatchSize: int
    validationBatchSize: int
    validationIouThreshold: int
    validationMetricType: 'string'
    warmupCosineLRCycles: int
    warmupCosineLRWarmupEpochs: int
    weightDecay: int
  }
  primaryMetric: 'string'
  searchSpace: [
    {
      amsGradient: 'string'
      augmentations: 'string'
      beta1: 'string'
      beta2: 'string'
      boxDetectionsPerImage: 'string'
      boxScoreThreshold: 'string'
      distributed: 'string'
      earlyStopping: 'string'
      earlyStoppingDelay: 'string'
      earlyStoppingPatience: 'string'
      enableOnnxNormalization: 'string'
      evaluationFrequency: 'string'
      gradientAccumulationStep: 'string'
      imageSize: 'string'
      layersToFreeze: 'string'
      learningRate: 'string'
      learningRateScheduler: 'string'
      maxSize: 'string'
      minSize: 'string'
      modelName: 'string'
      modelSize: 'string'
      momentum: 'string'
      multiScale: 'string'
      nesterov: 'string'
      nmsIouThreshold: 'string'
      numberOfEpochs: 'string'
      numberOfWorkers: 'string'
      optimizer: 'string'
      randomSeed: 'string'
      stepLRGamma: 'string'
      stepLRStepSize: 'string'
      tileGridSize: 'string'
      tileOverlapRatio: 'string'
      tilePredictionsNmsThreshold: 'string'
      trainingBatchSize: 'string'
      validationBatchSize: 'string'
      validationIouThreshold: 'string'
      validationMetricType: 'string'
      warmupCosineLRCycles: 'string'
      warmupCosineLRWarmupEpochs: 'string'
      weightDecay: 'string'
    }
  ]
  sweepSettings: {
    earlyTermination: {
      delayEvaluation: int
      evaluationInterval: int
      policyType: 'string'
      // For remaining properties, see EarlyTerminationPolicy objects
    }
    samplingAlgorithm: 'string'
  }
  taskType: 'ImageObjectDetection'
  validationData: {
    description: 'string'
    jobInputType: 'string'
    mode: 'string'
    uri: 'string'
  }
  validationDataSize: int
}

針對 回歸,請使用:

{
  cvSplitColumnNames: [
    'string'
  ]
  featurizationSettings: {
    blockedTransformers: [
      'string'
    ]
    columnNameAndTypes: {
      {customized property}: 'string'
    }
    datasetLanguage: 'string'
    enableDnnFeaturization: bool
    mode: 'string'
    transformerParams: {
      {customized property}: [
        {
          fields: [
            'string'
          ]
          parameters: any(...)
        }
      ]
    }
  }
  limitSettings: {
    enableEarlyTermination: bool
    exitScore: int
    maxConcurrentTrials: int
    maxCoresPerTrial: int
    maxTrials: int
    timeout: 'string'
    trialTimeout: 'string'
  }
  nCrossValidations: {
    mode: 'string'
    // For remaining properties, see NCrossValidations objects
  }
  primaryMetric: 'string'
  taskType: 'Regression'
  testData: {
    description: 'string'
    jobInputType: 'string'
    mode: 'string'
    uri: 'string'
  }
  testDataSize: int
  trainingSettings: {
    allowedTrainingAlgorithms: [
      'string'
    ]
    blockedTrainingAlgorithms: [
      'string'
    ]
    enableDnnTraining: bool
    enableModelExplainability: bool
    enableOnnxCompatibleModels: bool
    enableStackEnsemble: bool
    enableVoteEnsemble: bool
    ensembleModelDownloadTimeout: 'string'
    stackEnsembleSettings: {
      stackMetaLearnerKWargs: any(...)
      stackMetaLearnerTrainPercentage: int
      stackMetaLearnerType: 'string'
    }
  }
  validationData: {
    description: 'string'
    jobInputType: 'string'
    mode: 'string'
    uri: 'string'
  }
  validationDataSize: int
  weightColumnName: 'string'
}

針對 TextClassification,請使用:

{
  featurizationSettings: {
    datasetLanguage: 'string'
  }
  limitSettings: {
    maxConcurrentTrials: int
    maxTrials: int
    timeout: 'string'
  }
  primaryMetric: 'string'
  taskType: 'TextClassification'
  validationData: {
    description: 'string'
    jobInputType: 'string'
    mode: 'string'
    uri: 'string'
  }
}

針對 TextClassificationMultilabel,請使用:

{
  featurizationSettings: {
    datasetLanguage: 'string'
  }
  limitSettings: {
    maxConcurrentTrials: int
    maxTrials: int
    timeout: 'string'
  }
  taskType: 'TextClassificationMultilabel'
  validationData: {
    description: 'string'
    jobInputType: 'string'
    mode: 'string'
    uri: 'string'
  }
}

針對 TextNER,請使用:

{
  featurizationSettings: {
    datasetLanguage: 'string'
  }
  limitSettings: {
    maxConcurrentTrials: int
    maxTrials: int
    timeout: 'string'
  }
  taskType: 'TextNER'
  validationData: {
    description: 'string'
    jobInputType: 'string'
    mode: 'string'
    uri: 'string'
  }
}

屬性值

Microsoft.MachineLearningServices/workspaces/jobs

名稱 說明 價值觀
名稱 資源名稱 字串

約束:
模式 = ^[a-zA-Z0-9][a-zA-Z0-9\-_]{0,254}$ (必要)
父代 在 Bicep 中,您可以指定子資源的父資源。 只有在父資源外部宣告子資源時,才需要新增這個屬性。

如需詳細資訊,請參閱 父資源外部的子資源
類型資源的符號名稱: 工作區
屬性 [必要]實體的其他屬性。 JobBaseProperties (必要)

所有節點

名稱 說明 價值觀
nodesValueType [必要]Nodes 值的類型 'All' (必要)

AmlToken

名稱 說明 價值觀
身份類型 [必要]指定身分識別架構的類型。 'AMLToken' (必要)

自動預測地平線

名稱 說明 價值觀
模式 [必要]設定預測地平線值選取模式。 'Auto' (必要)

AutoMLJob

名稱 說明 價值觀
環境ID 作業之環境規格的 ARM 資源識別碼。
如果未提供,這是選擇性的值,AutoML 會在執行作業時,將此預設為 Production AutoML 策展環境版本。
字串
環境變數 作業中包含的環境變數。 AutoMLJobEnvironmentVariables
職位類型 [必要]指定作業的類型。 'AutoML' (必要)
輸出 對應作業中使用的輸出數據系結。 AutoMLJob 輸出
queueSettings 作業的佇列設定 佇列設置
資源 作業的計算資源組態。 JobResourceConfiguration
taskDetails [必要]這代表可以是其中一個數據表/NLP/Image 的案例 AutoMLVertical (必要)

AutoMLJobEnvironmentVariables

名稱 說明 價值觀

AutoMLJob 輸出

名稱 說明 價值觀

AutoMLVertical

名稱 說明 價值觀
logVerbosity (對數詳細程度) 作業的記錄詳細資訊。 “嚴重”
“調試”
“錯誤”
'資訊'
'NotSet'
“警告”
targetColumnName 目標數據行名稱:這是預測值數據行。
也稱為分類工作內容中的標籤數據行名稱。
字串
任務類型 設定為 [分類] 類型為 [ 分類]。 設定為 [預測] 類型為 [ 預測]。 設定為 ImageClassification 類型的 『ImageClassification』。 設定為 ImageClassificationMultilabel 類型的 'ImageClassificationMultilabel'。 針對 ImageInstanceSegmentation 類型,設定為 'ImageInstanceSegmentation'。 針對 ImageObjectDetection 類型,設定為 'ImageObjectDetection'。 將類型回歸設定為 [ 回歸]。 設定為 TextClassification 類型的 『TextClassification』。 針對 TextClassificationMultilabel 類型,設定為 『TextClassificationMultilabel』。 設定為 TextNer 類型的 『TextNER』。 '分類'
「預測」
'ImageClassification' (圖像分類)
'ImageClassificationMultilabel'
'ImageInstanceSegmentation' (圖像實例分段)
'ImageObjectDetection' (圖像對象檢測)
“回歸”
'TextClassification' (文本分類)
'TextClassificationMultilabel'
'TextNER' (必要)
trainingData (訓練數據) [必要]定型數據輸入。 MLTableJobInput (必要)

AutoNCross驗證

名稱 說明 價值觀
模式 [必要]判斷 N-Cross 驗證的模式。 'Auto' (必要)

自動季節性

名稱 說明 價值觀
模式 [必要]季節性模式。 'Auto' (必要)

AutoTargetLags 自動目標滯後

名稱 說明 價值觀
模式 [必要]設定目標延隔模式 - 自動/自定義 'Auto' (必要)

AutoTargetRollingWindowSize (自動目標滾動視窗大小)

名稱 說明 價值觀
模式 [必要]TargetRollingWindowSiz 偵測模式。 'Auto' (必要)

AzureDevOpsWebhook

名稱 說明 價值觀
webhook類型 [必要]指定要傳送回呼的服務類型 'AzureDevOps' (必要)

AzureOpenAiFineTuning

名稱 說明 價值觀
超參數 用於微調 Azure Open AI 模型的 HyperParameters。 AzureOpenAiHyperParameters
模型提供程式 [必要]用來判斷微調類型的列舉。 'AzureOpenAI' (必要)

AzureOpenAiHyperParameters

名稱 說明 價值觀
批次大小 每個批次中的範例數目。 批次大小越大,表示模型參數的更新頻率越低,而變異數越少。 整數 (int)
learningRateMultiplier 的 學習率的縮放比例。 較小的學習率可能有助於避免過度調整。 整數 (int)
nEpochs 訓練模型的週期數目。 Epoch 是指透過訓練資料集的一個完整週期。 整數 (int)

強盜政策 (BanditPolicy)

名稱 說明 價值觀
policyType [必要]原則設定的名稱 “強盜”(必要)
slackAmount 從最佳執行執行中允許的絕對距離。 整數 (int)
slackFactor 的 與最佳執行距離的允許距離比率。 整數 (int)

BayesianSamplingAlgorithm

名稱 說明 價值觀
採樣演算法類型 [必要]用來產生超參數值的演算法,以及組態屬性 “貝氏” (必要)

分類

名稱 說明 價值觀
cvSplitColumnNames 要用於 CVSplit 資料的數據行。 字串[]
featurizationSettings AutoML 作業所需的特徵化輸入。 TableVerticalFeaturizationSettings
limitSettings AutoMLJob 的執行條件約束。 TableVerticalLimitSettings
nCrossValidations 要套用至定型數據集的交叉驗證折疊數目
未提供驗證數據集時。
NCross驗證
積極標籤 二進位計量計算的正標籤。 字串
primaryMetric 工作的主要計量。 '精確度'
'AUCWeighted'
'AveragePrecisionScoreWeighted'
'NormMacroRecall' (規範巨集調用)
'PrecisionScoreWeighted'
任務類型 [必要]AutoMLJob 的工作類型。 '分類' (必要)
測試數據 測試數據輸入。 MLTableJobInput
testDataSize 必須為驗證目的而保留的測試數據集分數。
介於 (0.0 、 1.0) 之間的值
未提供驗證數據集時套用。
整數 (int)
trainingSettings AutoML 作業定型階段的輸入。 分類訓練設置
validationData (驗證資料) 驗證數據輸入。 MLTableJobInput
validationDataSize 需要為驗證目的預留的定型數據集分數。
介於 (0.0 、 1.0) 之間的值
未提供驗證數據集時套用。
整數 (int)
權重列名稱 範例權數數據行的名稱。 自動化 ML 支援加權數據行做為輸入,導致數據中的數據列向上或向下加權。 字串

分類訓練設置

名稱 說明 價值觀
allowedTrainingAlgorithms 分類工作的允許模型。 包含任何的字串數組:
'伯努利樸素貝葉斯'
“決策樹”
'ExtremeRandomTrees'
'梯度提升'
'KNN'
'LightGBM'
'線性 SVM'
'LogisticRegression' (邏輯回歸)
'多項式樸素貝葉斯'
'隨機森林'
'新幣'
“SVM”
'XGBoostClassifier'
blockedTrainingAlgorithms 分類工作的封鎖模型。 包含任何的字串數組:
'伯努利樸素貝葉斯'
“決策樹”
'ExtremeRandomTrees'
'梯度提升'
'KNN'
'LightGBM'
'線性 SVM'
'LogisticRegression' (邏輯回歸)
'多項式樸素貝葉斯'
'隨機森林'
'新幣'
“SVM”
'XGBoostClassifier'
enableDnnTraining 啟用 DNN 模型的建議。 布爾 (bool)
enableModelExplainability 在最佳模型上開啟可解釋性的旗標。 布爾 (bool)
enableOnnxCompatibleModels 啟用 onnx 相容模型的旗標。 布爾 (bool)
enableStackEnsemble 啟用堆疊合奏執行。 布爾 (bool)
enableVoteEnsemble 啟用投票合奏執行。 布爾 (bool)
ensembleModelDownloadTimeout 在 VotingEnsemble 和 StackEnsemble 模型產生期間,會下載先前子回合中的多個安裝模型。
如果需要更多時間,請使用高於 300 秒的值來設定此參數。
字串
stackEnsembleSettings 堆疊合奏執行的堆疊合奏設定。 StackEnsembleSettings

ColumnTransformer (列變換器)

名稱 說明 價值觀
領域 要套用轉換器邏輯的欄位。 字串[]
參數 要傳遞至轉換器的不同屬性。
輸入必須是 JSON 格式的索引鍵、值組字典。
任何

CommandJob

名稱 說明 價值觀
代碼Id 程式代碼資產的 ARM 資源識別碼。 字串
指令 [必要]在作業啟動時執行的命令。 例如 “python train.py” 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)
分配 作業的散發組態。 如果設定,這應該是其中一個 Mpi、Tensorflow、PyTorch 或 null。 分發配置
環境ID [必要]作業之環境規格的 ARM 資源識別碼。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)
環境變數 作業中包含的環境變數。 CommandJobEnvironmentVariables
輸入 對應作業中使用的輸入數據系結。 CommandJobInputs 命令作業輸入
職位類型 [必要]指定作業的類型。 'Command' (必要)
限制 命令作業限制。 CommandJobLimits 命令作業限制
輸出 對應作業中使用的輸出數據系結。 CommandJobOutputs 命令作業輸出
queueSettings 作業的佇列設定 佇列設置
資源 作業的計算資源組態。 JobResourceConfiguration

CommandJobEnvironmentVariables

名稱 說明 價值觀

CommandJobInputs 命令作業輸入

名稱 說明 價值觀

CommandJobLimits 命令作業限制

名稱 說明 價值觀
jobLimits類型 [必要]JobLimit 類型。 “命令”
'掃掠' (必要)
超時 ISO 8601 格式的最大執行持續時間,之後作業將會取消。 僅支援有效位數為秒數的持續時間。 字串

CommandJobOutputs 命令作業輸出

名稱 說明 價值觀

CustomForecastHorizon

名稱 說明 價值觀
模式 [必要]設定預測地平線值選取模式。 'Custom' (必要)
價值 [必要]預測地平線值。 int (必要)

CustomModelFineTuning 微調

名稱 說明 價值觀
超參數 微調自定義模型的 HyperParameters。 CustomModelFineTuningHyperParameters
模型提供程式 [必要]用來判斷微調類型的列舉。 'Custom' (必要)

CustomModelFineTuningHyperParameters

名稱 說明 價值觀

CustomModelJobInput (自定義模型作業輸入)

名稱 說明 價值觀
jobInputType (作業輸入類型) [必要]指定作業的類型。 'custom_model' (必要)
模式 輸入資產傳遞模式。 “直接”
'下載'
'EvalDownload'
'EvalMount'
'ReadOnlyMount'
“ReadWriteMount”
統一資源識別碼 (URI) [必要]輸入資產 URI。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)

CustomModelJobOutput

名稱 說明 價值觀
jobOutputType [必要]指定作業的類型。 'custom_model' (必要)
模式 輸出資產傳遞模式。 “直接”
“ReadWriteMount”
'上傳'
統一資源識別碼 (URI) 輸出資產 URI。 字串

CustomNCross驗證

名稱 說明 價值觀
模式 [必要]判斷 N-Cross 驗證的模式。 'Custom' (必要)
價值 [必要]N-Cross 驗證值。 int (必要)

CustomSeasonality (自定義季節性)

名稱 說明 價值觀
模式 [必要]季節性模式。 'Custom' (必要)
價值 [必要]季節性值。 int (必要)

CustomTargetLags (自定義目標滯後)

名稱 說明 價值觀
模式 [必要]設定目標延隔模式 - 自動/自定義 'Custom' (必要)
價值觀 [必要]設定目標延遲值。 int[] (必要)

CustomTargetRollingWindowSize

名稱 說明 價值觀
模式 [必要]TargetRollingWindowSiz 偵測模式。 'Custom' (必要)
價值 [必要]TargetRollingWindowSize 值。 int (必要)

分發配置

名稱 說明 價值觀
distributionType 設定為 Mpi 類型的 『Mpi』。 將類型 PyTorch 設定為 『PyTorch』。 將 設定為 TensorFlow 類型的 『TensorFlow』。 'Mpi'
“PyTorch”
'TensorFlow' (必要)

提早終止政策

名稱 說明 價值觀
delay評估 延遲第一次評估的間隔數目。 整數 (int)
evaluationInterval 原則評估之間的間隔(執行次數)。 整數 (int)
policyType 針對 BanditPolicy 類型,設定為 『Bandit』。 針對 MedianStoppingPolicy 類型,設定為 'MedianStoppingPolicy'。 針對 TruncationSelectionPolicy 類型,設定為 『TruncationSelection』 “強盜”
'MedianStopping' (中位數停止)
'TruncationSelection' (必要)

微調作業

名稱 說明 價值觀
微調細節 [必要] FineTuningVertical (必要)
職位類型 [必要]指定作業的類型。 'FineTuning' (必要)
輸出 [必要] FineTuningJobOutputs (必要)
queueSettings 作業的佇列設定 佇列設置
資源 作業的實例類型和其他資源 工作資源

FineTuningJobOutputs(微調作業輸出)

名稱 說明 價值觀

微調垂直

名稱 說明 價值觀
型號 [必要]用於微調的輸入模型。 MLFlowModelJobInput (必要)
模型提供程式 針對 AzureOpen AiFineTuning 類型,設定為 『AzureOpenAI』。 設定為 CustomModelFineTuning 類型的 [自定義]。 'AzureOpenAI'
'Custom' (必要)
任務類型 [必要]微調工作類型。 'ChatCompletion'
'ImageClassification' (圖像分類)
'ImageInstanceSegmentation' (圖像實例分段)
'ImageObjectDetection' (圖像對象檢測)
“問答”
'TextClassification' (文本分類)
'TextCompletion'
'TextSummarization' (文本摘要)
'TextTranslation'
'TokenClassification'
'VideoMultiObjectTracking' (必要)
trainingData (訓練數據) [必要]定型數據以進行微調。 JobInput (必要)
validationData (驗證資料) 微調的驗證數據。 作業輸入

預測地平線

名稱 說明 價值觀
模式 針對 AutoForecastHorizon 類型,設定為 'Auto'。 針對 CustomForecastHorizon 類型,設定為 'Custom'。 '自動'
'Custom' (必要)

預測

名稱 說明 價值觀
cvSplitColumnNames 要用於 CVSplit 資料的數據行。 字串[]
featurizationSettings AutoML 作業所需的特徵化輸入。 TableVerticalFeaturizationSettings
forecasting設置 預測工作特定輸入。 ForecastingSettings
limitSettings AutoMLJob 的執行條件約束。 TableVerticalLimitSettings
nCrossValidations 要套用至定型數據集的交叉驗證折疊數目
未提供驗證數據集時。
NCross驗證
primaryMetric 預測工作的主要計量。 'NormalizedMeanAbsoluteError'
'NormalizedRootMeanSquaredError'
'R2Score'
'SpearmanCorrelation'
任務類型 [必要]AutoMLJob 的工作類型。 “預測” (必要)
測試數據 測試數據輸入。 MLTableJobInput
testDataSize 必須為驗證目的而保留的測試數據集分數。
介於 (0.0 、 1.0) 之間的值
未提供驗證數據集時套用。
整數 (int)
trainingSettings AutoML 作業定型階段的輸入。 ForecastingTrainingSettings
validationData (驗證資料) 驗證數據輸入。 MLTableJobInput
validationDataSize 需要為驗證目的預留的定型數據集分數。
介於 (0.0 、 1.0) 之間的值
未提供驗證數據集時套用。
整數 (int)
權重列名稱 範例權數數據行的名稱。 自動化 ML 支援加權數據行做為輸入,導致數據中的數據列向上或向下加權。 字串

ForecastingSettings

名稱 說明 價值觀
國家/地區或地區為假期 用於預測工作的假日國家或地區。
這些應該是 ISO 3166 兩個字母的國家/地區代碼,例如 'US' 或 'GB'。
字串
cvStepSize 一個 CV 折疊的原點時間與下一個折疊之間的期間數。 為了
例如,如果 CVStepSize = 3 表示每日數據,則每個折疊的原點時間將會是
相隔三天。
整數 (int)
featureLags 使用 『auto』 或 null 產生數值特徵延遲的旗標。 '自動'
“無”
forecastHorizon (預測地平線) 所需的最大預測範圍,以時間序列頻率單位為單位。 預測地平線
頻率 預測時,此參數代表預期預測的期間,例如每日、每周、每年等。預測頻率預設為數據集頻率。 字串
季節性 將時間序列季節性設定為數列頻率的整數倍數。
如果季節性設定為 『auto』,則會推斷它。
季節性
shortSeriesHandlingConfig 定義 AutoML 是否應該處理短時間序列的參數。 '自動'
“掉落”
“無”
'墊'
targetAggregateFunction 要用來匯總時間序列目標數據行以符合使用者指定頻率的函式。
如果 TargetAggregateFunction 已設定,亦即不是 『None』,但未設定 freq 參數,則會引發錯誤。 可能的目標聚合函數包括:「sum」、“max”、“min” 和 “mean”。
“最大”
“卑鄙”
'敏'
“無”
'總和'
目標拉格 要從目標數據行延遲的過去期間數。 目標滯後
targetRollingWindowSize 用來建立目標數據行滾動視窗平均值的過去期間數目。 TargetRollingWindowSize
timeColumnName (時間列名稱) 時間數據行的名稱。 當預測以指定用於建置時間序列並推斷其頻率的輸入數據中的 datetime 數據行時,需要此參數。 字串
timeSeriesIdColumnNames 用來分組時間範圍的數據行名稱。 它可以用來建立多個數列。
如果未定義粒紋,則會假設數據集為一個時間序列。 此參數與工作類型預測搭配使用。
字串[]
useStl 設定時間序列目標數據行的 STL 分解。 “無”
“賽季”
'季節趨勢'

ForecastingTrainingSettings

名稱 說明 價值觀
allowedTrainingAlgorithms 用於預測工作的允許模型。 包含任何的字串數組:
“Arimax”
'AutoArima'
“平均”
“決策樹”
“彈性網”
'指數平滑'
'ExtremeRandomTrees'
'梯度提升'
'KNN'
“套索拉斯”
'LightGBM'
天真'
'先知'
'隨機森林'
'季節性平均值'
'季節性天真'
'新幣'
'TCNForecaster'
'XGBoostRegressor'
blockedTrainingAlgorithms 用於預測工作的封鎖模型。 包含任何的字串數組:
“Arimax”
'AutoArima'
“平均”
“決策樹”
“彈性網”
'指數平滑'
'ExtremeRandomTrees'
'梯度提升'
'KNN'
“套索拉斯”
'LightGBM'
天真'
'先知'
'隨機森林'
'季節性平均值'
'季節性天真'
'新幣'
'TCNForecaster'
'XGBoostRegressor'
enableDnnTraining 啟用 DNN 模型的建議。 布爾 (bool)
enableModelExplainability 在最佳模型上開啟可解釋性的旗標。 布爾 (bool)
enableOnnxCompatibleModels 啟用 onnx 相容模型的旗標。 布爾 (bool)
enableStackEnsemble 啟用堆疊合奏執行。 布爾 (bool)
enableVoteEnsemble 啟用投票合奏執行。 布爾 (bool)
ensembleModelDownloadTimeout 在 VotingEnsemble 和 StackEnsemble 模型產生期間,會下載先前子回合中的多個安裝模型。
如果需要更多時間,請使用高於 300 秒的值來設定此參數。
字串
stackEnsembleSettings 堆疊合奏執行的堆疊合奏設定。 StackEnsembleSettings

GridSampling演算法

名稱 說明 價值觀
採樣演算法類型 [必要]用來產生超參數值的演算法,以及組態屬性 'Grid' (必要)

身份配置

名稱 說明 價值觀
身份類型 將 設定為 AmlToken 類型的 『AMLToken』。 針對 ManagedIdentity 類型設定為 『Managed』。 針對 UserIdentity 類型,設定為 ' UserIdentity' 'AMLToken'
“託管”
'UserIdentity' (必要)

圖像分類

名稱 說明 價值觀
limitSettings [必要]限制 AutoML 作業的設定。 ImageLimitSettings (必要)
modelSettings 用於定型模型的設定。 ImageModelSettings分類
primaryMetric 要針對這項工作進行優化的主要計量。 '精確度'
'AUCWeighted'
'AveragePrecisionScoreWeighted'
'NormMacroRecall' (規範巨集調用)
'PrecisionScoreWeighted'
搜索空間 搜尋空間以取樣模型及其超參數的不同組合。 ImageModelDistributionSettingsClassification[]
掃描設置 模型掃掠和超參數掃掠相關設定。 ImageSweepSettings (圖像掃描設定)
任務類型 [必要]AutoMLJob 的工作類型。 'ImageClassification' (必要)
validationData (驗證資料) 驗證數據輸入。 MLTableJobInput
validationDataSize 需要為驗證目的預留的定型數據集分數。
介於 (0.0 、 1.0) 之間的值
未提供驗證數據集時套用。
整數 (int)

ImageClassificationMultilabel

名稱 說明 價值觀
limitSettings [必要]限制 AutoML 作業的設定。 ImageLimitSettings (必要)
modelSettings 用於定型模型的設定。 ImageModelSettings分類
primaryMetric 要針對這項工作進行優化的主要計量。 '精確度'
'AUCWeighted'
'AveragePrecisionScoreWeighted'
'欠條'
'NormMacroRecall' (規範巨集調用)
'PrecisionScoreWeighted'
搜索空間 搜尋空間以取樣模型及其超參數的不同組合。 ImageModelDistributionSettingsClassification[]
掃描設置 模型掃掠和超參數掃掠相關設定。 ImageSweepSettings (圖像掃描設定)
任務類型 [必要]AutoMLJob 的工作類型。 'ImageClassificationMultilabel' (必要)
validationData (驗證資料) 驗證數據輸入。 MLTableJobInput
validationDataSize 需要為驗證目的預留的定型數據集分數。
介於 (0.0 、 1.0) 之間的值
未提供驗證數據集時套用。
整數 (int)

ImageInstanceSegmentation

名稱 說明 價值觀
limitSettings [必要]限制 AutoML 作業的設定。 ImageLimitSettings (必要)
modelSettings 用於定型模型的設定。 ImageModelSettingsObjectDetection
primaryMetric 要針對這項工作進行優化的主要計量。 'MeanAveragePrecision'
搜索空間 搜尋空間以取樣模型及其超參數的不同組合。 ImageModelDistributionSettingsObjectDetection[]
掃描設置 模型掃掠和超參數掃掠相關設定。 ImageSweepSettings (圖像掃描設定)
任務類型 [必要]AutoMLJob 的工作類型。 'ImageInstanceSegmentation' (必要)
validationData (驗證資料) 驗證數據輸入。 MLTableJobInput
validationDataSize 需要為驗證目的預留的定型數據集分數。
介於 (0.0 、 1.0) 之間的值
未提供驗證數據集時套用。
整數 (int)

ImageLimitSettings

名稱 說明 價值觀
maxConcurrentTrials 並行 AutoML 反覆項目的數目上限。 整數 (int)
maxTrials AutoML 反覆項目的數目上限。 整數 (int)
超時 AutoML 作業逾時。 字串

ImageModelDistributionSettings分類

名稱 說明 價值觀
amsGradient 當優化器為 『adam』 或 'adamw' 時啟用AMSGrad。 字串
擴增 使用擴增的設定。 字串
貝塔1 優化器為 'adam' 或 'adamw' 時的 'beta1' 值。 必須是範圍 [0, 1] 內的浮點數。 字串
貝塔2 優化器為 'adam' 或 'adamw' 時的 'beta2' 值。 必須是範圍 [0, 1] 內的浮點數。 字串
分散式 是否要使用散發器定型。 字串
earlyStopping (早期停止) 在定型期間啟用早期停止邏輯。 字串
earlyStoppingDelay 在主要計量改進之前等待的最小 Epoch 或驗證評估數目
會追蹤早期停止。 必須是正整數。
字串
早點停止耐心 在之前沒有主要計量改進的最小 Epoch 或驗證評估數目
執行已停止。 必須是正整數。
字串
enableOnnxNormalization 匯出 ONNX 模型時啟用正規化。 字串
evaluationFrequency 評估驗證數據集以取得計量分數的頻率。 必須是正整數。 字串
梯度累積步驟 漸層累積表示執行未設定數目的 「GradAccumulationStep」 步驟
更新模型權數,同時累積這些步驟的漸層,然後使用
用來計算權數更新的累積漸層。 必須是正整數。
字串
layersToFreeze (圖層凍結) 要凍結模型的圖層數目。 必須是正整數。
例如,傳遞 2 作為 'seresnext' 值表示
凍結 layer0 和 layer1。 如需支援模型的完整清單,以及圖層凍結的詳細數據,請
請參閱:/azure/machine-learning/how-to-auto-train-image-models。
字串
learningRate 初始學習速率。 必須是範圍 [0, 1] 內的浮點數。 字串
learningRateScheduler 的 學習率排程器的類型。 必須是 『warmup_cosine』 或 'step'。 字串
型號名稱 要用於定型的模型名稱。
如需可用模型的詳細資訊,請流覽官方檔:
/azure/machine-learning/how-to-auto-train-image-models。
字串
動量 優化器為 『zn』 時的動量值。 必須是範圍 [0, 1] 內的浮點數。 字串
涅斯特羅夫 在優化器為 『zn』 時啟用 nesterov。 字串
紀元數 定型 Epoch 的數目。 必須是正整數。 字串
工人數 數據載入器背景工作角色的數目。 必須是非負整數。 字串
優化器 優化工具的類型。 必須是 『zn』、『adam』或 『adamw』。 字串
隨機種子 使用決定性定型時要使用的隨機種子。 字串
stepLR伽瑪 學習速率排程器為 『step』 時的 gamma 值。 必須是範圍 [0, 1] 內的浮點數。 字串
stepLRStepSize 學習速率排程器為 『step』 時的步驟大小值。 必須是正整數。 字串
trainingBatchSize 定型批次大小。 必須是正整數。 字串
trainingCropSize 影像裁剪大小,這是訓練數據集之類神經網路的輸入。 必須是正整數。 字串
validationBatchSize 驗證批次大小。 必須是正整數。 字串
validationCropSize 影像裁剪大小,這是用於驗證數據集之類神經網路的輸入。 必須是正整數。 字串
validationResizeSize 在裁剪驗證數據集之前要重設大小的影像大小。 必須是正整數。 字串
warmupCosineLRCycles 學習速率排程器為「warmup_cosine」時,餘弦迴圈的值。 必須是範圍 [0, 1] 內的浮點數。 字串
warmupCosineLRWarmupEpochs 學習速率排程器為「warmup_cosine」時,熱身 Epoch 的值。 必須是正整數。 字串
weightDecay (權重衰減) 優化器為 'zn'、'adam' 或 'adamw' 時的權數衰變值。 必須是範圍[0, 1] 中的浮點數。 字串
weightedLoss 加權減肥。 接受的值是 0,沒有加權損失。
1 用於與 sqrt 減肥。(class_weights)。 2 用於減肥與class_weights。 必須是 0 或 1 或 2。
字串

ImageModelDistributionSettingsObjectDetection

名稱 說明 價值觀
amsGradient 當優化器為 『adam』 或 'adamw' 時啟用AMSGrad。 字串
擴增 使用擴增的設定。 字串
貝塔1 優化器為 'adam' 或 'adamw' 時的 'beta1' 值。 必須是範圍 [0, 1] 內的浮點數。 字串
貝塔2 優化器為 'adam' 或 'adamw' 時的 'beta2' 值。 必須是範圍 [0, 1] 內的浮點數。 字串
boxDetectionsPerImage (每影像檢測) 所有類別的每個影像偵測數目上限。 必須是正整數。
注意:'yolov5' 演算法不支援此設定。
字串
boxScoreThreshold (盒分數閾值) 在推斷期間,只傳回大於分類分數的建議
BoxScoreThreshold。 必須是範圍[0, 1] 中的浮點數。
字串
分散式 是否要使用散發器定型。 字串
earlyStopping (早期停止) 在定型期間啟用早期停止邏輯。 字串
earlyStoppingDelay 在主要計量改進之前等待的最小 Epoch 或驗證評估數目
會追蹤早期停止。 必須是正整數。
字串
早點停止耐心 在之前沒有主要計量改進的最小 Epoch 或驗證評估數目
執行已停止。 必須是正整數。
字串
enableOnnxNormalization 匯出 ONNX 模型時啟用正規化。 字串
evaluationFrequency 評估驗證數據集以取得計量分數的頻率。 必須是正整數。 字串
梯度累積步驟 漸層累積表示執行未設定數目的 「GradAccumulationStep」 步驟
更新模型權數,同時累積這些步驟的漸層,然後使用
用來計算權數更新的累積漸層。 必須是正整數。
字串
影像大小 定型和驗證的影像大小。 必須是正整數。
注意:如果大小太大,定型回合可能會進入 CUDA OOM。
注意:只有 'yolov5' 演算法才支援此設定。
字串
layersToFreeze (圖層凍結) 要凍結模型的圖層數目。 必須是正整數。
例如,傳遞 2 作為 'seresnext' 值表示
凍結 layer0 和 layer1。 如需支援模型的完整清單,以及圖層凍結的詳細數據,請
請參閱:/azure/machine-learning/how-to-auto-train-image-models。
字串
learningRate 初始學習速率。 必須是範圍 [0, 1] 內的浮點數。 字串
learningRateScheduler 的 學習率排程器的類型。 必須是 『warmup_cosine』 或 'step'。 字串
最大尺寸 在將影像饋送至骨幹之前,要重新調整的影像大小上限。
必須是正整數。 注意:如果大小太大,定型執行可能會進入 CUDA OOM。
注意:'yolov5' 演算法不支援此設定。
字串
最小大小 將影像饋送至骨幹之前,要重新調整的影像大小下限。
必須是正整數。 注意:如果大小太大,定型執行可能會進入 CUDA OOM。
注意:'yolov5' 演算法不支援此設定。
字串
型號名稱 要用於定型的模型名稱。
如需可用模型的詳細資訊,請流覽官方檔:
/azure/machine-learning/how-to-auto-train-image-models。
字串
modelSize 模型大小。 必須是 'small'、'medium'、'large' 或 'xlarge'。
注意:如果模型大小太大,定型執行可能會進入 CUDA OOM。
注意:只有 'yolov5' 演算法才支援此設定。
字串
動量 優化器為 『zn』 時的動量值。 必須是範圍 [0, 1] 內的浮點數。 字串
多尺度 依 +/- 50%來啟用多重縮放影像大小。
注意:如果沒有足夠的 GPU 記憶體,定型執行可能會進入 CUDA OOM。
注意:只有 'yolov5' 演算法才支援此設定。
字串
涅斯特羅夫 在優化器為 『zn』 時啟用 nesterov。 字串
nmsIouThreshold 在NMS後置處理中的推斷期間使用的IOU臨界值。 必須是範圍 [0, 1] 中的 float。 字串
紀元數 定型 Epoch 的數目。 必須是正整數。 字串
工人數 數據載入器背景工作角色的數目。 必須是非負整數。 字串
優化器 優化工具的類型。 必須是 『zn』、『adam』或 『adamw』。 字串
隨機種子 使用決定性定型時要使用的隨機種子。 字串
stepLR伽瑪 學習速率排程器為 『step』 時的 gamma 值。 必須是範圍 [0, 1] 內的浮點數。 字串
stepLRStepSize 學習速率排程器為 『step』 時的步驟大小值。 必須是正整數。 字串
tileGridSize (平鋪網格大小) 用來貼上每個影像的網格線大小。 注意:TileGridSize 不得為
無可啟用小型物件偵測邏輯。 包含兩個整數的字串,其格式為一個。
注意:'yolov5' 演算法不支援此設定。
字串
tileOverlapRatio (平鋪重疊比率) 每個維度中相鄰磚之間的重疊比例。 必須是範圍 [0, 1] 中的浮點數。
注意:'yolov5' 演算法不支援此設定。
字串
tilePredictionsNmsThreshold 用來執行NMS的IOU閾值,同時合併磚和影像的預測。
用於驗證/推斷。 必須是範圍 [0, 1] 中的 float。
注意:'yolov5' 演算法不支援此設定。
NMS:非最大歸併
字串
trainingBatchSize 定型批次大小。 必須是正整數。 字串
validationBatchSize 驗證批次大小。 必須是正整數。 字串
validationIouThreshold 計算驗證計量時要使用的IOU閾值。 必須是範圍 [0, 1] 中的 float。 字串
validationMetricType 用於驗證計量的計量計算方法。 必須是 『none』、『coco』、'voc'或 'coco_voc'。 字串
warmupCosineLRCycles 學習速率排程器為「warmup_cosine」時,餘弦迴圈的值。 必須是範圍 [0, 1] 內的浮點數。 字串
warmupCosineLRWarmupEpochs 學習速率排程器為「warmup_cosine」時,熱身 Epoch 的值。 必須是正整數。 字串
weightDecay (權重衰減) 優化器為 'zn'、'adam' 或 'adamw' 時的權數衰變值。 必須是範圍[0, 1] 中的浮點數。 字串

ImageModelSettings分類

名稱 說明 價值觀
高級設置 進階案例的設定。 字串
amsGradient 當優化器為 『adam』 或 'adamw' 時啟用AMSGrad。 布爾 (bool)
擴增 使用擴增的設定。 字串
貝塔1 優化器為 'adam' 或 'adamw' 時的 'beta1' 值。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
貝塔2 優化器為 'adam' 或 'adamw' 時的 'beta2' 值。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
檢查點頻率 儲存模型檢查點的頻率。 必須是正整數。 整數 (int)
checkpoint模型 累加訓練的預先定型檢查點模型。 MLFlowModelJobInput
檢查點RunId 先前執行的標識碼,其具有累加式定型的預先定型檢查點。 字串
分散式 是否要使用分散式定型。 布爾 (bool)
earlyStopping (早期停止) 在定型期間啟用早期停止邏輯。 布爾 (bool)
earlyStoppingDelay 在主要計量改進之前等待的最小 Epoch 或驗證評估數目
會追蹤早期停止。 必須是正整數。
整數 (int)
早點停止耐心 在之前沒有主要計量改進的最小 Epoch 或驗證評估數目
執行已停止。 必須是正整數。
整數 (int)
enableOnnxNormalization 匯出 ONNX 模型時啟用正規化。 布爾 (bool)
evaluationFrequency 評估驗證數據集以取得計量分數的頻率。 必須是正整數。 整數 (int)
梯度累積步驟 漸層累積表示執行未設定數目的 「GradAccumulationStep」 步驟
更新模型權數,同時累積這些步驟的漸層,然後使用
用來計算權數更新的累積漸層。 必須是正整數。
整數 (int)
layersToFreeze (圖層凍結) 要凍結模型的圖層數目。 必須是正整數。
例如,傳遞 2 作為 'seresnext' 值表示
凍結 layer0 和 layer1。 如需支援模型的完整清單,以及圖層凍結的詳細數據,請
請參閱:/azure/machine-learning/how-to-auto-train-image-models。
整數 (int)
learningRate 初始學習速率。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
learningRateScheduler 的 學習率排程器的類型。 必須是 『warmup_cosine』 或 'step'。 “無”
“步驟”
'WarmupCosine'
型號名稱 要用於定型的模型名稱。
如需可用模型的詳細資訊,請流覽官方檔:
/azure/machine-learning/how-to-auto-train-image-models。
字串
動量 優化器為 『zn』 時的動量值。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
涅斯特羅夫 在優化器為 『zn』 時啟用 nesterov。 布爾 (bool)
紀元數 定型 Epoch 的數目。 必須是正整數。 整數 (int)
工人數 數據載入器背景工作角色的數目。 必須是非負整數。 整數 (int)
優化器 優化工具的類型。 “亞當”
“亞當”
“無”
'新元'
隨機種子 使用決定性定型時要使用的隨機種子。 整數 (int)
stepLR伽瑪 學習速率排程器為 『step』 時的 gamma 值。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
stepLRStepSize 學習速率排程器為 『step』 時的步驟大小值。 必須是正整數。 整數 (int)
trainingBatchSize 定型批次大小。 必須是正整數。 整數 (int)
trainingCropSize 影像裁剪大小,這是訓練數據集之類神經網路的輸入。 必須是正整數。 整數 (int)
validationBatchSize 驗證批次大小。 必須是正整數。 整數 (int)
validationCropSize 影像裁剪大小,這是用於驗證數據集之類神經網路的輸入。 必須是正整數。 整數 (int)
validationResizeSize 在裁剪驗證數據集之前要重設大小的影像大小。 必須是正整數。 整數 (int)
warmupCosineLRCycles 學習速率排程器為「warmup_cosine」時,餘弦迴圈的值。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
warmupCosineLRWarmupEpochs 學習速率排程器為「warmup_cosine」時,熱身 Epoch 的值。 必須是正整數。 整數 (int)
weightDecay (權重衰減) 優化器為 'zn'、'adam' 或 'adamw' 時的權數衰變值。 必須是範圍[0, 1] 中的浮點數。 整數 (int)
weightedLoss 加權減肥。 接受的值是 0,沒有加權損失。
1 用於與 sqrt 減肥。(class_weights)。 2 用於減肥與class_weights。 必須是 0 或 1 或 2。
整數 (int)

ImageModelSettingsObjectDetection

名稱 說明 價值觀
高級設置 進階案例的設定。 字串
amsGradient 當優化器為 『adam』 或 'adamw' 時啟用AMSGrad。 布爾 (bool)
擴增 使用擴增的設定。 字串
貝塔1 優化器為 'adam' 或 'adamw' 時的 'beta1' 值。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
貝塔2 優化器為 'adam' 或 'adamw' 時的 'beta2' 值。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
boxDetectionsPerImage (每影像檢測) 所有類別的每個影像偵測數目上限。 必須是正整數。
注意:'yolov5' 演算法不支援此設定。
整數 (int)
boxScoreThreshold (盒分數閾值) 在推斷期間,只傳回大於分類分數的建議
BoxScoreThreshold。 必須是範圍[0, 1] 中的浮點數。
整數 (int)
檢查點頻率 儲存模型檢查點的頻率。 必須是正整數。 整數 (int)
checkpoint模型 累加訓練的預先定型檢查點模型。 MLFlowModelJobInput
檢查點RunId 先前執行的標識碼,其具有累加式定型的預先定型檢查點。 字串
分散式 是否要使用分散式定型。 布爾 (bool)
earlyStopping (早期停止) 在定型期間啟用早期停止邏輯。 布爾 (bool)
earlyStoppingDelay 在主要計量改進之前等待的最小 Epoch 或驗證評估數目
會追蹤早期停止。 必須是正整數。
整數 (int)
早點停止耐心 在之前沒有主要計量改進的最小 Epoch 或驗證評估數目
執行已停止。 必須是正整數。
整數 (int)
enableOnnxNormalization 匯出 ONNX 模型時啟用正規化。 布爾 (bool)
evaluationFrequency 評估驗證數據集以取得計量分數的頻率。 必須是正整數。 整數 (int)
梯度累積步驟 漸層累積表示執行未設定數目的 「GradAccumulationStep」 步驟
更新模型權數,同時累積這些步驟的漸層,然後使用
用來計算權數更新的累積漸層。 必須是正整數。
整數 (int)
影像大小 定型和驗證的影像大小。 必須是正整數。
注意:如果大小太大,定型回合可能會進入 CUDA OOM。
注意:只有 'yolov5' 演算法才支援此設定。
整數 (int)
layersToFreeze (圖層凍結) 要凍結模型的圖層數目。 必須是正整數。
例如,傳遞 2 作為 'seresnext' 值表示
凍結 layer0 和 layer1。 如需支援模型的完整清單,以及圖層凍結的詳細數據,請
請參閱:/azure/machine-learning/how-to-auto-train-image-models。
整數 (int)
learningRate 初始學習速率。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
learningRateScheduler 的 學習率排程器的類型。 必須是 『warmup_cosine』 或 'step'。 “無”
“步驟”
'WarmupCosine'
最大尺寸 在將影像饋送至骨幹之前,要重新調整的影像大小上限。
必須是正整數。 注意:如果大小太大,定型執行可能會進入 CUDA OOM。
注意:'yolov5' 演算法不支援此設定。
整數 (int)
最小大小 將影像饋送至骨幹之前,要重新調整的影像大小下限。
必須是正整數。 注意:如果大小太大,定型執行可能會進入 CUDA OOM。
注意:'yolov5' 演算法不支援此設定。
整數 (int)
型號名稱 要用於定型的模型名稱。
如需可用模型的詳細資訊,請流覽官方檔:
/azure/machine-learning/how-to-auto-train-image-models。
字串
modelSize 模型大小。 必須是 'small'、'medium'、'large' 或 'xlarge'。
注意:如果模型大小太大,定型執行可能會進入 CUDA OOM。
注意:只有 'yolov5' 演算法才支援此設定。
“超大號”
“大”
'中等'
“無”
“小”
動量 優化器為 『zn』 時的動量值。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
多尺度 依 +/- 50%來啟用多重縮放影像大小。
注意:如果沒有足夠的 GPU 記憶體,定型執行可能會進入 CUDA OOM。
注意:只有 'yolov5' 演算法才支援此設定。
布爾 (bool)
涅斯特羅夫 在優化器為 『zn』 時啟用 nesterov。 布爾 (bool)
nmsIouThreshold 在NMS後置處理中的推斷期間使用的IOU臨界值。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
紀元數 定型 Epoch 的數目。 必須是正整數。 整數 (int)
工人數 數據載入器背景工作角色的數目。 必須是非負整數。 整數 (int)
優化器 優化工具的類型。 “亞當”
“亞當”
“無”
'新元'
隨機種子 使用決定性定型時要使用的隨機種子。 整數 (int)
stepLR伽瑪 學習速率排程器為 『step』 時的 gamma 值。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
stepLRStepSize 學習速率排程器為 『step』 時的步驟大小值。 必須是正整數。 整數 (int)
tileGridSize (平鋪網格大小) 用來貼上每個影像的網格線大小。 注意:TileGridSize 不得為
無可啟用小型物件偵測邏輯。 包含兩個整數的字串,其格式為一個。
注意:'yolov5' 演算法不支援此設定。
字串
tileOverlapRatio (平鋪重疊比率) 每個維度中相鄰磚之間的重疊比例。 必須是範圍 [0, 1] 中的浮點數。
注意:'yolov5' 演算法不支援此設定。
整數 (int)
tilePredictionsNmsThreshold 用來執行NMS的IOU閾值,同時合併磚和影像的預測。
用於驗證/推斷。 必須是範圍 [0, 1] 中的 float。
注意:'yolov5' 演算法不支援此設定。
整數 (int)
trainingBatchSize 定型批次大小。 必須是正整數。 整數 (int)
validationBatchSize 驗證批次大小。 必須是正整數。 整數 (int)
validationIouThreshold 計算驗證計量時要使用的IOU閾值。 必須是範圍 [0, 1] 中的 float。 整數 (int)
validationMetricType 用於驗證計量的計量計算方法。 《尋夢環遊記》
“可哥沃克”
“無”
'Voc'
warmupCosineLRCycles 學習速率排程器為「warmup_cosine」時,餘弦迴圈的值。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
warmupCosineLRWarmupEpochs 學習速率排程器為「warmup_cosine」時,熱身 Epoch 的值。 必須是正整數。 整數 (int)
weightDecay (權重衰減) 優化器為 'zn'、'adam' 或 'adamw' 時的權數衰變值。 必須是範圍[0, 1] 中的浮點數。 整數 (int)

ImageObjectDetection

名稱 說明 價值觀
limitSettings [必要]限制 AutoML 作業的設定。 ImageLimitSettings (必要)
modelSettings 用於定型模型的設定。 ImageModelSettingsObjectDetection
primaryMetric 要針對這項工作進行優化的主要計量。 'MeanAveragePrecision'
搜索空間 搜尋空間以取樣模型及其超參數的不同組合。 ImageModelDistributionSettingsObjectDetection[]
掃描設置 模型掃掠和超參數掃掠相關設定。 ImageSweepSettings (圖像掃描設定)
任務類型 [必要]AutoMLJob 的工作類型。 'ImageObjectDetection' (必要)
validationData (驗證資料) 驗證數據輸入。 MLTableJobInput
validationDataSize 需要為驗證目的預留的定型數據集分數。
介於 (0.0 、 1.0) 之間的值
未提供驗證數據集時套用。
整數 (int)

ImageSweepSettings (圖像掃描設定)

名稱 說明 價值觀
early終止 早期終止原則的類型。 提前終止政策
sampling演算法 [必要]超參數取樣演算法的類型。 “貝氏”
'網格'
'Random' (必要)

JobBaseProperties

名稱 說明 價值觀
元件ID 元件資源的 ARM 資源識別碼。 字串
計算ID 計算資源的 ARM 資源識別碼。 字串
說明 資產描述文字。 字串
顯示名稱 工作的顯示名稱。 字串
實驗名稱 作業所屬的實驗名稱。 如果未設定,作業會放在 「默認」實驗中。 字串
身分識別 身分識別設定。 如果設定,這應該是其中一個 AmlToken、ManagedIdentity、UserIdentity 或 null。
如果為 null,則預設為 AmlToken。
身份配置
isArchived 已存檔 資產是否已封存? 布爾 (bool)
職位類型 設定為 AutoMLJob 類型的 『AutoML』。 設定為 CommandJob 類型的 『Command』。 設定為 FineTuningJob 類型的 『FineTuning』。 設定為 PipelineJob 類型的 'Pipeline'。 針對 SparkJob 類型設定為 『Spark』。 設定為 掃掠Job 類型的 [掃掠]。 'AutoML'
“命令”
“微調”
“管道”
“火花”
'掃掠' (必要)
notification設置 作業的通知設定 通知設置
屬性 資產屬性字典。 ResourceBaseProperties
服務 JobEndpoints 的清單。
針對本機作業,作業端點會有 FileStreamObject 的端點值。
JobBase服務
標記 標記字典。 標記可以新增、移除和更新。 ResourceBaseTags

JobBase服務

名稱 說明 價值觀

作業輸入

名稱 說明 價值觀
說明 輸入的描述。 字串
jobInputType (作業輸入類型) 設定為 CustomModelJobInput 類型的 『custom_model』。 設定為 常值JobInput 類型的 『literal』。 設定為 MLFlowModelJobInput 類型的 『mlflow_model』。 設定為 MLTableJobInput 類型的 'mltable'。 設定為 TritonModelJobInput 類型的 『triton_model』。 針對 UriFileJobInput類型設定為 『uri_file』。 針對 UriFolderJobInput 類型設定為 『uri_folder』。 “custom_model”
'字面值'
“mlflow_model”
'mltable'
“triton_model”
“uri_file”
'uri_folder' (必要)

JobOutput (作業輸出)

名稱 說明 價值觀
說明 輸出的描述。 字串
jobOutputType 設定為 CustomModelJobOutput 類型的 『custom_model』。 設定為 MLFlowModelJobOutput 類型的 『mlflow_model』。 設定為 MLTableJobOutput 類型的 'mltable'。 將 設定為 TritonModelJobOutput 類型的 『triton_model』。 針對 UriFileJobOutput類型設定為 『uri_file』。 針對 UriFolderJobOutput類型設定為 『uri_folder』。 “custom_model”
“mlflow_model”
'mltable'
“triton_model”
“uri_file”
'uri_folder' (必要)

JobResourceConfiguration

名稱 說明 價值觀
dockerArgs 傳遞至 Docker run 命令的額外自變數。 這會覆寫系統或本節中已設定的任何參數。 此參數僅支援 Azure ML 計算類型。 字串
實例數量 計算目標所使用的實例或節點選擇性數目。 整數 (int)
instanceType 計算目標所支持的選擇性 VM 類型。 字串
屬性 其他屬性包。 ResourceConfigurationProperties
shmSize Docker 容器的共用記憶體區塊大小。 這應該是 (number)(unit) 的格式,其中數位大於 0,單位可以是 b(位元組)、k(KB)、m(MB)、或 g(GB)。 字串

約束:
圖案 = \d+[bBkKmMgG]

工作資源

名稱 說明 價值觀
instanceTypes 要從中選擇的實例類型清單。 字串[]

工作服務

名稱 說明 價值觀
端點 端點的 URL。 字串
jobServiceType 端點類型。 字串
節點 使用者想要啟動服務的節點。
如果未將 Nodes 設定或設定為 null,則服務只會在領導者節點上啟動。
節點
傳輸埠 端點的埠。 整數 (int)
屬性 在端點上設定的其他屬性。 JobServiceProperties

JobServiceProperties

名稱 說明 價值觀

LiteralJobInput (文字作業輸入)

名稱 說明 價值觀
jobInputType (作業輸入類型) [必要]指定作業的類型。 'literal' (必要)
價值 [必要]輸入的常值。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)

ManagedIdentity (託管身份)

名稱 說明 價值觀
用戶端ID 依用戶端識別元指定使用者指派的身分識別。 若為系統指派,請勿設定此欄位。 字串

約束:
最小長度 = 36
最大長度 = 36
圖案 = ^[0-9a-fA-F]{8}-([0-9a-fA-F]{4}-){3}[0-9a-fA-F]{12}$
身份類型 [必要]指定身分識別架構的類型。 'Managed' (必要)
物件ID 依物件識別元指定使用者指派的身分識別。 若為系統指派,請勿設定此欄位。 字串

約束:
最小長度 = 36
最大長度 = 36
圖案 = ^[0-9a-fA-F]{8}-([0-9a-fA-F]{4}-){3}[0-9a-fA-F]{12}$
資源ID 依 ARM 資源識別碼指定使用者指派的身分識別。 若為系統指派,請勿設定此欄位。 字串

中位數停止政策

名稱 說明 價值觀
policyType [必要]原則設定的名稱 'MedianStopping' (必要)

MLFlowModelJobInput

名稱 說明 價值觀
說明 輸入的描述。 字串
jobInputType (作業輸入類型) [必要]指定作業的類型。 “custom_model”
'字面值'
“mlflow_model”
'mltable'
“triton_model”
“uri_file”
'uri_folder' (必要)
模式 輸入資產傳遞模式。 “直接”
'下載'
'EvalDownload'
'EvalMount'
'ReadOnlyMount'
“ReadWriteMount”
統一資源識別碼 (URI) [必要]輸入資產 URI。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)

MLFlowModelJobInput

名稱 說明 價值觀
jobInputType (作業輸入類型) [必要]指定作業的類型。 'mlflow_model' (必要)
模式 輸入資產傳遞模式。 “直接”
'下載'
'EvalDownload'
'EvalMount'
'ReadOnlyMount'
“ReadWriteMount”
統一資源識別碼 (URI) [必要]輸入資產 URI。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)

MLFlowModelJobOutput

名稱 說明 價值觀
jobOutputType [必要]指定作業的類型。 'mlflow_model' (必要)
模式 輸出資產傳遞模式。 “直接”
“ReadWriteMount”
'上傳'
統一資源識別碼 (URI) 輸出資產 URI。 字串

MLTableJobInput

名稱 說明 價值觀
說明 輸入的描述。 字串
jobInputType (作業輸入類型) [必要]指定作業的類型。 “custom_model”
'字面值'
“mlflow_model”
'mltable'
“triton_model”
“uri_file”
'uri_folder' (必要)
模式 輸入資產傳遞模式。 “直接”
'下載'
'EvalDownload'
'EvalMount'
'ReadOnlyMount'
“ReadWriteMount”
統一資源識別碼 (URI) [必要]輸入資產 URI。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)

MLTableJobInput

名稱 說明 價值觀
jobInputType (作業輸入類型) [必要]指定作業的類型。 'mltable' (必要)
模式 輸入資產傳遞模式。 “直接”
'下載'
'EvalDownload'
'EvalMount'
'ReadOnlyMount'
“ReadWriteMount”
統一資源識別碼 (URI) [必要]輸入資產 URI。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)

MLTableJobOutput

名稱 說明 價值觀
jobOutputType [必要]指定作業的類型。 'mltable' (必要)
模式 輸出資產傳遞模式。 “直接”
“ReadWriteMount”
'上傳'
統一資源識別碼 (URI) 輸出資產 URI。 字串

MPI

名稱 說明 價值觀
distributionType [必要]指定散發架構的類型。 'Mpi' (必要)
processCountPerInstance 每個 MPI 節點的進程數目。 整數 (int)

NCross驗證

名稱 說明 價值觀
模式 針對 AutoNCrossValidations 類型,設定為 『Auto』。 設定為 CustomNCrossValidations 類型的 [自定義]。 '自動'
'Custom' (必要)

NlpVerticalFeaturizationSettings

名稱 說明 價值觀
datasetLanguage 數據集語言,適用於文字數據。 字串

NlpVerticalLimitSettings

名稱 說明 價值觀
maxConcurrentTrials 並行 AutoML 反覆專案上限。 整數 (int)
maxTrials AutoML 反覆項目的數目。 整數 (int)
超時 AutoML 作業逾時。 字串

節點

名稱 說明 價值觀
nodesValueType 針對 AllNodes 類型,設定為 'All'。 'All' (必要)

通知設置

名稱 說明 價值觀
電子郵件 在指定的通知類型上傳送電子郵件通知給使用者 包含任何的字串數組:
'JobCancelled' (工作已取消)
'JobCompleted'
“作業失敗”
電子郵件 這是電子郵件收件者清單,其總限製為499個字元,且具有逗號分隔符 字串[]
webhook 將 Webhook 回呼傳送至服務。 機碼是 Webhook 的使用者提供名稱。 NotificationSettingWebhook

NotificationSettingWebhook

名稱 說明 價值觀

目的

名稱 說明 價值觀
目標 [必要]定義超參數微調支援的計量目標 “最大化”
'最小化' (必要)
primaryMetric [必要]要優化之計量的名稱。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)

管道作業

名稱 說明 價值觀
輸入 管線作業的輸入。 管道作業輸入
工作 作業會建構管線作業。 PipelineJobJob 作業
職位類型 [必要]指定作業的類型。 'Pipeline' (必要)
輸出 管線作業的輸出 PipelineJobOutputs
設定 管線設定,適用於 ContinueRunOnStepFailure 等專案。 任何
源作業 ID 來源作業的 ARM 資源識別碼。 字串

管道作業輸入

名稱 說明 價值觀

PipelineJobJob 作業

名稱 說明 價值觀

PipelineJobOutputs

名稱 說明 價值觀

PyTorch

名稱 說明 價值觀
distributionType [必要]指定散發架構的類型。 'PyTorch' (必要)
processCountPerInstance 每個節點的進程數目。 整數 (int)

佇列設置

名稱 說明 價值觀
jobTier 控制計算作業層 “基本”
'Null'
“高級”
'Spot'
“標準”

RandomSamplingAlgorithm 演算法

名稱 說明 價值觀
規則 隨機演算法的特定類型 '隨機'
“索博爾”
採樣演算法類型 [必要]用來產生超參數值的演算法,以及組態屬性 'Random' (必要)
種子 要作為隨機數產生種子的選擇性整數 整數 (int)

迴歸

名稱 說明 價值觀
cvSplitColumnNames 要用於 CVSplit 資料的數據行。 字串[]
featurizationSettings AutoML 作業所需的特徵化輸入。 TableVerticalFeaturizationSettings
limitSettings AutoMLJob 的執行條件約束。 TableVerticalLimitSettings
nCrossValidations 要套用至定型數據集的交叉驗證折疊數目
未提供驗證數據集時。
NCross驗證
primaryMetric 回歸工作的主要計量。 'NormalizedMeanAbsoluteError'
'NormalizedRootMeanSquaredError'
'R2Score'
'SpearmanCorrelation'
任務類型 [必要]AutoMLJob 的工作類型。 '回歸' (必要)
測試數據 測試數據輸入。 MLTableJobInput
testDataSize 必須為驗證目的而保留的測試數據集分數。
介於 (0.0 、 1.0) 之間的值
未提供驗證數據集時套用。
整數 (int)
trainingSettings AutoML 作業定型階段的輸入。 回歸訓練設置
validationData (驗證資料) 驗證數據輸入。 MLTableJobInput
validationDataSize 需要為驗證目的預留的定型數據集分數。
介於 (0.0 、 1.0) 之間的值
未提供驗證數據集時套用。
整數 (int)
權重列名稱 範例權數數據行的名稱。 自動化 ML 支援加權數據行做為輸入,導致數據中的數據列向上或向下加權。 字串

回歸訓練設置

名稱 說明 價值觀
allowedTrainingAlgorithms 回歸工作的允許模型。 包含任何的字串數組:
“決策樹”
“彈性網”
'ExtremeRandomTrees'
'梯度提升'
'KNN'
“套索拉斯”
'LightGBM'
'隨機森林'
'新幣'
'XGBoostRegressor'
blockedTrainingAlgorithms 回歸工作的封鎖模型。 包含任何的字串數組:
“決策樹”
“彈性網”
'ExtremeRandomTrees'
'梯度提升'
'KNN'
“套索拉斯”
'LightGBM'
'隨機森林'
'新幣'
'XGBoostRegressor'
enableDnnTraining 啟用 DNN 模型的建議。 布爾 (bool)
enableModelExplainability 在最佳模型上開啟可解釋性的旗標。 布爾 (bool)
enableOnnxCompatibleModels 啟用 onnx 相容模型的旗標。 布爾 (bool)
enableStackEnsemble 啟用堆疊合奏執行。 布爾 (bool)
enableVoteEnsemble 啟用投票合奏執行。 布爾 (bool)
ensembleModelDownloadTimeout 在 VotingEnsemble 和 StackEnsemble 模型產生期間,會下載先前子回合中的多個安裝模型。
如果需要更多時間,請使用高於 300 秒的值來設定此參數。
字串
stackEnsembleSettings 堆疊合奏執行的堆疊合奏設定。 StackEnsembleSettings

ResourceBaseProperties

名稱 說明 價值觀

ResourceBaseTags

名稱 說明 價值觀

ResourceConfigurationProperties

名稱 說明 價值觀

採樣演算法

名稱 說明 價值觀
採樣演算法類型 針對 BayesianSamplingAlgorithm類型設定為 『Bayesian』。 設定為 GridSamplingAlgorithm 類型的 『Grid』。 設定為 RandomSamplingAlgorithm 類型的 『Random』。 “貝氏”
'網格'
'Random' (必要)

季節性

名稱 說明 價值觀
模式 針對 AutoSeasonality 類型,設定為 [自動]。 針對 CustomSeasonality 類型,設定為 [自定義]。 '自動'
'Custom' (必要)

SparkJob

名稱 說明 價值觀
檔案 封存作業中使用的檔案。 字串[]
參數 作業的自變數。 字串
代碼Id [必要] 程式代碼資產的arm識別碼。 字串 (必要)
會議 Spark 設定的屬性。 SparkJobConf 函數
輸入 [必要]在作業啟動時要執行的專案。 SparkJobEntry (必要)
環境ID 作業之環境規格的 ARM 資源識別碼。 字串
環境變數 作業中包含的環境變數。 SparkJobEnvironment變數
檔案 作業中使用的檔案。 字串[]
輸入 對應作業中使用的輸入數據系結。 SparkJobInputs
罐子 作業中使用的 Jar 檔案。 字串[]
職位類型 [必要]指定作業的類型。 'Spark' (必要)
輸出 對應作業中使用的輸出數據系結。 SparkJob輸出
py檔 作業中使用的 Python 檔案。 字串[]
queueSettings 作業的佇列設定 佇列設置
資源 作業的計算資源組態。 SparkResourceConfiguration

SparkJobConf 函數

名稱 說明 價值觀

SparkJobEntry

名稱 說明 價值觀
sparkJobEntryType 針對 SparkJobPythonEntry 類型,設定為 'SparkJobPythonEntry'。 針對 SparkJobScalaEntry 類型,設定為 'SparkJobScalaEntry'。 “SparkJobPythonEntry”
'SparkJobScalaEntry' (必要)

SparkJobEnvironment變數

名稱 說明 價值觀

SparkJobInputs

名稱 說明 價值觀

SparkJob輸出

名稱 說明 價值觀

SparkJobPython條目

名稱 說明 價值觀
檔案 [必要]作業進入點的相對 Python 檔案路徑。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)
sparkJobEntryType [必要]作業進入點的類型。 'SparkJobPythonEntry' (必要)

SparkJobScalaEntry

名稱 說明 價值觀
類別名稱 [必要]做為進入點的 Scala 類別名稱。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)
sparkJobEntryType [必要]作業進入點的類型。 'SparkJobScalaEntry' (必要)

SparkResourceConfiguration

名稱 說明 價值觀
instanceType 計算目標所支持的選擇性 VM 類型。 字串
runtimeVersion (運行時版本) 用於作業的 Spark 執行時間版本。 字串

StackEnsembleSettings

名稱 說明 價值觀
stackMetaLearnerKWargs 要傳遞至中繼學習模組初始化表達式的選擇性參數。 任何
stackMetaLearnerTrainPercentage 指定要保留用於訓練中繼學習工具的定型集比例(選擇定型和定型類型的訓練類型)。 預設值為 0.2。 整數 (int)
stackMetaLearnerType 中繼學習器是針對個別異質模型輸出定型的模型。 “彈性網”
'ElasticNetCV'
'LightGBMClassifier'
'LightGBMRegressor'
'線性回歸'
'LogisticRegression' (邏輯回歸)
'LogisticRegressionCV'
“無”

SweepJob (掃描作業)

名稱 說明 價值觀
early終止 早期終止原則可在完成之前取消執行效能不佳 提前終止政策
輸入 對應作業中使用的輸入數據系結。 SweepJobInputs
職位類型 [必要]指定作業的類型。 '掃掠' (必要)
限制 掃掠作業限制。 SweepJobLimits (掃描作業限制)
目的 [必要]優化目標。 目標 (必要)
輸出 對應作業中使用的輸出數據系結。 SweepJobOutputs (掃描作業輸出)
queueSettings 作業的佇列設定 佇列設置
sampling演算法 [必要]超參數取樣演算法 SamplingAlgorithm (必要)
搜索空間 [必要]包含每個參數及其散發的字典。 字典索引鍵是參數的名稱 任何 (必要)
試用 [必要]試用版元件定義。 TrialComponent (必要)

SweepJobInputs

名稱 說明 價值觀

SweepJobLimits (掃描作業限制)

名稱 說明 價值觀
jobLimits類型 [必要]JobLimit 類型。 “命令”
'掃掠' (必要)
maxConcurrentTrials 掃掠作業最大並行試用版。 整數 (int)
maxTotalTrials 掃掠作業最大總試用版。 整數 (int)
超時 ISO 8601 格式的最大執行持續時間,之後作業將會取消。 僅支援有效位數為秒數的持續時間。 字串
trialTimeout (試用超時) 掃掠作業試用版逾時值。 字串

SweepJobOutputs (掃描作業輸出)

名稱 說明 價值觀

TableVerticalFeaturizationSettings

名稱 說明 價值觀
阻塞變壓器 這些轉換器不得用於特徵化。 包含任何的字串數組:
'CatTargetEncoder' (貓目標編碼器)
'CountVectorizer'
'HashOneHotEncoder'
'標籤編碼器'
“天真貝葉斯”
'OneHotEncoder' (一熱編碼器)
'TextTargetEncoder'
'TfIdf'
'WoETargetEncoder' (WoETarget編碼器)
'WordEmbedding' (單詞嵌入)
columnNameAnd類型 數據行名稱及其類型字典(int、float、string、datetime 等)。 TableVerticalFeaturizationSettingsColumnNameAndTypes
datasetLanguage 數據集語言,適用於文字數據。 字串
enableDnnFeaturization 判斷是否要使用 Dnn 型特徵化程式進行數據特徵化。 布爾 (bool)
模式 特徵化模式 - 使用者可以保留預設的「自動」模式,而 AutoML 會負責在特徵化階段中轉換數據的必要轉換。
如果選取 [關閉],則不會完成特徵化。
如果選取 [自定義],則使用者可以指定其他輸入,以自定義特徵化的完成方式。
'自動'
“自定義”
“關閉”
transformer參數 用戶可以指定要搭配套用的其他轉換器,以及要套用它的數據行,以及轉換建構函式的參數。 TableVerticalFeaturizationSettingsTransformerParams

TableVerticalFeaturizationSettingsColumnNameAndTypes

名稱 說明 價值觀

TableVerticalFeaturizationSettingsTransformerParams

名稱 說明 價值觀

TableVerticalLimitSettings

名稱 說明 價值觀
enableEarlyTermination 啟用提早終止,判斷如果過去 20 個反覆項目沒有分數改善,AutoMLJob 是否會提早終止。 布爾 (bool)
退出評分 AutoML 作業的結束分數。 整數 (int)
maxConcurrentTrials 並行反覆運算數上限。 整數 (int)
maxCoresPerTrial 每個反覆專案的核心數上限。 整數 (int)
maxTrials 反覆項目的數目。 整數 (int)
超時 AutoML 作業逾時。 字串
trialTimeout (試用超時) 反覆專案逾時。 字串

目標滯後

名稱 說明 價值觀
模式 針對 AutoTargetLags 類型設定為 『Auto』。 設定為 CustomTargetLags 類型的 [自定義]。 '自動'
'Custom' (必要)

TargetRollingWindowSize

名稱 說明 價值觀
模式 針對 AutoTargetRollingWindowSize 類型,設定為 『Auto』。 針對 CustomTargetRollingWindowSize 類型設定為 『Custom』。 '自動'
'Custom' (必要)

TensorFlow

名稱 說明 價值觀
distributionType [必要]指定散發架構的類型。 'TensorFlow' (必要)
parameterServerCount 參數伺服器工作的數目。 整數 (int)
workerCount 工人計數 工人人數。 如果未指定,則會預設為實例計數。 整數 (int)

文本分類

名稱 說明 價值觀
featurizationSettings AutoML 作業所需的特徵化輸入。 NlpVerticalFeaturizationSettings
limitSettings AutoMLJob 的執行條件約束。 NlpVerticalLimitSettings
primaryMetric Text-Classification 工作的主要計量。 '精確度'
'AUCWeighted'
'AveragePrecisionScoreWeighted'
'NormMacroRecall' (規範巨集調用)
'PrecisionScoreWeighted'
任務類型 [必要]AutoMLJob 的工作類型。 'TextClassification' (必要)
validationData (驗證資料) 驗證數據輸入。 MLTableJobInput

TextClassificationMultilabel

名稱 說明 價值觀
featurizationSettings AutoML 作業所需的特徵化輸入。 NlpVerticalFeaturizationSettings
limitSettings AutoMLJob 的執行條件約束。 NlpVerticalLimitSettings
任務類型 [必要]AutoMLJob 的工作類型。 'TextClassificationMultilabel' (必要)
validationData (驗證資料) 驗證數據輸入。 MLTableJobInput

TextNer (文字產生器)

名稱 說明 價值觀
featurizationSettings AutoML 作業所需的特徵化輸入。 NlpVerticalFeaturizationSettings
limitSettings AutoMLJob 的執行條件約束。 NlpVerticalLimitSettings
任務類型 [必要]AutoMLJob 的工作類型。 'TextNER' (必要)
validationData (驗證資料) 驗證數據輸入。 MLTableJobInput

TrialComponent (試用元件)

名稱 說明 價值觀
代碼Id 程式代碼資產的 ARM 資源識別碼。 字串
指令 [必要]在作業啟動時執行的命令。 例如 “python train.py” 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)
分配 作業的散發組態。 如果設定,這應該是其中一個 Mpi、Tensorflow、PyTorch 或 null。 分發配置
環境ID [必要]作業之環境規格的 ARM 資源識別碼。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)
環境變數 作業中包含的環境變數。 TrialComponentEnvironmentVariables
資源 作業的計算資源組態。 JobResourceConfiguration

TrialComponentEnvironmentVariables

名稱 說明 價值觀

TritonModelJobInput

名稱 說明 價值觀
jobInputType (作業輸入類型) [必要]指定作業的類型。 'triton_model' (必要)
模式 輸入資產傳遞模式。 “直接”
'下載'
'EvalDownload'
'EvalMount'
'ReadOnlyMount'
“ReadWriteMount”
統一資源識別碼 (URI) [必要]輸入資產 URI。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)

TritonModelJobOutput

名稱 說明 價值觀
jobOutputType [必要]指定作業的類型。 'triton_model' (必要)
模式 輸出資產傳遞模式。 “直接”
“ReadWriteMount”
'上傳'
統一資源識別碼 (URI) 輸出資產 URI。 字串

截斷選擇策略 (TruncationSelectionPolicy)

名稱 說明 價值觀
policyType [必要]原則設定的名稱 'TruncationSelection' (必要)
截斷百分比 要在每個評估間隔取消的執行百分比。 整數 (int)

UriFileJobInput

名稱 說明 價值觀
jobInputType (作業輸入類型) [必要]指定作業的類型。 'uri_file' (必要)
模式 輸入資產傳遞模式。 “直接”
'下載'
'EvalDownload'
'EvalMount'
'ReadOnlyMount'
“ReadWriteMount”
統一資源識別碼 (URI) [必要]輸入資產 URI。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)

UriFileJobOutput

名稱 說明 價值觀
jobOutputType [必要]指定作業的類型。 'uri_file' (必要)
模式 輸出資產傳遞模式。 “直接”
“ReadWriteMount”
'上傳'
統一資源識別碼 (URI) 輸出資產 URI。 字串

UriFolderJobInput

名稱 說明 價值觀
jobInputType (作業輸入類型) [必要]指定作業的類型。 'uri_folder' (必要)
模式 輸入資產傳遞模式。 “直接”
'下載'
'EvalDownload'
'EvalMount'
'ReadOnlyMount'
“ReadWriteMount”
統一資源識別碼 (URI) [必要]輸入資產 URI。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)

UriFolderJobOutput

名稱 說明 價值觀
jobOutputType [必要]指定作業的類型。 'uri_folder' (必要)
模式 輸出資產傳遞模式。 “直接”
“ReadWriteMount”
'上傳'
統一資源識別碼 (URI) 輸出資產 URI。 字串

使用者身份

名稱 說明 價值觀
身份類型 [必要]指定身分識別架構的類型。 'UserIdentity' (必要)

Webhook

名稱 說明 價值觀
事件類型 在指定的通知事件上傳送回呼 字串
webhook類型 針對 AzureDevOpsWebhook 類型,設定為 『AzureDevOps』。 'AzureDevOps' (必要)

使用範例

Azure 快速入門範例

下列 Azure 快速入門範本 包含用於部署此資源類型的 Bicep 範例。

Bicep 檔案 說明
建立 Azure Machine Learning AutoML 分類作業 此範本會建立 Azure Machine Learning AutoML 分類作業,以找出最佳模型,以預測客戶是否會訂閱與金融機構的固定期存款。
建立 Azure Machine Learning 命令作業 此範本會建立具有基本hello_world腳本的 Azure Machine Learning 命令作業
建立 Azure Machine Learning 掃掠作業 此範本會建立 Azure Machine Learning 掃掠作業以進行超參數微調。

ARM 樣本資源定義

工作區/作業資源類型可以使用目標作業來部署:

如需每個 API 版本中已變更屬性的清單,請參閱 變更記錄檔。

資源格式

若要建立 Microsoft.MachineLearningServices/workspaces/jobs 資源,請將下列 JSON 新增至範本。

{
  "type": "Microsoft.MachineLearningServices/workspaces/jobs",
  "apiVersion": "2024-07-01-preview",
  "name": "string",
  "properties": {
    "componentId": "string",
    "computeId": "string",
    "description": "string",
    "displayName": "string",
    "experimentName": "string",
    "identity": {
      "identityType": "string"
      // For remaining properties, see IdentityConfiguration objects
    },
    "isArchived": "bool",
    "notificationSetting": {
      "emailOn": [ "string" ],
      "emails": [ "string" ],
      "webhooks": {
        "{customized property}": {
          "eventType": "string",
          "webhookType": "string"
          // For remaining properties, see Webhook objects
        }
      }
    },
    "properties": {
      "{customized property}": "string"
    },
    "services": {
      "{customized property}": {
        "endpoint": "string",
        "jobServiceType": "string",
        "nodes": {
          "nodesValueType": "string"
          // For remaining properties, see Nodes objects
        },
        "port": "int",
        "properties": {
          "{customized property}": "string"
        }
      }
    },
    "tags": {
      "{customized property}": "string"
    },
    "jobType": "string"
    // For remaining properties, see JobBaseProperties objects
  }
}

FineTuningVertical 物件

設定 modelProvider 屬性以指定物件的類型。

針對 AzureOpenAI,請使用:

{
  "hyperParameters": {
    "batchSize": "int",
    "learningRateMultiplier": "int",
    "nEpochs": "int"
  },
  "modelProvider": "AzureOpenAI"
}

針對 [自定義],請使用:

{
  "hyperParameters": {
    "{customized property}": "string"
  },
  "modelProvider": "Custom"
}

TargetRollingWindowSize 物件

設定 mode 屬性以指定物件的類型。

針對 [自動],請使用:

{
  "mode": "Auto"
}

針對 [自定義],請使用:

{
  "mode": "Custom",
  "value": "int"
}

JobInput 物件

設定 jobInputType 屬性以指定物件的類型。

針對 custom_model,請使用:

{
  "jobInputType": "custom_model",
  "mode": "string",
  "uri": "string"
}

針對 常值,請使用:

{
  "jobInputType": "literal",
  "value": "string"
}

針對 mlflow_model,請使用:

{
  "jobInputType": "mlflow_model",
  "mode": "string",
  "uri": "string"
}

針對 mltable,請使用:

{
  "jobInputType": "mltable",
  "mode": "string",
  "uri": "string"
}

針對 triton_model,請使用:

{
  "jobInputType": "triton_model",
  "mode": "string",
  "uri": "string"
}

針對 uri_file,請使用:

{
  "jobInputType": "uri_file",
  "mode": "string",
  "uri": "string"
}

針對 uri_folder,請使用:

{
  "jobInputType": "uri_folder",
  "mode": "string",
  "uri": "string"
}

JobBaseProperties 物件

設定 jobType 屬性以指定物件的類型。

針對 AutoML,請使用:

{
  "environmentId": "string",
  "environmentVariables": {
    "{customized property}": "string"
  },
  "jobType": "AutoML",
  "outputs": {
    "{customized property}": {
      "description": "string",
      "jobOutputType": "string"
      // For remaining properties, see JobOutput objects
    }
  },
  "queueSettings": {
    "jobTier": "string"
  },
  "resources": {
    "dockerArgs": "string",
    "instanceCount": "int",
    "instanceType": "string",
    "properties": {
      "{customized property}": {}
    },
    "shmSize": "string"
  },
  "taskDetails": {
    "logVerbosity": "string",
    "targetColumnName": "string",
    "trainingData": {
      "description": "string",
      "jobInputType": "string",
      "mode": "string",
      "uri": "string"
    },
    "taskType": "string"
    // For remaining properties, see AutoMLVertical objects
  }
}

針對 [命令],請使用:

{
  "codeId": "string",
  "command": "string",
  "distribution": {
    "distributionType": "string"
    // For remaining properties, see DistributionConfiguration objects
  },
  "environmentId": "string",
  "environmentVariables": {
    "{customized property}": "string"
  },
  "inputs": {
    "{customized property}": {
      "description": "string",
      "jobInputType": "string"
      // For remaining properties, see JobInput objects
    }
  },
  "jobType": "Command",
  "limits": {
    "jobLimitsType": "string",
    "timeout": "string"
  },
  "outputs": {
    "{customized property}": {
      "description": "string",
      "jobOutputType": "string"
      // For remaining properties, see JobOutput objects
    }
  },
  "queueSettings": {
    "jobTier": "string"
  },
  "resources": {
    "dockerArgs": "string",
    "instanceCount": "int",
    "instanceType": "string",
    "properties": {
      "{customized property}": {}
    },
    "shmSize": "string"
  }
}

針對 FineTuning,請使用:

{
  "fineTuningDetails": {
    "model": {
      "description": "string",
      "jobInputType": "string",
      "mode": "string",
      "uri": "string"
    },
    "taskType": "string",
    "trainingData": {
      "description": "string",
      "jobInputType": "string"
      // For remaining properties, see JobInput objects
    },
    "validationData": {
      "description": "string",
      "jobInputType": "string"
      // For remaining properties, see JobInput objects
    },
    "modelProvider": "string"
    // For remaining properties, see FineTuningVertical objects
  },
  "jobType": "FineTuning",
  "outputs": {
    "{customized property}": {
      "description": "string",
      "jobOutputType": "string"
      // For remaining properties, see JobOutput objects
    }
  },
  "queueSettings": {
    "jobTier": "string"
  },
  "resources": {
    "instanceTypes": [ "string" ]
  }
}

針對 管線,請使用:

{
  "inputs": {
    "{customized property}": {
      "description": "string",
      "jobInputType": "string"
      // For remaining properties, see JobInput objects
    }
  },
  "jobs": {
    "{customized property}": {}
  },
  "jobType": "Pipeline",
  "outputs": {
    "{customized property}": {
      "description": "string",
      "jobOutputType": "string"
      // For remaining properties, see JobOutput objects
    }
  },
  "settings": {},
  "sourceJobId": "string"
}

針對 Spark,請使用:

{
  "archives": [ "string" ],
  "args": "string",
  "codeId": "string",
  "conf": {
    "{customized property}": "string"
  },
  "entry": {
    "sparkJobEntryType": "string"
    // For remaining properties, see SparkJobEntry objects
  },
  "environmentId": "string",
  "environmentVariables": {
    "{customized property}": "string"
  },
  "files": [ "string" ],
  "inputs": {
    "{customized property}": {
      "description": "string",
      "jobInputType": "string"
      // For remaining properties, see JobInput objects
    }
  },
  "jars": [ "string" ],
  "jobType": "Spark",
  "outputs": {
    "{customized property}": {
      "description": "string",
      "jobOutputType": "string"
      // For remaining properties, see JobOutput objects
    }
  },
  "pyFiles": [ "string" ],
  "queueSettings": {
    "jobTier": "string"
  },
  "resources": {
    "instanceType": "string",
    "runtimeVersion": "string"
  }
}

針對 掃掠,請使用:

{
  "earlyTermination": {
    "delayEvaluation": "int",
    "evaluationInterval": "int",
    "policyType": "string"
    // For remaining properties, see EarlyTerminationPolicy objects
  },
  "inputs": {
    "{customized property}": {
      "description": "string",
      "jobInputType": "string"
      // For remaining properties, see JobInput objects
    }
  },
  "jobType": "Sweep",
  "limits": {
    "jobLimitsType": "string",
    "maxConcurrentTrials": "int",
    "maxTotalTrials": "int",
    "timeout": "string",
    "trialTimeout": "string"
  },
  "objective": {
    "goal": "string",
    "primaryMetric": "string"
  },
  "outputs": {
    "{customized property}": {
      "description": "string",
      "jobOutputType": "string"
      // For remaining properties, see JobOutput objects
    }
  },
  "queueSettings": {
    "jobTier": "string"
  },
  "samplingAlgorithm": {
    "samplingAlgorithmType": "string"
    // For remaining properties, see SamplingAlgorithm objects
  },
  "searchSpace": {},
  "trial": {
    "codeId": "string",
    "command": "string",
    "distribution": {
      "distributionType": "string"
      // For remaining properties, see DistributionConfiguration objects
    },
    "environmentId": "string",
    "environmentVariables": {
      "{customized property}": "string"
    },
    "resources": {
      "dockerArgs": "string",
      "instanceCount": "int",
      "instanceType": "string",
      "properties": {
        "{customized property}": {}
      },
      "shmSize": "string"
    }
  }
}

IdentityConfiguration 物件

設定 identityType 屬性以指定物件的類型。

針對 AMLToken,請使用:

{
  "identityType": "AMLToken"
}

針對 [受控],請使用:

{
  "clientId": "string",
  "identityType": "Managed",
  "objectId": "string",
  "resourceId": "string"
}

針對 UserIdentity,請使用:

{
  "identityType": "UserIdentity"
}

ForecastHorizon 物件

設定 mode 屬性以指定物件的類型。

針對 [自動],請使用:

{
  "mode": "Auto"
}

針對 [自定義],請使用:

{
  "mode": "Custom",
  "value": "int"
}

EarlyTerminationPolicy 物件

設定 policyType 屬性以指定物件的類型。

針對 Bandit,請使用:

{
  "policyType": "Bandit",
  "slackAmount": "int",
  "slackFactor": "int"
}

針對 MedianStopping,請使用:

{
  "policyType": "MedianStopping"
}

針對 [截斷][選取],請使用:

{
  "policyType": "TruncationSelection",
  "truncationPercentage": "int"
}

季節性物件

設定 mode 屬性以指定物件的類型。

針對 [自動],請使用:

{
  "mode": "Auto"
}

針對 [自定義],請使用:

{
  "mode": "Custom",
  "value": "int"
}

JobOutput 物件

設定 jobOutputType 屬性以指定對象的類型。

針對 custom_model,請使用:

{
  "jobOutputType": "custom_model",
  "mode": "string",
  "uri": "string"
}

針對 mlflow_model,請使用:

{
  "jobOutputType": "mlflow_model",
  "mode": "string",
  "uri": "string"
}

針對 mltable,請使用:

{
  "jobOutputType": "mltable",
  "mode": "string",
  "uri": "string"
}

針對 triton_model,請使用:

{
  "jobOutputType": "triton_model",
  "mode": "string",
  "uri": "string"
}

針對 uri_file,請使用:

{
  "jobOutputType": "uri_file",
  "mode": "string",
  "uri": "string"
}

針對 uri_folder,請使用:

{
  "jobOutputType": "uri_folder",
  "mode": "string",
  "uri": "string"
}

TargetLags 物件

設定 mode 屬性以指定物件的類型。

針對 [自動],請使用:

{
  "mode": "Auto"
}

針對 [自定義],請使用:

{
  "mode": "Custom",
  "values": [ "int" ]
}

Webhook 物件

設定 webhookType 屬性以指定物件的類型。

針對 AzureDevOps,請使用:

{
  "webhookType": "AzureDevOps"
}

SamplingAlgorithm 物件

設定 samplingAlgorithmType 屬性以指定物件的類型。

針對 貝氏,請使用:

{
  "samplingAlgorithmType": "Bayesian"
}

針對 Grid,請使用:

{
  "samplingAlgorithmType": "Grid"
}

針對 [隨機],請使用:

{
  "rule": "string",
  "samplingAlgorithmType": "Random",
  "seed": "int"
}

SparkJobEntry 物件

設定 sparkJobEntryType 屬性以指定物件的類型。

針對 SparkJobPythonEntry,請使用:

{
  "file": "string",
  "sparkJobEntryType": "SparkJobPythonEntry"
}

針對 SparkJobScalaEntry,請使用:

{
  "className": "string",
  "sparkJobEntryType": "SparkJobScalaEntry"
}

DistributionConfiguration 物件

設定 distributionType 屬性以指定物件的類型。

針對 Mpi,請使用:

{
  "distributionType": "Mpi",
  "processCountPerInstance": "int"
}

針對 PyTorch,請使用:

{
  "distributionType": "PyTorch",
  "processCountPerInstance": "int"
}

針對 TensorFlow,請使用:

{
  "distributionType": "TensorFlow",
  "parameterServerCount": "int",
  "workerCount": "int"
}

NCrossValidations 物件

設定 mode 屬性以指定物件的類型。

針對 [自動],請使用:

{
  "mode": "Auto"
}

針對 [自定義],請使用:

{
  "mode": "Custom",
  "value": "int"
}

Nodes 物件

設定 nodesValueType 屬性以指定對象的類型。

針對 [全部],請使用:

{
  "nodesValueType": "All"
}

AutoMLVertical 物件

設定 taskType 屬性以指定物件的類型。

針對 [分類],請使用:

{
  "cvSplitColumnNames": [ "string" ],
  "featurizationSettings": {
    "blockedTransformers": [ "string" ],
    "columnNameAndTypes": {
      "{customized property}": "string"
    },
    "datasetLanguage": "string",
    "enableDnnFeaturization": "bool",
    "mode": "string",
    "transformerParams": {
      "{customized property}": [
        {
          "fields": [ "string" ],
          "parameters": {}
        }
      ]
    }
  },
  "limitSettings": {
    "enableEarlyTermination": "bool",
    "exitScore": "int",
    "maxConcurrentTrials": "int",
    "maxCoresPerTrial": "int",
    "maxTrials": "int",
    "timeout": "string",
    "trialTimeout": "string"
  },
  "nCrossValidations": {
    "mode": "string"
    // For remaining properties, see NCrossValidations objects
  },
  "positiveLabel": "string",
  "primaryMetric": "string",
  "taskType": "Classification",
  "testData": {
    "description": "string",
    "jobInputType": "string",
    "mode": "string",
    "uri": "string"
  },
  "testDataSize": "int",
  "trainingSettings": {
    "allowedTrainingAlgorithms": [ "string" ],
    "blockedTrainingAlgorithms": [ "string" ],
    "enableDnnTraining": "bool",
    "enableModelExplainability": "bool",
    "enableOnnxCompatibleModels": "bool",
    "enableStackEnsemble": "bool",
    "enableVoteEnsemble": "bool",
    "ensembleModelDownloadTimeout": "string",
    "stackEnsembleSettings": {
      "stackMetaLearnerKWargs": {},
      "stackMetaLearnerTrainPercentage": "int",
      "stackMetaLearnerType": "string"
    }
  },
  "validationData": {
    "description": "string",
    "jobInputType": "string",
    "mode": "string",
    "uri": "string"
  },
  "validationDataSize": "int",
  "weightColumnName": "string"
}

針對 預測,請使用:

{
  "cvSplitColumnNames": [ "string" ],
  "featurizationSettings": {
    "blockedTransformers": [ "string" ],
    "columnNameAndTypes": {
      "{customized property}": "string"
    },
    "datasetLanguage": "string",
    "enableDnnFeaturization": "bool",
    "mode": "string",
    "transformerParams": {
      "{customized property}": [
        {
          "fields": [ "string" ],
          "parameters": {}
        }
      ]
    }
  },
  "forecastingSettings": {
    "countryOrRegionForHolidays": "string",
    "cvStepSize": "int",
    "featureLags": "string",
    "forecastHorizon": {
      "mode": "string"
      // For remaining properties, see ForecastHorizon objects
    },
    "frequency": "string",
    "seasonality": {
      "mode": "string"
      // For remaining properties, see Seasonality objects
    },
    "shortSeriesHandlingConfig": "string",
    "targetAggregateFunction": "string",
    "targetLags": {
      "mode": "string"
      // For remaining properties, see TargetLags objects
    },
    "targetRollingWindowSize": {
      "mode": "string"
      // For remaining properties, see TargetRollingWindowSize objects
    },
    "timeColumnName": "string",
    "timeSeriesIdColumnNames": [ "string" ],
    "useStl": "string"
  },
  "limitSettings": {
    "enableEarlyTermination": "bool",
    "exitScore": "int",
    "maxConcurrentTrials": "int",
    "maxCoresPerTrial": "int",
    "maxTrials": "int",
    "timeout": "string",
    "trialTimeout": "string"
  },
  "nCrossValidations": {
    "mode": "string"
    // For remaining properties, see NCrossValidations objects
  },
  "primaryMetric": "string",
  "taskType": "Forecasting",
  "testData": {
    "description": "string",
    "jobInputType": "string",
    "mode": "string",
    "uri": "string"
  },
  "testDataSize": "int",
  "trainingSettings": {
    "allowedTrainingAlgorithms": [ "string" ],
    "blockedTrainingAlgorithms": [ "string" ],
    "enableDnnTraining": "bool",
    "enableModelExplainability": "bool",
    "enableOnnxCompatibleModels": "bool",
    "enableStackEnsemble": "bool",
    "enableVoteEnsemble": "bool",
    "ensembleModelDownloadTimeout": "string",
    "stackEnsembleSettings": {
      "stackMetaLearnerKWargs": {},
      "stackMetaLearnerTrainPercentage": "int",
      "stackMetaLearnerType": "string"
    }
  },
  "validationData": {
    "description": "string",
    "jobInputType": "string",
    "mode": "string",
    "uri": "string"
  },
  "validationDataSize": "int",
  "weightColumnName": "string"
}

針對 ImageClassification,請使用:

{
  "limitSettings": {
    "maxConcurrentTrials": "int",
    "maxTrials": "int",
    "timeout": "string"
  },
  "modelSettings": {
    "advancedSettings": "string",
    "amsGradient": "bool",
    "augmentations": "string",
    "beta1": "int",
    "beta2": "int",
    "checkpointFrequency": "int",
    "checkpointModel": {
      "description": "string",
      "jobInputType": "string",
      "mode": "string",
      "uri": "string"
    },
    "checkpointRunId": "string",
    "distributed": "bool",
    "earlyStopping": "bool",
    "earlyStoppingDelay": "int",
    "earlyStoppingPatience": "int",
    "enableOnnxNormalization": "bool",
    "evaluationFrequency": "int",
    "gradientAccumulationStep": "int",
    "layersToFreeze": "int",
    "learningRate": "int",
    "learningRateScheduler": "string",
    "modelName": "string",
    "momentum": "int",
    "nesterov": "bool",
    "numberOfEpochs": "int",
    "numberOfWorkers": "int",
    "optimizer": "string",
    "randomSeed": "int",
    "stepLRGamma": "int",
    "stepLRStepSize": "int",
    "trainingBatchSize": "int",
    "trainingCropSize": "int",
    "validationBatchSize": "int",
    "validationCropSize": "int",
    "validationResizeSize": "int",
    "warmupCosineLRCycles": "int",
    "warmupCosineLRWarmupEpochs": "int",
    "weightDecay": "int",
    "weightedLoss": "int"
  },
  "primaryMetric": "string",
  "searchSpace": [
    {
      "amsGradient": "string",
      "augmentations": "string",
      "beta1": "string",
      "beta2": "string",
      "distributed": "string",
      "earlyStopping": "string",
      "earlyStoppingDelay": "string",
      "earlyStoppingPatience": "string",
      "enableOnnxNormalization": "string",
      "evaluationFrequency": "string",
      "gradientAccumulationStep": "string",
      "layersToFreeze": "string",
      "learningRate": "string",
      "learningRateScheduler": "string",
      "modelName": "string",
      "momentum": "string",
      "nesterov": "string",
      "numberOfEpochs": "string",
      "numberOfWorkers": "string",
      "optimizer": "string",
      "randomSeed": "string",
      "stepLRGamma": "string",
      "stepLRStepSize": "string",
      "trainingBatchSize": "string",
      "trainingCropSize": "string",
      "validationBatchSize": "string",
      "validationCropSize": "string",
      "validationResizeSize": "string",
      "warmupCosineLRCycles": "string",
      "warmupCosineLRWarmupEpochs": "string",
      "weightDecay": "string",
      "weightedLoss": "string"
    }
  ],
  "sweepSettings": {
    "earlyTermination": {
      "delayEvaluation": "int",
      "evaluationInterval": "int",
      "policyType": "string"
      // For remaining properties, see EarlyTerminationPolicy objects
    },
    "samplingAlgorithm": "string"
  },
  "taskType": "ImageClassification",
  "validationData": {
    "description": "string",
    "jobInputType": "string",
    "mode": "string",
    "uri": "string"
  },
  "validationDataSize": "int"
}

針對 ImageClassificationMultilabel,請使用:

{
  "limitSettings": {
    "maxConcurrentTrials": "int",
    "maxTrials": "int",
    "timeout": "string"
  },
  "modelSettings": {
    "advancedSettings": "string",
    "amsGradient": "bool",
    "augmentations": "string",
    "beta1": "int",
    "beta2": "int",
    "checkpointFrequency": "int",
    "checkpointModel": {
      "description": "string",
      "jobInputType": "string",
      "mode": "string",
      "uri": "string"
    },
    "checkpointRunId": "string",
    "distributed": "bool",
    "earlyStopping": "bool",
    "earlyStoppingDelay": "int",
    "earlyStoppingPatience": "int",
    "enableOnnxNormalization": "bool",
    "evaluationFrequency": "int",
    "gradientAccumulationStep": "int",
    "layersToFreeze": "int",
    "learningRate": "int",
    "learningRateScheduler": "string",
    "modelName": "string",
    "momentum": "int",
    "nesterov": "bool",
    "numberOfEpochs": "int",
    "numberOfWorkers": "int",
    "optimizer": "string",
    "randomSeed": "int",
    "stepLRGamma": "int",
    "stepLRStepSize": "int",
    "trainingBatchSize": "int",
    "trainingCropSize": "int",
    "validationBatchSize": "int",
    "validationCropSize": "int",
    "validationResizeSize": "int",
    "warmupCosineLRCycles": "int",
    "warmupCosineLRWarmupEpochs": "int",
    "weightDecay": "int",
    "weightedLoss": "int"
  },
  "primaryMetric": "string",
  "searchSpace": [
    {
      "amsGradient": "string",
      "augmentations": "string",
      "beta1": "string",
      "beta2": "string",
      "distributed": "string",
      "earlyStopping": "string",
      "earlyStoppingDelay": "string",
      "earlyStoppingPatience": "string",
      "enableOnnxNormalization": "string",
      "evaluationFrequency": "string",
      "gradientAccumulationStep": "string",
      "layersToFreeze": "string",
      "learningRate": "string",
      "learningRateScheduler": "string",
      "modelName": "string",
      "momentum": "string",
      "nesterov": "string",
      "numberOfEpochs": "string",
      "numberOfWorkers": "string",
      "optimizer": "string",
      "randomSeed": "string",
      "stepLRGamma": "string",
      "stepLRStepSize": "string",
      "trainingBatchSize": "string",
      "trainingCropSize": "string",
      "validationBatchSize": "string",
      "validationCropSize": "string",
      "validationResizeSize": "string",
      "warmupCosineLRCycles": "string",
      "warmupCosineLRWarmupEpochs": "string",
      "weightDecay": "string",
      "weightedLoss": "string"
    }
  ],
  "sweepSettings": {
    "earlyTermination": {
      "delayEvaluation": "int",
      "evaluationInterval": "int",
      "policyType": "string"
      // For remaining properties, see EarlyTerminationPolicy objects
    },
    "samplingAlgorithm": "string"
  },
  "taskType": "ImageClassificationMultilabel",
  "validationData": {
    "description": "string",
    "jobInputType": "string",
    "mode": "string",
    "uri": "string"
  },
  "validationDataSize": "int"
}

針對 ImageInstanceSegmentation,請使用:

{
  "limitSettings": {
    "maxConcurrentTrials": "int",
    "maxTrials": "int",
    "timeout": "string"
  },
  "modelSettings": {
    "advancedSettings": "string",
    "amsGradient": "bool",
    "augmentations": "string",
    "beta1": "int",
    "beta2": "int",
    "boxDetectionsPerImage": "int",
    "boxScoreThreshold": "int",
    "checkpointFrequency": "int",
    "checkpointModel": {
      "description": "string",
      "jobInputType": "string",
      "mode": "string",
      "uri": "string"
    },
    "checkpointRunId": "string",
    "distributed": "bool",
    "earlyStopping": "bool",
    "earlyStoppingDelay": "int",
    "earlyStoppingPatience": "int",
    "enableOnnxNormalization": "bool",
    "evaluationFrequency": "int",
    "gradientAccumulationStep": "int",
    "imageSize": "int",
    "layersToFreeze": "int",
    "learningRate": "int",
    "learningRateScheduler": "string",
    "maxSize": "int",
    "minSize": "int",
    "modelName": "string",
    "modelSize": "string",
    "momentum": "int",
    "multiScale": "bool",
    "nesterov": "bool",
    "nmsIouThreshold": "int",
    "numberOfEpochs": "int",
    "numberOfWorkers": "int",
    "optimizer": "string",
    "randomSeed": "int",
    "stepLRGamma": "int",
    "stepLRStepSize": "int",
    "tileGridSize": "string",
    "tileOverlapRatio": "int",
    "tilePredictionsNmsThreshold": "int",
    "trainingBatchSize": "int",
    "validationBatchSize": "int",
    "validationIouThreshold": "int",
    "validationMetricType": "string",
    "warmupCosineLRCycles": "int",
    "warmupCosineLRWarmupEpochs": "int",
    "weightDecay": "int"
  },
  "primaryMetric": "string",
  "searchSpace": [
    {
      "amsGradient": "string",
      "augmentations": "string",
      "beta1": "string",
      "beta2": "string",
      "boxDetectionsPerImage": "string",
      "boxScoreThreshold": "string",
      "distributed": "string",
      "earlyStopping": "string",
      "earlyStoppingDelay": "string",
      "earlyStoppingPatience": "string",
      "enableOnnxNormalization": "string",
      "evaluationFrequency": "string",
      "gradientAccumulationStep": "string",
      "imageSize": "string",
      "layersToFreeze": "string",
      "learningRate": "string",
      "learningRateScheduler": "string",
      "maxSize": "string",
      "minSize": "string",
      "modelName": "string",
      "modelSize": "string",
      "momentum": "string",
      "multiScale": "string",
      "nesterov": "string",
      "nmsIouThreshold": "string",
      "numberOfEpochs": "string",
      "numberOfWorkers": "string",
      "optimizer": "string",
      "randomSeed": "string",
      "stepLRGamma": "string",
      "stepLRStepSize": "string",
      "tileGridSize": "string",
      "tileOverlapRatio": "string",
      "tilePredictionsNmsThreshold": "string",
      "trainingBatchSize": "string",
      "validationBatchSize": "string",
      "validationIouThreshold": "string",
      "validationMetricType": "string",
      "warmupCosineLRCycles": "string",
      "warmupCosineLRWarmupEpochs": "string",
      "weightDecay": "string"
    }
  ],
  "sweepSettings": {
    "earlyTermination": {
      "delayEvaluation": "int",
      "evaluationInterval": "int",
      "policyType": "string"
      // For remaining properties, see EarlyTerminationPolicy objects
    },
    "samplingAlgorithm": "string"
  },
  "taskType": "ImageInstanceSegmentation",
  "validationData": {
    "description": "string",
    "jobInputType": "string",
    "mode": "string",
    "uri": "string"
  },
  "validationDataSize": "int"
}

針對 ImageObjectDetection,請使用:

{
  "limitSettings": {
    "maxConcurrentTrials": "int",
    "maxTrials": "int",
    "timeout": "string"
  },
  "modelSettings": {
    "advancedSettings": "string",
    "amsGradient": "bool",
    "augmentations": "string",
    "beta1": "int",
    "beta2": "int",
    "boxDetectionsPerImage": "int",
    "boxScoreThreshold": "int",
    "checkpointFrequency": "int",
    "checkpointModel": {
      "description": "string",
      "jobInputType": "string",
      "mode": "string",
      "uri": "string"
    },
    "checkpointRunId": "string",
    "distributed": "bool",
    "earlyStopping": "bool",
    "earlyStoppingDelay": "int",
    "earlyStoppingPatience": "int",
    "enableOnnxNormalization": "bool",
    "evaluationFrequency": "int",
    "gradientAccumulationStep": "int",
    "imageSize": "int",
    "layersToFreeze": "int",
    "learningRate": "int",
    "learningRateScheduler": "string",
    "maxSize": "int",
    "minSize": "int",
    "modelName": "string",
    "modelSize": "string",
    "momentum": "int",
    "multiScale": "bool",
    "nesterov": "bool",
    "nmsIouThreshold": "int",
    "numberOfEpochs": "int",
    "numberOfWorkers": "int",
    "optimizer": "string",
    "randomSeed": "int",
    "stepLRGamma": "int",
    "stepLRStepSize": "int",
    "tileGridSize": "string",
    "tileOverlapRatio": "int",
    "tilePredictionsNmsThreshold": "int",
    "trainingBatchSize": "int",
    "validationBatchSize": "int",
    "validationIouThreshold": "int",
    "validationMetricType": "string",
    "warmupCosineLRCycles": "int",
    "warmupCosineLRWarmupEpochs": "int",
    "weightDecay": "int"
  },
  "primaryMetric": "string",
  "searchSpace": [
    {
      "amsGradient": "string",
      "augmentations": "string",
      "beta1": "string",
      "beta2": "string",
      "boxDetectionsPerImage": "string",
      "boxScoreThreshold": "string",
      "distributed": "string",
      "earlyStopping": "string",
      "earlyStoppingDelay": "string",
      "earlyStoppingPatience": "string",
      "enableOnnxNormalization": "string",
      "evaluationFrequency": "string",
      "gradientAccumulationStep": "string",
      "imageSize": "string",
      "layersToFreeze": "string",
      "learningRate": "string",
      "learningRateScheduler": "string",
      "maxSize": "string",
      "minSize": "string",
      "modelName": "string",
      "modelSize": "string",
      "momentum": "string",
      "multiScale": "string",
      "nesterov": "string",
      "nmsIouThreshold": "string",
      "numberOfEpochs": "string",
      "numberOfWorkers": "string",
      "optimizer": "string",
      "randomSeed": "string",
      "stepLRGamma": "string",
      "stepLRStepSize": "string",
      "tileGridSize": "string",
      "tileOverlapRatio": "string",
      "tilePredictionsNmsThreshold": "string",
      "trainingBatchSize": "string",
      "validationBatchSize": "string",
      "validationIouThreshold": "string",
      "validationMetricType": "string",
      "warmupCosineLRCycles": "string",
      "warmupCosineLRWarmupEpochs": "string",
      "weightDecay": "string"
    }
  ],
  "sweepSettings": {
    "earlyTermination": {
      "delayEvaluation": "int",
      "evaluationInterval": "int",
      "policyType": "string"
      // For remaining properties, see EarlyTerminationPolicy objects
    },
    "samplingAlgorithm": "string"
  },
  "taskType": "ImageObjectDetection",
  "validationData": {
    "description": "string",
    "jobInputType": "string",
    "mode": "string",
    "uri": "string"
  },
  "validationDataSize": "int"
}

針對 回歸,請使用:

{
  "cvSplitColumnNames": [ "string" ],
  "featurizationSettings": {
    "blockedTransformers": [ "string" ],
    "columnNameAndTypes": {
      "{customized property}": "string"
    },
    "datasetLanguage": "string",
    "enableDnnFeaturization": "bool",
    "mode": "string",
    "transformerParams": {
      "{customized property}": [
        {
          "fields": [ "string" ],
          "parameters": {}
        }
      ]
    }
  },
  "limitSettings": {
    "enableEarlyTermination": "bool",
    "exitScore": "int",
    "maxConcurrentTrials": "int",
    "maxCoresPerTrial": "int",
    "maxTrials": "int",
    "timeout": "string",
    "trialTimeout": "string"
  },
  "nCrossValidations": {
    "mode": "string"
    // For remaining properties, see NCrossValidations objects
  },
  "primaryMetric": "string",
  "taskType": "Regression",
  "testData": {
    "description": "string",
    "jobInputType": "string",
    "mode": "string",
    "uri": "string"
  },
  "testDataSize": "int",
  "trainingSettings": {
    "allowedTrainingAlgorithms": [ "string" ],
    "blockedTrainingAlgorithms": [ "string" ],
    "enableDnnTraining": "bool",
    "enableModelExplainability": "bool",
    "enableOnnxCompatibleModels": "bool",
    "enableStackEnsemble": "bool",
    "enableVoteEnsemble": "bool",
    "ensembleModelDownloadTimeout": "string",
    "stackEnsembleSettings": {
      "stackMetaLearnerKWargs": {},
      "stackMetaLearnerTrainPercentage": "int",
      "stackMetaLearnerType": "string"
    }
  },
  "validationData": {
    "description": "string",
    "jobInputType": "string",
    "mode": "string",
    "uri": "string"
  },
  "validationDataSize": "int",
  "weightColumnName": "string"
}

針對 TextClassification,請使用:

{
  "featurizationSettings": {
    "datasetLanguage": "string"
  },
  "limitSettings": {
    "maxConcurrentTrials": "int",
    "maxTrials": "int",
    "timeout": "string"
  },
  "primaryMetric": "string",
  "taskType": "TextClassification",
  "validationData": {
    "description": "string",
    "jobInputType": "string",
    "mode": "string",
    "uri": "string"
  }
}

針對 TextClassificationMultilabel,請使用:

{
  "featurizationSettings": {
    "datasetLanguage": "string"
  },
  "limitSettings": {
    "maxConcurrentTrials": "int",
    "maxTrials": "int",
    "timeout": "string"
  },
  "taskType": "TextClassificationMultilabel",
  "validationData": {
    "description": "string",
    "jobInputType": "string",
    "mode": "string",
    "uri": "string"
  }
}

針對 TextNER,請使用:

{
  "featurizationSettings": {
    "datasetLanguage": "string"
  },
  "limitSettings": {
    "maxConcurrentTrials": "int",
    "maxTrials": "int",
    "timeout": "string"
  },
  "taskType": "TextNER",
  "validationData": {
    "description": "string",
    "jobInputType": "string",
    "mode": "string",
    "uri": "string"
  }
}

屬性值

Microsoft.MachineLearningServices/workspaces/jobs

名稱 說明 價值觀
apiVersion API 版本 “2024-07-01-預覽”
名稱 資源名稱 字串

約束:
模式 = ^[a-zA-Z0-9][a-zA-Z0-9\-_]{0,254}$ (必要)
屬性 [必要]實體的其他屬性。 JobBaseProperties (必要)
型別 資源類型 “Microsoft.MachineLearningServices/workspaces/jobs”

所有節點

名稱 說明 價值觀
nodesValueType [必要]Nodes 值的類型 'All' (必要)

AmlToken

名稱 說明 價值觀
身份類型 [必要]指定身分識別架構的類型。 'AMLToken' (必要)

自動預測地平線

名稱 說明 價值觀
模式 [必要]設定預測地平線值選取模式。 'Auto' (必要)

AutoMLJob

名稱 說明 價值觀
環境ID 作業之環境規格的 ARM 資源識別碼。
如果未提供,這是選擇性的值,AutoML 會在執行作業時,將此預設為 Production AutoML 策展環境版本。
字串
環境變數 作業中包含的環境變數。 AutoMLJobEnvironmentVariables
職位類型 [必要]指定作業的類型。 'AutoML' (必要)
輸出 對應作業中使用的輸出數據系結。 AutoMLJob 輸出
queueSettings 作業的佇列設定 佇列設置
資源 作業的計算資源組態。 JobResourceConfiguration
taskDetails [必要]這代表可以是其中一個數據表/NLP/Image 的案例 AutoMLVertical (必要)

AutoMLJobEnvironmentVariables

名稱 說明 價值觀

AutoMLJob 輸出

名稱 說明 價值觀

AutoMLVertical

名稱 說明 價值觀
logVerbosity (對數詳細程度) 作業的記錄詳細資訊。 “嚴重”
“調試”
“錯誤”
'資訊'
'NotSet'
“警告”
targetColumnName 目標數據行名稱:這是預測值數據行。
也稱為分類工作內容中的標籤數據行名稱。
字串
任務類型 設定為 [分類] 類型為 [ 分類]。 設定為 [預測] 類型為 [ 預測]。 設定為 ImageClassification 類型的 『ImageClassification』。 設定為 ImageClassificationMultilabel 類型的 'ImageClassificationMultilabel'。 針對 ImageInstanceSegmentation 類型,設定為 'ImageInstanceSegmentation'。 針對 ImageObjectDetection 類型,設定為 'ImageObjectDetection'。 將類型回歸設定為 [ 回歸]。 設定為 TextClassification 類型的 『TextClassification』。 針對 TextClassificationMultilabel 類型,設定為 『TextClassificationMultilabel』。 設定為 TextNer 類型的 『TextNER』。 '分類'
「預測」
'ImageClassification' (圖像分類)
'ImageClassificationMultilabel'
'ImageInstanceSegmentation' (圖像實例分段)
'ImageObjectDetection' (圖像對象檢測)
“回歸”
'TextClassification' (文本分類)
'TextClassificationMultilabel'
'TextNER' (必要)
trainingData (訓練數據) [必要]定型數據輸入。 MLTableJobInput (必要)

AutoNCross驗證

名稱 說明 價值觀
模式 [必要]判斷 N-Cross 驗證的模式。 'Auto' (必要)

自動季節性

名稱 說明 價值觀
模式 [必要]季節性模式。 'Auto' (必要)

AutoTargetLags 自動目標滯後

名稱 說明 價值觀
模式 [必要]設定目標延隔模式 - 自動/自定義 'Auto' (必要)

AutoTargetRollingWindowSize (自動目標滾動視窗大小)

名稱 說明 價值觀
模式 [必要]TargetRollingWindowSiz 偵測模式。 'Auto' (必要)

AzureDevOpsWebhook

名稱 說明 價值觀
webhook類型 [必要]指定要傳送回呼的服務類型 'AzureDevOps' (必要)

AzureOpenAiFineTuning

名稱 說明 價值觀
超參數 用於微調 Azure Open AI 模型的 HyperParameters。 AzureOpenAiHyperParameters
模型提供程式 [必要]用來判斷微調類型的列舉。 'AzureOpenAI' (必要)

AzureOpenAiHyperParameters

名稱 說明 價值觀
批次大小 每個批次中的範例數目。 批次大小越大,表示模型參數的更新頻率越低,而變異數越少。 整數 (int)
learningRateMultiplier 的 學習率的縮放比例。 較小的學習率可能有助於避免過度調整。 整數 (int)
nEpochs 訓練模型的週期數目。 Epoch 是指透過訓練資料集的一個完整週期。 整數 (int)

強盜政策 (BanditPolicy)

名稱 說明 價值觀
policyType [必要]原則設定的名稱 “強盜”(必要)
slackAmount 從最佳執行執行中允許的絕對距離。 整數 (int)
slackFactor 的 與最佳執行距離的允許距離比率。 整數 (int)

BayesianSamplingAlgorithm

名稱 說明 價值觀
採樣演算法類型 [必要]用來產生超參數值的演算法,以及組態屬性 “貝氏” (必要)

分類

名稱 說明 價值觀
cvSplitColumnNames 要用於 CVSplit 資料的數據行。 字串[]
featurizationSettings AutoML 作業所需的特徵化輸入。 TableVerticalFeaturizationSettings
limitSettings AutoMLJob 的執行條件約束。 TableVerticalLimitSettings
nCrossValidations 要套用至定型數據集的交叉驗證折疊數目
未提供驗證數據集時。
NCross驗證
積極標籤 二進位計量計算的正標籤。 字串
primaryMetric 工作的主要計量。 '精確度'
'AUCWeighted'
'AveragePrecisionScoreWeighted'
'NormMacroRecall' (規範巨集調用)
'PrecisionScoreWeighted'
任務類型 [必要]AutoMLJob 的工作類型。 '分類' (必要)
測試數據 測試數據輸入。 MLTableJobInput
testDataSize 必須為驗證目的而保留的測試數據集分數。
介於 (0.0 、 1.0) 之間的值
未提供驗證數據集時套用。
整數 (int)
trainingSettings AutoML 作業定型階段的輸入。 分類訓練設置
validationData (驗證資料) 驗證數據輸入。 MLTableJobInput
validationDataSize 需要為驗證目的預留的定型數據集分數。
介於 (0.0 、 1.0) 之間的值
未提供驗證數據集時套用。
整數 (int)
權重列名稱 範例權數數據行的名稱。 自動化 ML 支援加權數據行做為輸入,導致數據中的數據列向上或向下加權。 字串

分類訓練設置

名稱 說明 價值觀
allowedTrainingAlgorithms 分類工作的允許模型。 包含任何的字串數組:
'伯努利樸素貝葉斯'
“決策樹”
'ExtremeRandomTrees'
'梯度提升'
'KNN'
'LightGBM'
'線性 SVM'
'LogisticRegression' (邏輯回歸)
'多項式樸素貝葉斯'
'隨機森林'
'新幣'
“SVM”
'XGBoostClassifier'
blockedTrainingAlgorithms 分類工作的封鎖模型。 包含任何的字串數組:
'伯努利樸素貝葉斯'
“決策樹”
'ExtremeRandomTrees'
'梯度提升'
'KNN'
'LightGBM'
'線性 SVM'
'LogisticRegression' (邏輯回歸)
'多項式樸素貝葉斯'
'隨機森林'
'新幣'
“SVM”
'XGBoostClassifier'
enableDnnTraining 啟用 DNN 模型的建議。 布爾 (bool)
enableModelExplainability 在最佳模型上開啟可解釋性的旗標。 布爾 (bool)
enableOnnxCompatibleModels 啟用 onnx 相容模型的旗標。 布爾 (bool)
enableStackEnsemble 啟用堆疊合奏執行。 布爾 (bool)
enableVoteEnsemble 啟用投票合奏執行。 布爾 (bool)
ensembleModelDownloadTimeout 在 VotingEnsemble 和 StackEnsemble 模型產生期間,會下載先前子回合中的多個安裝模型。
如果需要更多時間,請使用高於 300 秒的值來設定此參數。
字串
stackEnsembleSettings 堆疊合奏執行的堆疊合奏設定。 StackEnsembleSettings

ColumnTransformer (列變換器)

名稱 說明 價值觀
領域 要套用轉換器邏輯的欄位。 字串[]
參數 要傳遞至轉換器的不同屬性。
輸入必須是 JSON 格式的索引鍵、值組字典。
任何

CommandJob

名稱 說明 價值觀
代碼Id 程式代碼資產的 ARM 資源識別碼。 字串
指令 [必要]在作業啟動時執行的命令。 例如 “python train.py” 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)
分配 作業的散發組態。 如果設定,這應該是其中一個 Mpi、Tensorflow、PyTorch 或 null。 分發配置
環境ID [必要]作業之環境規格的 ARM 資源識別碼。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)
環境變數 作業中包含的環境變數。 CommandJobEnvironmentVariables
輸入 對應作業中使用的輸入數據系結。 CommandJobInputs 命令作業輸入
職位類型 [必要]指定作業的類型。 'Command' (必要)
限制 命令作業限制。 CommandJobLimits 命令作業限制
輸出 對應作業中使用的輸出數據系結。 CommandJobOutputs 命令作業輸出
queueSettings 作業的佇列設定 佇列設置
資源 作業的計算資源組態。 JobResourceConfiguration

CommandJobEnvironmentVariables

名稱 說明 價值觀

CommandJobInputs 命令作業輸入

名稱 說明 價值觀

CommandJobLimits 命令作業限制

名稱 說明 價值觀
jobLimits類型 [必要]JobLimit 類型。 “命令”
'掃掠' (必要)
超時 ISO 8601 格式的最大執行持續時間,之後作業將會取消。 僅支援有效位數為秒數的持續時間。 字串

CommandJobOutputs 命令作業輸出

名稱 說明 價值觀

CustomForecastHorizon

名稱 說明 價值觀
模式 [必要]設定預測地平線值選取模式。 'Custom' (必要)
價值 [必要]預測地平線值。 int (必要)

CustomModelFineTuning 微調

名稱 說明 價值觀
超參數 微調自定義模型的 HyperParameters。 CustomModelFineTuningHyperParameters
模型提供程式 [必要]用來判斷微調類型的列舉。 'Custom' (必要)

CustomModelFineTuningHyperParameters

名稱 說明 價值觀

CustomModelJobInput (自定義模型作業輸入)

名稱 說明 價值觀
jobInputType (作業輸入類型) [必要]指定作業的類型。 'custom_model' (必要)
模式 輸入資產傳遞模式。 “直接”
'下載'
'EvalDownload'
'EvalMount'
'ReadOnlyMount'
“ReadWriteMount”
統一資源識別碼 (URI) [必要]輸入資產 URI。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)

CustomModelJobOutput

名稱 說明 價值觀
jobOutputType [必要]指定作業的類型。 'custom_model' (必要)
模式 輸出資產傳遞模式。 “直接”
“ReadWriteMount”
'上傳'
統一資源識別碼 (URI) 輸出資產 URI。 字串

CustomNCross驗證

名稱 說明 價值觀
模式 [必要]判斷 N-Cross 驗證的模式。 'Custom' (必要)
價值 [必要]N-Cross 驗證值。 int (必要)

CustomSeasonality (自定義季節性)

名稱 說明 價值觀
模式 [必要]季節性模式。 'Custom' (必要)
價值 [必要]季節性值。 int (必要)

CustomTargetLags (自定義目標滯後)

名稱 說明 價值觀
模式 [必要]設定目標延隔模式 - 自動/自定義 'Custom' (必要)
價值觀 [必要]設定目標延遲值。 int[] (必要)

CustomTargetRollingWindowSize

名稱 說明 價值觀
模式 [必要]TargetRollingWindowSiz 偵測模式。 'Custom' (必要)
價值 [必要]TargetRollingWindowSize 值。 int (必要)

分發配置

名稱 說明 價值觀
distributionType 設定為 Mpi 類型的 『Mpi』。 將類型 PyTorch 設定為 『PyTorch』。 將 設定為 TensorFlow 類型的 『TensorFlow』。 'Mpi'
“PyTorch”
'TensorFlow' (必要)

提早終止政策

名稱 說明 價值觀
delay評估 延遲第一次評估的間隔數目。 整數 (int)
evaluationInterval 原則評估之間的間隔(執行次數)。 整數 (int)
policyType 針對 BanditPolicy 類型,設定為 『Bandit』。 針對 MedianStoppingPolicy 類型,設定為 'MedianStoppingPolicy'。 針對 TruncationSelectionPolicy 類型,設定為 『TruncationSelection』 “強盜”
'MedianStopping' (中位數停止)
'TruncationSelection' (必要)

微調作業

名稱 說明 價值觀
微調細節 [必要] FineTuningVertical (必要)
職位類型 [必要]指定作業的類型。 'FineTuning' (必要)
輸出 [必要] FineTuningJobOutputs (必要)
queueSettings 作業的佇列設定 佇列設置
資源 作業的實例類型和其他資源 工作資源

FineTuningJobOutputs(微調作業輸出)

名稱 說明 價值觀

微調垂直

名稱 說明 價值觀
型號 [必要]用於微調的輸入模型。 MLFlowModelJobInput (必要)
模型提供程式 針對 AzureOpen AiFineTuning 類型,設定為 『AzureOpenAI』。 設定為 CustomModelFineTuning 類型的 [自定義]。 'AzureOpenAI'
'Custom' (必要)
任務類型 [必要]微調工作類型。 'ChatCompletion'
'ImageClassification' (圖像分類)
'ImageInstanceSegmentation' (圖像實例分段)
'ImageObjectDetection' (圖像對象檢測)
“問答”
'TextClassification' (文本分類)
'TextCompletion'
'TextSummarization' (文本摘要)
'TextTranslation'
'TokenClassification'
'VideoMultiObjectTracking' (必要)
trainingData (訓練數據) [必要]定型數據以進行微調。 JobInput (必要)
validationData (驗證資料) 微調的驗證數據。 作業輸入

預測地平線

名稱 說明 價值觀
模式 針對 AutoForecastHorizon 類型,設定為 'Auto'。 針對 CustomForecastHorizon 類型,設定為 'Custom'。 '自動'
'Custom' (必要)

預測

名稱 說明 價值觀
cvSplitColumnNames 要用於 CVSplit 資料的數據行。 字串[]
featurizationSettings AutoML 作業所需的特徵化輸入。 TableVerticalFeaturizationSettings
forecasting設置 預測工作特定輸入。 ForecastingSettings
limitSettings AutoMLJob 的執行條件約束。 TableVerticalLimitSettings
nCrossValidations 要套用至定型數據集的交叉驗證折疊數目
未提供驗證數據集時。
NCross驗證
primaryMetric 預測工作的主要計量。 'NormalizedMeanAbsoluteError'
'NormalizedRootMeanSquaredError'
'R2Score'
'SpearmanCorrelation'
任務類型 [必要]AutoMLJob 的工作類型。 “預測” (必要)
測試數據 測試數據輸入。 MLTableJobInput
testDataSize 必須為驗證目的而保留的測試數據集分數。
介於 (0.0 、 1.0) 之間的值
未提供驗證數據集時套用。
整數 (int)
trainingSettings AutoML 作業定型階段的輸入。 ForecastingTrainingSettings
validationData (驗證資料) 驗證數據輸入。 MLTableJobInput
validationDataSize 需要為驗證目的預留的定型數據集分數。
介於 (0.0 、 1.0) 之間的值
未提供驗證數據集時套用。
整數 (int)
權重列名稱 範例權數數據行的名稱。 自動化 ML 支援加權數據行做為輸入,導致數據中的數據列向上或向下加權。 字串

ForecastingSettings

名稱 說明 價值觀
國家/地區或地區為假期 用於預測工作的假日國家或地區。
這些應該是 ISO 3166 兩個字母的國家/地區代碼,例如 'US' 或 'GB'。
字串
cvStepSize 一個 CV 折疊的原點時間與下一個折疊之間的期間數。 為了
例如,如果 CVStepSize = 3 表示每日數據,則每個折疊的原點時間將會是
相隔三天。
整數 (int)
featureLags 使用 『auto』 或 null 產生數值特徵延遲的旗標。 '自動'
“無”
forecastHorizon (預測地平線) 所需的最大預測範圍,以時間序列頻率單位為單位。 預測地平線
頻率 預測時,此參數代表預期預測的期間,例如每日、每周、每年等。預測頻率預設為數據集頻率。 字串
季節性 將時間序列季節性設定為數列頻率的整數倍數。
如果季節性設定為 『auto』,則會推斷它。
季節性
shortSeriesHandlingConfig 定義 AutoML 是否應該處理短時間序列的參數。 '自動'
“掉落”
“無”
'墊'
targetAggregateFunction 要用來匯總時間序列目標數據行以符合使用者指定頻率的函式。
如果 TargetAggregateFunction 已設定,亦即不是 『None』,但未設定 freq 參數,則會引發錯誤。 可能的目標聚合函數包括:「sum」、“max”、“min” 和 “mean”。
“最大”
“卑鄙”
'敏'
“無”
'總和'
目標拉格 要從目標數據行延遲的過去期間數。 目標滯後
targetRollingWindowSize 用來建立目標數據行滾動視窗平均值的過去期間數目。 TargetRollingWindowSize
timeColumnName (時間列名稱) 時間數據行的名稱。 當預測以指定用於建置時間序列並推斷其頻率的輸入數據中的 datetime 數據行時,需要此參數。 字串
timeSeriesIdColumnNames 用來分組時間範圍的數據行名稱。 它可以用來建立多個數列。
如果未定義粒紋,則會假設數據集為一個時間序列。 此參數與工作類型預測搭配使用。
字串[]
useStl 設定時間序列目標數據行的 STL 分解。 “無”
“賽季”
'季節趨勢'

ForecastingTrainingSettings

名稱 說明 價值觀
allowedTrainingAlgorithms 用於預測工作的允許模型。 包含任何的字串數組:
“Arimax”
'AutoArima'
“平均”
“決策樹”
“彈性網”
'指數平滑'
'ExtremeRandomTrees'
'梯度提升'
'KNN'
“套索拉斯”
'LightGBM'
天真'
'先知'
'隨機森林'
'季節性平均值'
'季節性天真'
'新幣'
'TCNForecaster'
'XGBoostRegressor'
blockedTrainingAlgorithms 用於預測工作的封鎖模型。 包含任何的字串數組:
“Arimax”
'AutoArima'
“平均”
“決策樹”
“彈性網”
'指數平滑'
'ExtremeRandomTrees'
'梯度提升'
'KNN'
“套索拉斯”
'LightGBM'
天真'
'先知'
'隨機森林'
'季節性平均值'
'季節性天真'
'新幣'
'TCNForecaster'
'XGBoostRegressor'
enableDnnTraining 啟用 DNN 模型的建議。 布爾 (bool)
enableModelExplainability 在最佳模型上開啟可解釋性的旗標。 布爾 (bool)
enableOnnxCompatibleModels 啟用 onnx 相容模型的旗標。 布爾 (bool)
enableStackEnsemble 啟用堆疊合奏執行。 布爾 (bool)
enableVoteEnsemble 啟用投票合奏執行。 布爾 (bool)
ensembleModelDownloadTimeout 在 VotingEnsemble 和 StackEnsemble 模型產生期間,會下載先前子回合中的多個安裝模型。
如果需要更多時間,請使用高於 300 秒的值來設定此參數。
字串
stackEnsembleSettings 堆疊合奏執行的堆疊合奏設定。 StackEnsembleSettings

GridSampling演算法

名稱 說明 價值觀
採樣演算法類型 [必要]用來產生超參數值的演算法,以及組態屬性 'Grid' (必要)

身份配置

名稱 說明 價值觀
身份類型 將 設定為 AmlToken 類型的 『AMLToken』。 針對 ManagedIdentity 類型設定為 『Managed』。 針對 UserIdentity 類型,設定為 ' UserIdentity' 'AMLToken'
“託管”
'UserIdentity' (必要)

圖像分類

名稱 說明 價值觀
limitSettings [必要]限制 AutoML 作業的設定。 ImageLimitSettings (必要)
modelSettings 用於定型模型的設定。 ImageModelSettings分類
primaryMetric 要針對這項工作進行優化的主要計量。 '精確度'
'AUCWeighted'
'AveragePrecisionScoreWeighted'
'NormMacroRecall' (規範巨集調用)
'PrecisionScoreWeighted'
搜索空間 搜尋空間以取樣模型及其超參數的不同組合。 ImageModelDistributionSettingsClassification[]
掃描設置 模型掃掠和超參數掃掠相關設定。 ImageSweepSettings (圖像掃描設定)
任務類型 [必要]AutoMLJob 的工作類型。 'ImageClassification' (必要)
validationData (驗證資料) 驗證數據輸入。 MLTableJobInput
validationDataSize 需要為驗證目的預留的定型數據集分數。
介於 (0.0 、 1.0) 之間的值
未提供驗證數據集時套用。
整數 (int)

ImageClassificationMultilabel

名稱 說明 價值觀
limitSettings [必要]限制 AutoML 作業的設定。 ImageLimitSettings (必要)
modelSettings 用於定型模型的設定。 ImageModelSettings分類
primaryMetric 要針對這項工作進行優化的主要計量。 '精確度'
'AUCWeighted'
'AveragePrecisionScoreWeighted'
'欠條'
'NormMacroRecall' (規範巨集調用)
'PrecisionScoreWeighted'
搜索空間 搜尋空間以取樣模型及其超參數的不同組合。 ImageModelDistributionSettingsClassification[]
掃描設置 模型掃掠和超參數掃掠相關設定。 ImageSweepSettings (圖像掃描設定)
任務類型 [必要]AutoMLJob 的工作類型。 'ImageClassificationMultilabel' (必要)
validationData (驗證資料) 驗證數據輸入。 MLTableJobInput
validationDataSize 需要為驗證目的預留的定型數據集分數。
介於 (0.0 、 1.0) 之間的值
未提供驗證數據集時套用。
整數 (int)

ImageInstanceSegmentation

名稱 說明 價值觀
limitSettings [必要]限制 AutoML 作業的設定。 ImageLimitSettings (必要)
modelSettings 用於定型模型的設定。 ImageModelSettingsObjectDetection
primaryMetric 要針對這項工作進行優化的主要計量。 'MeanAveragePrecision'
搜索空間 搜尋空間以取樣模型及其超參數的不同組合。 ImageModelDistributionSettingsObjectDetection[]
掃描設置 模型掃掠和超參數掃掠相關設定。 ImageSweepSettings (圖像掃描設定)
任務類型 [必要]AutoMLJob 的工作類型。 'ImageInstanceSegmentation' (必要)
validationData (驗證資料) 驗證數據輸入。 MLTableJobInput
validationDataSize 需要為驗證目的預留的定型數據集分數。
介於 (0.0 、 1.0) 之間的值
未提供驗證數據集時套用。
整數 (int)

ImageLimitSettings

名稱 說明 價值觀
maxConcurrentTrials 並行 AutoML 反覆項目的數目上限。 整數 (int)
maxTrials AutoML 反覆項目的數目上限。 整數 (int)
超時 AutoML 作業逾時。 字串

ImageModelDistributionSettings分類

名稱 說明 價值觀
amsGradient 當優化器為 『adam』 或 'adamw' 時啟用AMSGrad。 字串
擴增 使用擴增的設定。 字串
貝塔1 優化器為 'adam' 或 'adamw' 時的 'beta1' 值。 必須是範圍 [0, 1] 內的浮點數。 字串
貝塔2 優化器為 'adam' 或 'adamw' 時的 'beta2' 值。 必須是範圍 [0, 1] 內的浮點數。 字串
分散式 是否要使用散發器定型。 字串
earlyStopping (早期停止) 在定型期間啟用早期停止邏輯。 字串
earlyStoppingDelay 在主要計量改進之前等待的最小 Epoch 或驗證評估數目
會追蹤早期停止。 必須是正整數。
字串
早點停止耐心 在之前沒有主要計量改進的最小 Epoch 或驗證評估數目
執行已停止。 必須是正整數。
字串
enableOnnxNormalization 匯出 ONNX 模型時啟用正規化。 字串
evaluationFrequency 評估驗證數據集以取得計量分數的頻率。 必須是正整數。 字串
梯度累積步驟 漸層累積表示執行未設定數目的 「GradAccumulationStep」 步驟
更新模型權數,同時累積這些步驟的漸層,然後使用
用來計算權數更新的累積漸層。 必須是正整數。
字串
layersToFreeze (圖層凍結) 要凍結模型的圖層數目。 必須是正整數。
例如,傳遞 2 作為 'seresnext' 值表示
凍結 layer0 和 layer1。 如需支援模型的完整清單,以及圖層凍結的詳細數據,請
請參閱:/azure/machine-learning/how-to-auto-train-image-models。
字串
learningRate 初始學習速率。 必須是範圍 [0, 1] 內的浮點數。 字串
learningRateScheduler 的 學習率排程器的類型。 必須是 『warmup_cosine』 或 'step'。 字串
型號名稱 要用於定型的模型名稱。
如需可用模型的詳細資訊,請流覽官方檔:
/azure/machine-learning/how-to-auto-train-image-models。
字串
動量 優化器為 『zn』 時的動量值。 必須是範圍 [0, 1] 內的浮點數。 字串
涅斯特羅夫 在優化器為 『zn』 時啟用 nesterov。 字串
紀元數 定型 Epoch 的數目。 必須是正整數。 字串
工人數 數據載入器背景工作角色的數目。 必須是非負整數。 字串
優化器 優化工具的類型。 必須是 『zn』、『adam』或 『adamw』。 字串
隨機種子 使用決定性定型時要使用的隨機種子。 字串
stepLR伽瑪 學習速率排程器為 『step』 時的 gamma 值。 必須是範圍 [0, 1] 內的浮點數。 字串
stepLRStepSize 學習速率排程器為 『step』 時的步驟大小值。 必須是正整數。 字串
trainingBatchSize 定型批次大小。 必須是正整數。 字串
trainingCropSize 影像裁剪大小,這是訓練數據集之類神經網路的輸入。 必須是正整數。 字串
validationBatchSize 驗證批次大小。 必須是正整數。 字串
validationCropSize 影像裁剪大小,這是用於驗證數據集之類神經網路的輸入。 必須是正整數。 字串
validationResizeSize 在裁剪驗證數據集之前要重設大小的影像大小。 必須是正整數。 字串
warmupCosineLRCycles 學習速率排程器為「warmup_cosine」時,餘弦迴圈的值。 必須是範圍 [0, 1] 內的浮點數。 字串
warmupCosineLRWarmupEpochs 學習速率排程器為「warmup_cosine」時,熱身 Epoch 的值。 必須是正整數。 字串
weightDecay (權重衰減) 優化器為 'zn'、'adam' 或 'adamw' 時的權數衰變值。 必須是範圍[0, 1] 中的浮點數。 字串
weightedLoss 加權減肥。 接受的值是 0,沒有加權損失。
1 用於與 sqrt 減肥。(class_weights)。 2 用於減肥與class_weights。 必須是 0 或 1 或 2。
字串

ImageModelDistributionSettingsObjectDetection

名稱 說明 價值觀
amsGradient 當優化器為 『adam』 或 'adamw' 時啟用AMSGrad。 字串
擴增 使用擴增的設定。 字串
貝塔1 優化器為 'adam' 或 'adamw' 時的 'beta1' 值。 必須是範圍 [0, 1] 內的浮點數。 字串
貝塔2 優化器為 'adam' 或 'adamw' 時的 'beta2' 值。 必須是範圍 [0, 1] 內的浮點數。 字串
boxDetectionsPerImage (每影像檢測) 所有類別的每個影像偵測數目上限。 必須是正整數。
注意:'yolov5' 演算法不支援此設定。
字串
boxScoreThreshold (盒分數閾值) 在推斷期間,只傳回大於分類分數的建議
BoxScoreThreshold。 必須是範圍[0, 1] 中的浮點數。
字串
分散式 是否要使用散發器定型。 字串
earlyStopping (早期停止) 在定型期間啟用早期停止邏輯。 字串
earlyStoppingDelay 在主要計量改進之前等待的最小 Epoch 或驗證評估數目
會追蹤早期停止。 必須是正整數。
字串
早點停止耐心 在之前沒有主要計量改進的最小 Epoch 或驗證評估數目
執行已停止。 必須是正整數。
字串
enableOnnxNormalization 匯出 ONNX 模型時啟用正規化。 字串
evaluationFrequency 評估驗證數據集以取得計量分數的頻率。 必須是正整數。 字串
梯度累積步驟 漸層累積表示執行未設定數目的 「GradAccumulationStep」 步驟
更新模型權數,同時累積這些步驟的漸層,然後使用
用來計算權數更新的累積漸層。 必須是正整數。
字串
影像大小 定型和驗證的影像大小。 必須是正整數。
注意:如果大小太大,定型回合可能會進入 CUDA OOM。
注意:只有 'yolov5' 演算法才支援此設定。
字串
layersToFreeze (圖層凍結) 要凍結模型的圖層數目。 必須是正整數。
例如,傳遞 2 作為 'seresnext' 值表示
凍結 layer0 和 layer1。 如需支援模型的完整清單,以及圖層凍結的詳細數據,請
請參閱:/azure/machine-learning/how-to-auto-train-image-models。
字串
learningRate 初始學習速率。 必須是範圍 [0, 1] 內的浮點數。 字串
learningRateScheduler 的 學習率排程器的類型。 必須是 『warmup_cosine』 或 'step'。 字串
最大尺寸 在將影像饋送至骨幹之前,要重新調整的影像大小上限。
必須是正整數。 注意:如果大小太大,定型執行可能會進入 CUDA OOM。
注意:'yolov5' 演算法不支援此設定。
字串
最小大小 將影像饋送至骨幹之前,要重新調整的影像大小下限。
必須是正整數。 注意:如果大小太大,定型執行可能會進入 CUDA OOM。
注意:'yolov5' 演算法不支援此設定。
字串
型號名稱 要用於定型的模型名稱。
如需可用模型的詳細資訊,請流覽官方檔:
/azure/machine-learning/how-to-auto-train-image-models。
字串
modelSize 模型大小。 必須是 'small'、'medium'、'large' 或 'xlarge'。
注意:如果模型大小太大,定型執行可能會進入 CUDA OOM。
注意:只有 'yolov5' 演算法才支援此設定。
字串
動量 優化器為 『zn』 時的動量值。 必須是範圍 [0, 1] 內的浮點數。 字串
多尺度 依 +/- 50%來啟用多重縮放影像大小。
注意:如果沒有足夠的 GPU 記憶體,定型執行可能會進入 CUDA OOM。
注意:只有 'yolov5' 演算法才支援此設定。
字串
涅斯特羅夫 在優化器為 『zn』 時啟用 nesterov。 字串
nmsIouThreshold 在NMS後置處理中的推斷期間使用的IOU臨界值。 必須是範圍 [0, 1] 中的 float。 字串
紀元數 定型 Epoch 的數目。 必須是正整數。 字串
工人數 數據載入器背景工作角色的數目。 必須是非負整數。 字串
優化器 優化工具的類型。 必須是 『zn』、『adam』或 『adamw』。 字串
隨機種子 使用決定性定型時要使用的隨機種子。 字串
stepLR伽瑪 學習速率排程器為 『step』 時的 gamma 值。 必須是範圍 [0, 1] 內的浮點數。 字串
stepLRStepSize 學習速率排程器為 『step』 時的步驟大小值。 必須是正整數。 字串
tileGridSize (平鋪網格大小) 用來貼上每個影像的網格線大小。 注意:TileGridSize 不得為
無可啟用小型物件偵測邏輯。 包含兩個整數的字串,其格式為一個。
注意:'yolov5' 演算法不支援此設定。
字串
tileOverlapRatio (平鋪重疊比率) 每個維度中相鄰磚之間的重疊比例。 必須是範圍 [0, 1] 中的浮點數。
注意:'yolov5' 演算法不支援此設定。
字串
tilePredictionsNmsThreshold 用來執行NMS的IOU閾值,同時合併磚和影像的預測。
用於驗證/推斷。 必須是範圍 [0, 1] 中的 float。
注意:'yolov5' 演算法不支援此設定。
NMS:非最大歸併
字串
trainingBatchSize 定型批次大小。 必須是正整數。 字串
validationBatchSize 驗證批次大小。 必須是正整數。 字串
validationIouThreshold 計算驗證計量時要使用的IOU閾值。 必須是範圍 [0, 1] 中的 float。 字串
validationMetricType 用於驗證計量的計量計算方法。 必須是 『none』、『coco』、'voc'或 'coco_voc'。 字串
warmupCosineLRCycles 學習速率排程器為「warmup_cosine」時,餘弦迴圈的值。 必須是範圍 [0, 1] 內的浮點數。 字串
warmupCosineLRWarmupEpochs 學習速率排程器為「warmup_cosine」時,熱身 Epoch 的值。 必須是正整數。 字串
weightDecay (權重衰減) 優化器為 'zn'、'adam' 或 'adamw' 時的權數衰變值。 必須是範圍[0, 1] 中的浮點數。 字串

ImageModelSettings分類

名稱 說明 價值觀
高級設置 進階案例的設定。 字串
amsGradient 當優化器為 『adam』 或 'adamw' 時啟用AMSGrad。 布爾 (bool)
擴增 使用擴增的設定。 字串
貝塔1 優化器為 'adam' 或 'adamw' 時的 'beta1' 值。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
貝塔2 優化器為 'adam' 或 'adamw' 時的 'beta2' 值。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
檢查點頻率 儲存模型檢查點的頻率。 必須是正整數。 整數 (int)
checkpoint模型 累加訓練的預先定型檢查點模型。 MLFlowModelJobInput
檢查點RunId 先前執行的標識碼,其具有累加式定型的預先定型檢查點。 字串
分散式 是否要使用分散式定型。 布爾 (bool)
earlyStopping (早期停止) 在定型期間啟用早期停止邏輯。 布爾 (bool)
earlyStoppingDelay 在主要計量改進之前等待的最小 Epoch 或驗證評估數目
會追蹤早期停止。 必須是正整數。
整數 (int)
早點停止耐心 在之前沒有主要計量改進的最小 Epoch 或驗證評估數目
執行已停止。 必須是正整數。
整數 (int)
enableOnnxNormalization 匯出 ONNX 模型時啟用正規化。 布爾 (bool)
evaluationFrequency 評估驗證數據集以取得計量分數的頻率。 必須是正整數。 整數 (int)
梯度累積步驟 漸層累積表示執行未設定數目的 「GradAccumulationStep」 步驟
更新模型權數,同時累積這些步驟的漸層,然後使用
用來計算權數更新的累積漸層。 必須是正整數。
整數 (int)
layersToFreeze (圖層凍結) 要凍結模型的圖層數目。 必須是正整數。
例如,傳遞 2 作為 'seresnext' 值表示
凍結 layer0 和 layer1。 如需支援模型的完整清單,以及圖層凍結的詳細數據,請
請參閱:/azure/machine-learning/how-to-auto-train-image-models。
整數 (int)
learningRate 初始學習速率。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
learningRateScheduler 的 學習率排程器的類型。 必須是 『warmup_cosine』 或 'step'。 “無”
“步驟”
'WarmupCosine'
型號名稱 要用於定型的模型名稱。
如需可用模型的詳細資訊,請流覽官方檔:
/azure/machine-learning/how-to-auto-train-image-models。
字串
動量 優化器為 『zn』 時的動量值。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
涅斯特羅夫 在優化器為 『zn』 時啟用 nesterov。 布爾 (bool)
紀元數 定型 Epoch 的數目。 必須是正整數。 整數 (int)
工人數 數據載入器背景工作角色的數目。 必須是非負整數。 整數 (int)
優化器 優化工具的類型。 “亞當”
“亞當”
“無”
'新元'
隨機種子 使用決定性定型時要使用的隨機種子。 整數 (int)
stepLR伽瑪 學習速率排程器為 『step』 時的 gamma 值。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
stepLRStepSize 學習速率排程器為 『step』 時的步驟大小值。 必須是正整數。 整數 (int)
trainingBatchSize 定型批次大小。 必須是正整數。 整數 (int)
trainingCropSize 影像裁剪大小,這是訓練數據集之類神經網路的輸入。 必須是正整數。 整數 (int)
validationBatchSize 驗證批次大小。 必須是正整數。 整數 (int)
validationCropSize 影像裁剪大小,這是用於驗證數據集之類神經網路的輸入。 必須是正整數。 整數 (int)
validationResizeSize 在裁剪驗證數據集之前要重設大小的影像大小。 必須是正整數。 整數 (int)
warmupCosineLRCycles 學習速率排程器為「warmup_cosine」時,餘弦迴圈的值。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
warmupCosineLRWarmupEpochs 學習速率排程器為「warmup_cosine」時,熱身 Epoch 的值。 必須是正整數。 整數 (int)
weightDecay (權重衰減) 優化器為 'zn'、'adam' 或 'adamw' 時的權數衰變值。 必須是範圍[0, 1] 中的浮點數。 整數 (int)
weightedLoss 加權減肥。 接受的值是 0,沒有加權損失。
1 用於與 sqrt 減肥。(class_weights)。 2 用於減肥與class_weights。 必須是 0 或 1 或 2。
整數 (int)

ImageModelSettingsObjectDetection

名稱 說明 價值觀
高級設置 進階案例的設定。 字串
amsGradient 當優化器為 『adam』 或 'adamw' 時啟用AMSGrad。 布爾 (bool)
擴增 使用擴增的設定。 字串
貝塔1 優化器為 'adam' 或 'adamw' 時的 'beta1' 值。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
貝塔2 優化器為 'adam' 或 'adamw' 時的 'beta2' 值。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
boxDetectionsPerImage (每影像檢測) 所有類別的每個影像偵測數目上限。 必須是正整數。
注意:'yolov5' 演算法不支援此設定。
整數 (int)
boxScoreThreshold (盒分數閾值) 在推斷期間,只傳回大於分類分數的建議
BoxScoreThreshold。 必須是範圍[0, 1] 中的浮點數。
整數 (int)
檢查點頻率 儲存模型檢查點的頻率。 必須是正整數。 整數 (int)
checkpoint模型 累加訓練的預先定型檢查點模型。 MLFlowModelJobInput
檢查點RunId 先前執行的標識碼,其具有累加式定型的預先定型檢查點。 字串
分散式 是否要使用分散式定型。 布爾 (bool)
earlyStopping (早期停止) 在定型期間啟用早期停止邏輯。 布爾 (bool)
earlyStoppingDelay 在主要計量改進之前等待的最小 Epoch 或驗證評估數目
會追蹤早期停止。 必須是正整數。
整數 (int)
早點停止耐心 在之前沒有主要計量改進的最小 Epoch 或驗證評估數目
執行已停止。 必須是正整數。
整數 (int)
enableOnnxNormalization 匯出 ONNX 模型時啟用正規化。 布爾 (bool)
evaluationFrequency 評估驗證數據集以取得計量分數的頻率。 必須是正整數。 整數 (int)
梯度累積步驟 漸層累積表示執行未設定數目的 「GradAccumulationStep」 步驟
更新模型權數,同時累積這些步驟的漸層,然後使用
用來計算權數更新的累積漸層。 必須是正整數。
整數 (int)
影像大小 定型和驗證的影像大小。 必須是正整數。
注意:如果大小太大,定型回合可能會進入 CUDA OOM。
注意:只有 'yolov5' 演算法才支援此設定。
整數 (int)
layersToFreeze (圖層凍結) 要凍結模型的圖層數目。 必須是正整數。
例如,傳遞 2 作為 'seresnext' 值表示
凍結 layer0 和 layer1。 如需支援模型的完整清單,以及圖層凍結的詳細數據,請
請參閱:/azure/machine-learning/how-to-auto-train-image-models。
整數 (int)
learningRate 初始學習速率。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
learningRateScheduler 的 學習率排程器的類型。 必須是 『warmup_cosine』 或 'step'。 “無”
“步驟”
'WarmupCosine'
最大尺寸 在將影像饋送至骨幹之前,要重新調整的影像大小上限。
必須是正整數。 注意:如果大小太大,定型執行可能會進入 CUDA OOM。
注意:'yolov5' 演算法不支援此設定。
整數 (int)
最小大小 將影像饋送至骨幹之前,要重新調整的影像大小下限。
必須是正整數。 注意:如果大小太大,定型執行可能會進入 CUDA OOM。
注意:'yolov5' 演算法不支援此設定。
整數 (int)
型號名稱 要用於定型的模型名稱。
如需可用模型的詳細資訊,請流覽官方檔:
/azure/machine-learning/how-to-auto-train-image-models。
字串
modelSize 模型大小。 必須是 'small'、'medium'、'large' 或 'xlarge'。
注意:如果模型大小太大,定型執行可能會進入 CUDA OOM。
注意:只有 'yolov5' 演算法才支援此設定。
“超大號”
“大”
'中等'
“無”
“小”
動量 優化器為 『zn』 時的動量值。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
多尺度 依 +/- 50%來啟用多重縮放影像大小。
注意:如果沒有足夠的 GPU 記憶體,定型執行可能會進入 CUDA OOM。
注意:只有 'yolov5' 演算法才支援此設定。
布爾 (bool)
涅斯特羅夫 在優化器為 『zn』 時啟用 nesterov。 布爾 (bool)
nmsIouThreshold 在NMS後置處理中的推斷期間使用的IOU臨界值。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
紀元數 定型 Epoch 的數目。 必須是正整數。 整數 (int)
工人數 數據載入器背景工作角色的數目。 必須是非負整數。 整數 (int)
優化器 優化工具的類型。 “亞當”
“亞當”
“無”
'新元'
隨機種子 使用決定性定型時要使用的隨機種子。 整數 (int)
stepLR伽瑪 學習速率排程器為 『step』 時的 gamma 值。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
stepLRStepSize 學習速率排程器為 『step』 時的步驟大小值。 必須是正整數。 整數 (int)
tileGridSize (平鋪網格大小) 用來貼上每個影像的網格線大小。 注意:TileGridSize 不得為
無可啟用小型物件偵測邏輯。 包含兩個整數的字串,其格式為一個。
注意:'yolov5' 演算法不支援此設定。
字串
tileOverlapRatio (平鋪重疊比率) 每個維度中相鄰磚之間的重疊比例。 必須是範圍 [0, 1] 中的浮點數。
注意:'yolov5' 演算法不支援此設定。
整數 (int)
tilePredictionsNmsThreshold 用來執行NMS的IOU閾值,同時合併磚和影像的預測。
用於驗證/推斷。 必須是範圍 [0, 1] 中的 float。
注意:'yolov5' 演算法不支援此設定。
整數 (int)
trainingBatchSize 定型批次大小。 必須是正整數。 整數 (int)
validationBatchSize 驗證批次大小。 必須是正整數。 整數 (int)
validationIouThreshold 計算驗證計量時要使用的IOU閾值。 必須是範圍 [0, 1] 中的 float。 整數 (int)
validationMetricType 用於驗證計量的計量計算方法。 《尋夢環遊記》
“可哥沃克”
“無”
'Voc'
warmupCosineLRCycles 學習速率排程器為「warmup_cosine」時,餘弦迴圈的值。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
warmupCosineLRWarmupEpochs 學習速率排程器為「warmup_cosine」時,熱身 Epoch 的值。 必須是正整數。 整數 (int)
weightDecay (權重衰減) 優化器為 'zn'、'adam' 或 'adamw' 時的權數衰變值。 必須是範圍[0, 1] 中的浮點數。 整數 (int)

ImageObjectDetection

名稱 說明 價值觀
limitSettings [必要]限制 AutoML 作業的設定。 ImageLimitSettings (必要)
modelSettings 用於定型模型的設定。 ImageModelSettingsObjectDetection
primaryMetric 要針對這項工作進行優化的主要計量。 'MeanAveragePrecision'
搜索空間 搜尋空間以取樣模型及其超參數的不同組合。 ImageModelDistributionSettingsObjectDetection[]
掃描設置 模型掃掠和超參數掃掠相關設定。 ImageSweepSettings (圖像掃描設定)
任務類型 [必要]AutoMLJob 的工作類型。 'ImageObjectDetection' (必要)
validationData (驗證資料) 驗證數據輸入。 MLTableJobInput
validationDataSize 需要為驗證目的預留的定型數據集分數。
介於 (0.0 、 1.0) 之間的值
未提供驗證數據集時套用。
整數 (int)

ImageSweepSettings (圖像掃描設定)

名稱 說明 價值觀
early終止 早期終止原則的類型。 提前終止政策
sampling演算法 [必要]超參數取樣演算法的類型。 “貝氏”
'網格'
'Random' (必要)

JobBaseProperties

名稱 說明 價值觀
元件ID 元件資源的 ARM 資源識別碼。 字串
計算ID 計算資源的 ARM 資源識別碼。 字串
說明 資產描述文字。 字串
顯示名稱 工作的顯示名稱。 字串
實驗名稱 作業所屬的實驗名稱。 如果未設定,作業會放在 「默認」實驗中。 字串
身分識別 身分識別設定。 如果設定,這應該是其中一個 AmlToken、ManagedIdentity、UserIdentity 或 null。
如果為 null,則預設為 AmlToken。
身份配置
isArchived 已存檔 資產是否已封存? 布爾 (bool)
職位類型 設定為 AutoMLJob 類型的 『AutoML』。 設定為 CommandJob 類型的 『Command』。 設定為 FineTuningJob 類型的 『FineTuning』。 設定為 PipelineJob 類型的 'Pipeline'。 針對 SparkJob 類型設定為 『Spark』。 設定為 掃掠Job 類型的 [掃掠]。 'AutoML'
“命令”
“微調”
“管道”
“火花”
'掃掠' (必要)
notification設置 作業的通知設定 通知設置
屬性 資產屬性字典。 ResourceBaseProperties
服務 JobEndpoints 的清單。
針對本機作業,作業端點會有 FileStreamObject 的端點值。
JobBase服務
標記 標記字典。 標記可以新增、移除和更新。 ResourceBaseTags

JobBase服務

名稱 說明 價值觀

作業輸入

名稱 說明 價值觀
說明 輸入的描述。 字串
jobInputType (作業輸入類型) 設定為 CustomModelJobInput 類型的 『custom_model』。 設定為 常值JobInput 類型的 『literal』。 設定為 MLFlowModelJobInput 類型的 『mlflow_model』。 設定為 MLTableJobInput 類型的 'mltable'。 設定為 TritonModelJobInput 類型的 『triton_model』。 針對 UriFileJobInput類型設定為 『uri_file』。 針對 UriFolderJobInput 類型設定為 『uri_folder』。 “custom_model”
'字面值'
“mlflow_model”
'mltable'
“triton_model”
“uri_file”
'uri_folder' (必要)

JobOutput (作業輸出)

名稱 說明 價值觀
說明 輸出的描述。 字串
jobOutputType 設定為 CustomModelJobOutput 類型的 『custom_model』。 設定為 MLFlowModelJobOutput 類型的 『mlflow_model』。 設定為 MLTableJobOutput 類型的 'mltable'。 將 設定為 TritonModelJobOutput 類型的 『triton_model』。 針對 UriFileJobOutput類型設定為 『uri_file』。 針對 UriFolderJobOutput類型設定為 『uri_folder』。 “custom_model”
“mlflow_model”
'mltable'
“triton_model”
“uri_file”
'uri_folder' (必要)

JobResourceConfiguration

名稱 說明 價值觀
dockerArgs 傳遞至 Docker run 命令的額外自變數。 這會覆寫系統或本節中已設定的任何參數。 此參數僅支援 Azure ML 計算類型。 字串
實例數量 計算目標所使用的實例或節點選擇性數目。 整數 (int)
instanceType 計算目標所支持的選擇性 VM 類型。 字串
屬性 其他屬性包。 ResourceConfigurationProperties
shmSize Docker 容器的共用記憶體區塊大小。 這應該是 (number)(unit) 的格式,其中數位大於 0,單位可以是 b(位元組)、k(KB)、m(MB)、或 g(GB)。 字串

約束:
圖案 = \d+[bBkKmMgG]

工作資源

名稱 說明 價值觀
instanceTypes 要從中選擇的實例類型清單。 字串[]

工作服務

名稱 說明 價值觀
端點 端點的 URL。 字串
jobServiceType 端點類型。 字串
節點 使用者想要啟動服務的節點。
如果未將 Nodes 設定或設定為 null,則服務只會在領導者節點上啟動。
節點
傳輸埠 端點的埠。 整數 (int)
屬性 在端點上設定的其他屬性。 JobServiceProperties

JobServiceProperties

名稱 說明 價值觀

LiteralJobInput (文字作業輸入)

名稱 說明 價值觀
jobInputType (作業輸入類型) [必要]指定作業的類型。 'literal' (必要)
價值 [必要]輸入的常值。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)

ManagedIdentity (託管身份)

名稱 說明 價值觀
用戶端ID 依用戶端識別元指定使用者指派的身分識別。 若為系統指派,請勿設定此欄位。 字串

約束:
最小長度 = 36
最大長度 = 36
圖案 = ^[0-9a-fA-F]{8}-([0-9a-fA-F]{4}-){3}[0-9a-fA-F]{12}$
身份類型 [必要]指定身分識別架構的類型。 'Managed' (必要)
物件ID 依物件識別元指定使用者指派的身分識別。 若為系統指派,請勿設定此欄位。 字串

約束:
最小長度 = 36
最大長度 = 36
圖案 = ^[0-9a-fA-F]{8}-([0-9a-fA-F]{4}-){3}[0-9a-fA-F]{12}$
資源ID 依 ARM 資源識別碼指定使用者指派的身分識別。 若為系統指派,請勿設定此欄位。 字串

中位數停止政策

名稱 說明 價值觀
policyType [必要]原則設定的名稱 'MedianStopping' (必要)

MLFlowModelJobInput

名稱 說明 價值觀
說明 輸入的描述。 字串
jobInputType (作業輸入類型) [必要]指定作業的類型。 “custom_model”
'字面值'
“mlflow_model”
'mltable'
“triton_model”
“uri_file”
'uri_folder' (必要)
模式 輸入資產傳遞模式。 “直接”
'下載'
'EvalDownload'
'EvalMount'
'ReadOnlyMount'
“ReadWriteMount”
統一資源識別碼 (URI) [必要]輸入資產 URI。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)

MLFlowModelJobInput

名稱 說明 價值觀
jobInputType (作業輸入類型) [必要]指定作業的類型。 'mlflow_model' (必要)
模式 輸入資產傳遞模式。 “直接”
'下載'
'EvalDownload'
'EvalMount'
'ReadOnlyMount'
“ReadWriteMount”
統一資源識別碼 (URI) [必要]輸入資產 URI。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)

MLFlowModelJobOutput

名稱 說明 價值觀
jobOutputType [必要]指定作業的類型。 'mlflow_model' (必要)
模式 輸出資產傳遞模式。 “直接”
“ReadWriteMount”
'上傳'
統一資源識別碼 (URI) 輸出資產 URI。 字串

MLTableJobInput

名稱 說明 價值觀
說明 輸入的描述。 字串
jobInputType (作業輸入類型) [必要]指定作業的類型。 “custom_model”
'字面值'
“mlflow_model”
'mltable'
“triton_model”
“uri_file”
'uri_folder' (必要)
模式 輸入資產傳遞模式。 “直接”
'下載'
'EvalDownload'
'EvalMount'
'ReadOnlyMount'
“ReadWriteMount”
統一資源識別碼 (URI) [必要]輸入資產 URI。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)

MLTableJobInput

名稱 說明 價值觀
jobInputType (作業輸入類型) [必要]指定作業的類型。 'mltable' (必要)
模式 輸入資產傳遞模式。 “直接”
'下載'
'EvalDownload'
'EvalMount'
'ReadOnlyMount'
“ReadWriteMount”
統一資源識別碼 (URI) [必要]輸入資產 URI。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)

MLTableJobOutput

名稱 說明 價值觀
jobOutputType [必要]指定作業的類型。 'mltable' (必要)
模式 輸出資產傳遞模式。 “直接”
“ReadWriteMount”
'上傳'
統一資源識別碼 (URI) 輸出資產 URI。 字串

MPI

名稱 說明 價值觀
distributionType [必要]指定散發架構的類型。 'Mpi' (必要)
processCountPerInstance 每個 MPI 節點的進程數目。 整數 (int)

NCross驗證

名稱 說明 價值觀
模式 針對 AutoNCrossValidations 類型,設定為 『Auto』。 設定為 CustomNCrossValidations 類型的 [自定義]。 '自動'
'Custom' (必要)

NlpVerticalFeaturizationSettings

名稱 說明 價值觀
datasetLanguage 數據集語言,適用於文字數據。 字串

NlpVerticalLimitSettings

名稱 說明 價值觀
maxConcurrentTrials 並行 AutoML 反覆專案上限。 整數 (int)
maxTrials AutoML 反覆項目的數目。 整數 (int)
超時 AutoML 作業逾時。 字串

節點

名稱 說明 價值觀
nodesValueType 針對 AllNodes 類型,設定為 'All'。 'All' (必要)

通知設置

名稱 說明 價值觀
電子郵件 在指定的通知類型上傳送電子郵件通知給使用者 包含任何的字串數組:
'JobCancelled' (工作已取消)
'JobCompleted'
“作業失敗”
電子郵件 這是電子郵件收件者清單,其總限製為499個字元,且具有逗號分隔符 字串[]
webhook 將 Webhook 回呼傳送至服務。 機碼是 Webhook 的使用者提供名稱。 NotificationSettingWebhook

NotificationSettingWebhook

名稱 說明 價值觀

目的

名稱 說明 價值觀
目標 [必要]定義超參數微調支援的計量目標 “最大化”
'最小化' (必要)
primaryMetric [必要]要優化之計量的名稱。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)

管道作業

名稱 說明 價值觀
輸入 管線作業的輸入。 管道作業輸入
工作 作業會建構管線作業。 PipelineJobJob 作業
職位類型 [必要]指定作業的類型。 'Pipeline' (必要)
輸出 管線作業的輸出 PipelineJobOutputs
設定 管線設定,適用於 ContinueRunOnStepFailure 等專案。 任何
源作業 ID 來源作業的 ARM 資源識別碼。 字串

管道作業輸入

名稱 說明 價值觀

PipelineJobJob 作業

名稱 說明 價值觀

PipelineJobOutputs

名稱 說明 價值觀

PyTorch

名稱 說明 價值觀
distributionType [必要]指定散發架構的類型。 'PyTorch' (必要)
processCountPerInstance 每個節點的進程數目。 整數 (int)

佇列設置

名稱 說明 價值觀
jobTier 控制計算作業層 “基本”
'Null'
“高級”
'Spot'
“標準”

RandomSamplingAlgorithm 演算法

名稱 說明 價值觀
規則 隨機演算法的特定類型 '隨機'
“索博爾”
採樣演算法類型 [必要]用來產生超參數值的演算法,以及組態屬性 'Random' (必要)
種子 要作為隨機數產生種子的選擇性整數 整數 (int)

迴歸

名稱 說明 價值觀
cvSplitColumnNames 要用於 CVSplit 資料的數據行。 字串[]
featurizationSettings AutoML 作業所需的特徵化輸入。 TableVerticalFeaturizationSettings
limitSettings AutoMLJob 的執行條件約束。 TableVerticalLimitSettings
nCrossValidations 要套用至定型數據集的交叉驗證折疊數目
未提供驗證數據集時。
NCross驗證
primaryMetric 回歸工作的主要計量。 'NormalizedMeanAbsoluteError'
'NormalizedRootMeanSquaredError'
'R2Score'
'SpearmanCorrelation'
任務類型 [必要]AutoMLJob 的工作類型。 '回歸' (必要)
測試數據 測試數據輸入。 MLTableJobInput
testDataSize 必須為驗證目的而保留的測試數據集分數。
介於 (0.0 、 1.0) 之間的值
未提供驗證數據集時套用。
整數 (int)
trainingSettings AutoML 作業定型階段的輸入。 回歸訓練設置
validationData (驗證資料) 驗證數據輸入。 MLTableJobInput
validationDataSize 需要為驗證目的預留的定型數據集分數。
介於 (0.0 、 1.0) 之間的值
未提供驗證數據集時套用。
整數 (int)
權重列名稱 範例權數數據行的名稱。 自動化 ML 支援加權數據行做為輸入,導致數據中的數據列向上或向下加權。 字串

回歸訓練設置

名稱 說明 價值觀
allowedTrainingAlgorithms 回歸工作的允許模型。 包含任何的字串數組:
“決策樹”
“彈性網”
'ExtremeRandomTrees'
'梯度提升'
'KNN'
“套索拉斯”
'LightGBM'
'隨機森林'
'新幣'
'XGBoostRegressor'
blockedTrainingAlgorithms 回歸工作的封鎖模型。 包含任何的字串數組:
“決策樹”
“彈性網”
'ExtremeRandomTrees'
'梯度提升'
'KNN'
“套索拉斯”
'LightGBM'
'隨機森林'
'新幣'
'XGBoostRegressor'
enableDnnTraining 啟用 DNN 模型的建議。 布爾 (bool)
enableModelExplainability 在最佳模型上開啟可解釋性的旗標。 布爾 (bool)
enableOnnxCompatibleModels 啟用 onnx 相容模型的旗標。 布爾 (bool)
enableStackEnsemble 啟用堆疊合奏執行。 布爾 (bool)
enableVoteEnsemble 啟用投票合奏執行。 布爾 (bool)
ensembleModelDownloadTimeout 在 VotingEnsemble 和 StackEnsemble 模型產生期間,會下載先前子回合中的多個安裝模型。
如果需要更多時間,請使用高於 300 秒的值來設定此參數。
字串
stackEnsembleSettings 堆疊合奏執行的堆疊合奏設定。 StackEnsembleSettings

ResourceBaseProperties

名稱 說明 價值觀

ResourceBaseTags

名稱 說明 價值觀

ResourceConfigurationProperties

名稱 說明 價值觀

採樣演算法

名稱 說明 價值觀
採樣演算法類型 針對 BayesianSamplingAlgorithm類型設定為 『Bayesian』。 設定為 GridSamplingAlgorithm 類型的 『Grid』。 設定為 RandomSamplingAlgorithm 類型的 『Random』。 “貝氏”
'網格'
'Random' (必要)

季節性

名稱 說明 價值觀
模式 針對 AutoSeasonality 類型,設定為 [自動]。 針對 CustomSeasonality 類型,設定為 [自定義]。 '自動'
'Custom' (必要)

SparkJob

名稱 說明 價值觀
檔案 封存作業中使用的檔案。 字串[]
參數 作業的自變數。 字串
代碼Id [必要] 程式代碼資產的arm識別碼。 字串 (必要)
會議 Spark 設定的屬性。 SparkJobConf 函數
輸入 [必要]在作業啟動時要執行的專案。 SparkJobEntry (必要)
環境ID 作業之環境規格的 ARM 資源識別碼。 字串
環境變數 作業中包含的環境變數。 SparkJobEnvironment變數
檔案 作業中使用的檔案。 字串[]
輸入 對應作業中使用的輸入數據系結。 SparkJobInputs
罐子 作業中使用的 Jar 檔案。 字串[]
職位類型 [必要]指定作業的類型。 'Spark' (必要)
輸出 對應作業中使用的輸出數據系結。 SparkJob輸出
py檔 作業中使用的 Python 檔案。 字串[]
queueSettings 作業的佇列設定 佇列設置
資源 作業的計算資源組態。 SparkResourceConfiguration

SparkJobConf 函數

名稱 說明 價值觀

SparkJobEntry

名稱 說明 價值觀
sparkJobEntryType 針對 SparkJobPythonEntry 類型,設定為 'SparkJobPythonEntry'。 針對 SparkJobScalaEntry 類型,設定為 'SparkJobScalaEntry'。 “SparkJobPythonEntry”
'SparkJobScalaEntry' (必要)

SparkJobEnvironment變數

名稱 說明 價值觀

SparkJobInputs

名稱 說明 價值觀

SparkJob輸出

名稱 說明 價值觀

SparkJobPython條目

名稱 說明 價值觀
檔案 [必要]作業進入點的相對 Python 檔案路徑。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)
sparkJobEntryType [必要]作業進入點的類型。 'SparkJobPythonEntry' (必要)

SparkJobScalaEntry

名稱 說明 價值觀
類別名稱 [必要]做為進入點的 Scala 類別名稱。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)
sparkJobEntryType [必要]作業進入點的類型。 'SparkJobScalaEntry' (必要)

SparkResourceConfiguration

名稱 說明 價值觀
instanceType 計算目標所支持的選擇性 VM 類型。 字串
runtimeVersion (運行時版本) 用於作業的 Spark 執行時間版本。 字串

StackEnsembleSettings

名稱 說明 價值觀
stackMetaLearnerKWargs 要傳遞至中繼學習模組初始化表達式的選擇性參數。 任何
stackMetaLearnerTrainPercentage 指定要保留用於訓練中繼學習工具的定型集比例(選擇定型和定型類型的訓練類型)。 預設值為 0.2。 整數 (int)
stackMetaLearnerType 中繼學習器是針對個別異質模型輸出定型的模型。 “彈性網”
'ElasticNetCV'
'LightGBMClassifier'
'LightGBMRegressor'
'線性回歸'
'LogisticRegression' (邏輯回歸)
'LogisticRegressionCV'
“無”

SweepJob (掃描作業)

名稱 說明 價值觀
early終止 早期終止原則可在完成之前取消執行效能不佳 提前終止政策
輸入 對應作業中使用的輸入數據系結。 SweepJobInputs
職位類型 [必要]指定作業的類型。 '掃掠' (必要)
限制 掃掠作業限制。 SweepJobLimits (掃描作業限制)
目的 [必要]優化目標。 目標 (必要)
輸出 對應作業中使用的輸出數據系結。 SweepJobOutputs (掃描作業輸出)
queueSettings 作業的佇列設定 佇列設置
sampling演算法 [必要]超參數取樣演算法 SamplingAlgorithm (必要)
搜索空間 [必要]包含每個參數及其散發的字典。 字典索引鍵是參數的名稱 任何 (必要)
試用 [必要]試用版元件定義。 TrialComponent (必要)

SweepJobInputs

名稱 說明 價值觀

SweepJobLimits (掃描作業限制)

名稱 說明 價值觀
jobLimits類型 [必要]JobLimit 類型。 “命令”
'掃掠' (必要)
maxConcurrentTrials 掃掠作業最大並行試用版。 整數 (int)
maxTotalTrials 掃掠作業最大總試用版。 整數 (int)
超時 ISO 8601 格式的最大執行持續時間,之後作業將會取消。 僅支援有效位數為秒數的持續時間。 字串
trialTimeout (試用超時) 掃掠作業試用版逾時值。 字串

SweepJobOutputs (掃描作業輸出)

名稱 說明 價值觀

TableVerticalFeaturizationSettings

名稱 說明 價值觀
阻塞變壓器 這些轉換器不得用於特徵化。 包含任何的字串數組:
'CatTargetEncoder' (貓目標編碼器)
'CountVectorizer'
'HashOneHotEncoder'
'標籤編碼器'
“天真貝葉斯”
'OneHotEncoder' (一熱編碼器)
'TextTargetEncoder'
'TfIdf'
'WoETargetEncoder' (WoETarget編碼器)
'WordEmbedding' (單詞嵌入)
columnNameAnd類型 數據行名稱及其類型字典(int、float、string、datetime 等)。 TableVerticalFeaturizationSettingsColumnNameAndTypes
datasetLanguage 數據集語言,適用於文字數據。 字串
enableDnnFeaturization 判斷是否要使用 Dnn 型特徵化程式進行數據特徵化。 布爾 (bool)
模式 特徵化模式 - 使用者可以保留預設的「自動」模式,而 AutoML 會負責在特徵化階段中轉換數據的必要轉換。
如果選取 [關閉],則不會完成特徵化。
如果選取 [自定義],則使用者可以指定其他輸入,以自定義特徵化的完成方式。
'自動'
“自定義”
“關閉”
transformer參數 用戶可以指定要搭配套用的其他轉換器,以及要套用它的數據行,以及轉換建構函式的參數。 TableVerticalFeaturizationSettingsTransformerParams

TableVerticalFeaturizationSettingsColumnNameAndTypes

名稱 說明 價值觀

TableVerticalFeaturizationSettingsTransformerParams

名稱 說明 價值觀

TableVerticalLimitSettings

名稱 說明 價值觀
enableEarlyTermination 啟用提早終止,判斷如果過去 20 個反覆項目沒有分數改善,AutoMLJob 是否會提早終止。 布爾 (bool)
退出評分 AutoML 作業的結束分數。 整數 (int)
maxConcurrentTrials 並行反覆運算數上限。 整數 (int)
maxCoresPerTrial 每個反覆專案的核心數上限。 整數 (int)
maxTrials 反覆項目的數目。 整數 (int)
超時 AutoML 作業逾時。 字串
trialTimeout (試用超時) 反覆專案逾時。 字串

目標滯後

名稱 說明 價值觀
模式 針對 AutoTargetLags 類型設定為 『Auto』。 設定為 CustomTargetLags 類型的 [自定義]。 '自動'
'Custom' (必要)

TargetRollingWindowSize

名稱 說明 價值觀
模式 針對 AutoTargetRollingWindowSize 類型,設定為 『Auto』。 針對 CustomTargetRollingWindowSize 類型設定為 『Custom』。 '自動'
'Custom' (必要)

TensorFlow

名稱 說明 價值觀
distributionType [必要]指定散發架構的類型。 'TensorFlow' (必要)
parameterServerCount 參數伺服器工作的數目。 整數 (int)
workerCount 工人計數 工人人數。 如果未指定,則會預設為實例計數。 整數 (int)

文本分類

名稱 說明 價值觀
featurizationSettings AutoML 作業所需的特徵化輸入。 NlpVerticalFeaturizationSettings
limitSettings AutoMLJob 的執行條件約束。 NlpVerticalLimitSettings
primaryMetric Text-Classification 工作的主要計量。 '精確度'
'AUCWeighted'
'AveragePrecisionScoreWeighted'
'NormMacroRecall' (規範巨集調用)
'PrecisionScoreWeighted'
任務類型 [必要]AutoMLJob 的工作類型。 'TextClassification' (必要)
validationData (驗證資料) 驗證數據輸入。 MLTableJobInput

TextClassificationMultilabel

名稱 說明 價值觀
featurizationSettings AutoML 作業所需的特徵化輸入。 NlpVerticalFeaturizationSettings
limitSettings AutoMLJob 的執行條件約束。 NlpVerticalLimitSettings
任務類型 [必要]AutoMLJob 的工作類型。 'TextClassificationMultilabel' (必要)
validationData (驗證資料) 驗證數據輸入。 MLTableJobInput

TextNer (文字產生器)

名稱 說明 價值觀
featurizationSettings AutoML 作業所需的特徵化輸入。 NlpVerticalFeaturizationSettings
limitSettings AutoMLJob 的執行條件約束。 NlpVerticalLimitSettings
任務類型 [必要]AutoMLJob 的工作類型。 'TextNER' (必要)
validationData (驗證資料) 驗證數據輸入。 MLTableJobInput

TrialComponent (試用元件)

名稱 說明 價值觀
代碼Id 程式代碼資產的 ARM 資源識別碼。 字串
指令 [必要]在作業啟動時執行的命令。 例如 “python train.py” 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)
分配 作業的散發組態。 如果設定,這應該是其中一個 Mpi、Tensorflow、PyTorch 或 null。 分發配置
環境ID [必要]作業之環境規格的 ARM 資源識別碼。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)
環境變數 作業中包含的環境變數。 TrialComponentEnvironmentVariables
資源 作業的計算資源組態。 JobResourceConfiguration

TrialComponentEnvironmentVariables

名稱 說明 價值觀

TritonModelJobInput

名稱 說明 價值觀
jobInputType (作業輸入類型) [必要]指定作業的類型。 'triton_model' (必要)
模式 輸入資產傳遞模式。 “直接”
'下載'
'EvalDownload'
'EvalMount'
'ReadOnlyMount'
“ReadWriteMount”
統一資源識別碼 (URI) [必要]輸入資產 URI。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)

TritonModelJobOutput

名稱 說明 價值觀
jobOutputType [必要]指定作業的類型。 'triton_model' (必要)
模式 輸出資產傳遞模式。 “直接”
“ReadWriteMount”
'上傳'
統一資源識別碼 (URI) 輸出資產 URI。 字串

截斷選擇策略 (TruncationSelectionPolicy)

名稱 說明 價值觀
policyType [必要]原則設定的名稱 'TruncationSelection' (必要)
截斷百分比 要在每個評估間隔取消的執行百分比。 整數 (int)

UriFileJobInput

名稱 說明 價值觀
jobInputType (作業輸入類型) [必要]指定作業的類型。 'uri_file' (必要)
模式 輸入資產傳遞模式。 “直接”
'下載'
'EvalDownload'
'EvalMount'
'ReadOnlyMount'
“ReadWriteMount”
統一資源識別碼 (URI) [必要]輸入資產 URI。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)

UriFileJobOutput

名稱 說明 價值觀
jobOutputType [必要]指定作業的類型。 'uri_file' (必要)
模式 輸出資產傳遞模式。 “直接”
“ReadWriteMount”
'上傳'
統一資源識別碼 (URI) 輸出資產 URI。 字串

UriFolderJobInput

名稱 說明 價值觀
jobInputType (作業輸入類型) [必要]指定作業的類型。 'uri_folder' (必要)
模式 輸入資產傳遞模式。 “直接”
'下載'
'EvalDownload'
'EvalMount'
'ReadOnlyMount'
“ReadWriteMount”
統一資源識別碼 (URI) [必要]輸入資產 URI。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)

UriFolderJobOutput

名稱 說明 價值觀
jobOutputType [必要]指定作業的類型。 'uri_folder' (必要)
模式 輸出資產傳遞模式。 “直接”
“ReadWriteMount”
'上傳'
統一資源識別碼 (URI) 輸出資產 URI。 字串

使用者身份

名稱 說明 價值觀
身份類型 [必要]指定身分識別架構的類型。 'UserIdentity' (必要)

Webhook

名稱 說明 價值觀
事件類型 在指定的通知事件上傳送回呼 字串
webhook類型 針對 AzureDevOpsWebhook 類型,設定為 『AzureDevOps』。 'AzureDevOps' (必要)

使用範例

Azure 快速入門範本

下列 Azure 快速入門範本 會部署此資源類型。

範本 說明
建立 Azure Machine Learning AutoML 分類作業

部署至 Azure
此範本會建立 Azure Machine Learning AutoML 分類作業,以找出最佳模型,以預測客戶是否會訂閱與金融機構的固定期存款。
建立 Azure Machine Learning 命令作業

部署至 Azure
此範本會建立具有基本hello_world腳本的 Azure Machine Learning 命令作業
建立 Azure Machine Learning 掃掠作業

部署至 Azure
此範本會建立 Azure Machine Learning 掃掠作業以進行超參數微調。

Terraform (AzAPI 提供者) 資源定義

工作區/作業資源類型可以使用目標作業來部署:

  • 資源群組

如需每個 API 版本中已變更屬性的清單,請參閱 變更記錄檔。

資源格式

若要建立 Microsoft.MachineLearningServices/workspaces/jobs 資源,請將下列 Terraform 新增至範本。

resource "azapi_resource" "symbolicname" {
  type = "Microsoft.MachineLearningServices/workspaces/jobs@2024-07-01-preview"
  name = "string"
  parent_id = "string"
  body = {
    properties = {
      componentId = "string"
      computeId = "string"
      description = "string"
      displayName = "string"
      experimentName = "string"
      identity = {
        identityType = "string"
        // For remaining properties, see IdentityConfiguration objects
      }
      isArchived = bool
      notificationSetting = {
        emailOn = [
          "string"
        ]
        emails = [
          "string"
        ]
        webhooks = {
          {customized property} = {
            eventType = "string"
            webhookType = "string"
            // For remaining properties, see Webhook objects
          }
        }
      }
      properties = {
        {customized property} = "string"
      }
      services = {
        {customized property} = {
          endpoint = "string"
          jobServiceType = "string"
          nodes = {
            nodesValueType = "string"
            // For remaining properties, see Nodes objects
          }
          port = int
          properties = {
            {customized property} = "string"
          }
        }
      }
      tags = {
        {customized property} = "string"
      }
      jobType = "string"
      // For remaining properties, see JobBaseProperties objects
    }
  }
}

FineTuningVertical 物件

設定 modelProvider 屬性以指定物件的類型。

針對 AzureOpenAI,請使用:

{
  hyperParameters = {
    batchSize = int
    learningRateMultiplier = int
    nEpochs = int
  }
  modelProvider = "AzureOpenAI"
}

針對 [自定義],請使用:

{
  hyperParameters = {
    {customized property} = "string"
  }
  modelProvider = "Custom"
}

TargetRollingWindowSize 物件

設定 mode 屬性以指定物件的類型。

針對 [自動],請使用:

{
  mode = "Auto"
}

針對 [自定義],請使用:

{
  mode = "Custom"
  value = int
}

JobInput 物件

設定 jobInputType 屬性以指定物件的類型。

針對 custom_model,請使用:

{
  jobInputType = "custom_model"
  mode = "string"
  uri = "string"
}

針對 常值,請使用:

{
  jobInputType = "literal"
  value = "string"
}

針對 mlflow_model,請使用:

{
  jobInputType = "mlflow_model"
  mode = "string"
  uri = "string"
}

針對 mltable,請使用:

{
  jobInputType = "mltable"
  mode = "string"
  uri = "string"
}

針對 triton_model,請使用:

{
  jobInputType = "triton_model"
  mode = "string"
  uri = "string"
}

針對 uri_file,請使用:

{
  jobInputType = "uri_file"
  mode = "string"
  uri = "string"
}

針對 uri_folder,請使用:

{
  jobInputType = "uri_folder"
  mode = "string"
  uri = "string"
}

JobBaseProperties 物件

設定 jobType 屬性以指定物件的類型。

針對 AutoML,請使用:

{
  environmentId = "string"
  environmentVariables = {
    {customized property} = "string"
  }
  jobType = "AutoML"
  outputs = {
    {customized property} = {
      description = "string"
      jobOutputType = "string"
      // For remaining properties, see JobOutput objects
    }
  }
  queueSettings = {
    jobTier = "string"
  }
  resources = {
    dockerArgs = "string"
    instanceCount = int
    instanceType = "string"
    properties = {
      {customized property} = ?
    }
    shmSize = "string"
  }
  taskDetails = {
    logVerbosity = "string"
    targetColumnName = "string"
    trainingData = {
      description = "string"
      jobInputType = "string"
      mode = "string"
      uri = "string"
    }
    taskType = "string"
    // For remaining properties, see AutoMLVertical objects
  }
}

針對 [命令],請使用:

{
  codeId = "string"
  command = "string"
  distribution = {
    distributionType = "string"
    // For remaining properties, see DistributionConfiguration objects
  }
  environmentId = "string"
  environmentVariables = {
    {customized property} = "string"
  }
  inputs = {
    {customized property} = {
      description = "string"
      jobInputType = "string"
      // For remaining properties, see JobInput objects
    }
  }
  jobType = "Command"
  limits = {
    jobLimitsType = "string"
    timeout = "string"
  }
  outputs = {
    {customized property} = {
      description = "string"
      jobOutputType = "string"
      // For remaining properties, see JobOutput objects
    }
  }
  queueSettings = {
    jobTier = "string"
  }
  resources = {
    dockerArgs = "string"
    instanceCount = int
    instanceType = "string"
    properties = {
      {customized property} = ?
    }
    shmSize = "string"
  }
}

針對 FineTuning,請使用:

{
  fineTuningDetails = {
    model = {
      description = "string"
      jobInputType = "string"
      mode = "string"
      uri = "string"
    }
    taskType = "string"
    trainingData = {
      description = "string"
      jobInputType = "string"
      // For remaining properties, see JobInput objects
    }
    validationData = {
      description = "string"
      jobInputType = "string"
      // For remaining properties, see JobInput objects
    }
    modelProvider = "string"
    // For remaining properties, see FineTuningVertical objects
  }
  jobType = "FineTuning"
  outputs = {
    {customized property} = {
      description = "string"
      jobOutputType = "string"
      // For remaining properties, see JobOutput objects
    }
  }
  queueSettings = {
    jobTier = "string"
  }
  resources = {
    instanceTypes = [
      "string"
    ]
  }
}

針對 管線,請使用:

{
  inputs = {
    {customized property} = {
      description = "string"
      jobInputType = "string"
      // For remaining properties, see JobInput objects
    }
  }
  jobs = {
    {customized property} = ?
  }
  jobType = "Pipeline"
  outputs = {
    {customized property} = {
      description = "string"
      jobOutputType = "string"
      // For remaining properties, see JobOutput objects
    }
  }
  settings = ?
  sourceJobId = "string"
}

針對 Spark,請使用:

{
  archives = [
    "string"
  ]
  args = "string"
  codeId = "string"
  conf = {
    {customized property} = "string"
  }
  entry = {
    sparkJobEntryType = "string"
    // For remaining properties, see SparkJobEntry objects
  }
  environmentId = "string"
  environmentVariables = {
    {customized property} = "string"
  }
  files = [
    "string"
  ]
  inputs = {
    {customized property} = {
      description = "string"
      jobInputType = "string"
      // For remaining properties, see JobInput objects
    }
  }
  jars = [
    "string"
  ]
  jobType = "Spark"
  outputs = {
    {customized property} = {
      description = "string"
      jobOutputType = "string"
      // For remaining properties, see JobOutput objects
    }
  }
  pyFiles = [
    "string"
  ]
  queueSettings = {
    jobTier = "string"
  }
  resources = {
    instanceType = "string"
    runtimeVersion = "string"
  }
}

針對 掃掠,請使用:

{
  earlyTermination = {
    delayEvaluation = int
    evaluationInterval = int
    policyType = "string"
    // For remaining properties, see EarlyTerminationPolicy objects
  }
  inputs = {
    {customized property} = {
      description = "string"
      jobInputType = "string"
      // For remaining properties, see JobInput objects
    }
  }
  jobType = "Sweep"
  limits = {
    jobLimitsType = "string"
    maxConcurrentTrials = int
    maxTotalTrials = int
    timeout = "string"
    trialTimeout = "string"
  }
  objective = {
    goal = "string"
    primaryMetric = "string"
  }
  outputs = {
    {customized property} = {
      description = "string"
      jobOutputType = "string"
      // For remaining properties, see JobOutput objects
    }
  }
  queueSettings = {
    jobTier = "string"
  }
  samplingAlgorithm = {
    samplingAlgorithmType = "string"
    // For remaining properties, see SamplingAlgorithm objects
  }
  searchSpace = ?
  trial = {
    codeId = "string"
    command = "string"
    distribution = {
      distributionType = "string"
      // For remaining properties, see DistributionConfiguration objects
    }
    environmentId = "string"
    environmentVariables = {
      {customized property} = "string"
    }
    resources = {
      dockerArgs = "string"
      instanceCount = int
      instanceType = "string"
      properties = {
        {customized property} = ?
      }
      shmSize = "string"
    }
  }
}

IdentityConfiguration 物件

設定 identityType 屬性以指定物件的類型。

針對 AMLToken,請使用:

{
  identityType = "AMLToken"
}

針對 [受控],請使用:

{
  clientId = "string"
  identityType = "Managed"
  objectId = "string"
  resourceId = "string"
}

針對 UserIdentity,請使用:

{
  identityType = "UserIdentity"
}

ForecastHorizon 物件

設定 mode 屬性以指定物件的類型。

針對 [自動],請使用:

{
  mode = "Auto"
}

針對 [自定義],請使用:

{
  mode = "Custom"
  value = int
}

EarlyTerminationPolicy 物件

設定 policyType 屬性以指定物件的類型。

針對 Bandit,請使用:

{
  policyType = "Bandit"
  slackAmount = int
  slackFactor = int
}

針對 MedianStopping,請使用:

{
  policyType = "MedianStopping"
}

針對 [截斷][選取],請使用:

{
  policyType = "TruncationSelection"
  truncationPercentage = int
}

季節性物件

設定 mode 屬性以指定物件的類型。

針對 [自動],請使用:

{
  mode = "Auto"
}

針對 [自定義],請使用:

{
  mode = "Custom"
  value = int
}

JobOutput 物件

設定 jobOutputType 屬性以指定對象的類型。

針對 custom_model,請使用:

{
  jobOutputType = "custom_model"
  mode = "string"
  uri = "string"
}

針對 mlflow_model,請使用:

{
  jobOutputType = "mlflow_model"
  mode = "string"
  uri = "string"
}

針對 mltable,請使用:

{
  jobOutputType = "mltable"
  mode = "string"
  uri = "string"
}

針對 triton_model,請使用:

{
  jobOutputType = "triton_model"
  mode = "string"
  uri = "string"
}

針對 uri_file,請使用:

{
  jobOutputType = "uri_file"
  mode = "string"
  uri = "string"
}

針對 uri_folder,請使用:

{
  jobOutputType = "uri_folder"
  mode = "string"
  uri = "string"
}

TargetLags 物件

設定 mode 屬性以指定物件的類型。

針對 [自動],請使用:

{
  mode = "Auto"
}

針對 [自定義],請使用:

{
  mode = "Custom"
  values = [
    int
  ]
}

Webhook 物件

設定 webhookType 屬性以指定物件的類型。

針對 AzureDevOps,請使用:

{
  webhookType = "AzureDevOps"
}

SamplingAlgorithm 物件

設定 samplingAlgorithmType 屬性以指定物件的類型。

針對 貝氏,請使用:

{
  samplingAlgorithmType = "Bayesian"
}

針對 Grid,請使用:

{
  samplingAlgorithmType = "Grid"
}

針對 [隨機],請使用:

{
  rule = "string"
  samplingAlgorithmType = "Random"
  seed = int
}

SparkJobEntry 物件

設定 sparkJobEntryType 屬性以指定物件的類型。

針對 SparkJobPythonEntry,請使用:

{
  file = "string"
  sparkJobEntryType = "SparkJobPythonEntry"
}

針對 SparkJobScalaEntry,請使用:

{
  className = "string"
  sparkJobEntryType = "SparkJobScalaEntry"
}

DistributionConfiguration 物件

設定 distributionType 屬性以指定物件的類型。

針對 Mpi,請使用:

{
  distributionType = "Mpi"
  processCountPerInstance = int
}

針對 PyTorch,請使用:

{
  distributionType = "PyTorch"
  processCountPerInstance = int
}

針對 TensorFlow,請使用:

{
  distributionType = "TensorFlow"
  parameterServerCount = int
  workerCount = int
}

NCrossValidations 物件

設定 mode 屬性以指定物件的類型。

針對 [自動],請使用:

{
  mode = "Auto"
}

針對 [自定義],請使用:

{
  mode = "Custom"
  value = int
}

Nodes 物件

設定 nodesValueType 屬性以指定對象的類型。

針對 [全部],請使用:

{
  nodesValueType = "All"
}

AutoMLVertical 物件

設定 taskType 屬性以指定物件的類型。

針對 [分類],請使用:

{
  cvSplitColumnNames = [
    "string"
  ]
  featurizationSettings = {
    blockedTransformers = [
      "string"
    ]
    columnNameAndTypes = {
      {customized property} = "string"
    }
    datasetLanguage = "string"
    enableDnnFeaturization = bool
    mode = "string"
    transformerParams = {
      {customized property} = [
        {
          fields = [
            "string"
          ]
          parameters = ?
        }
      ]
    }
  }
  limitSettings = {
    enableEarlyTermination = bool
    exitScore = int
    maxConcurrentTrials = int
    maxCoresPerTrial = int
    maxTrials = int
    timeout = "string"
    trialTimeout = "string"
  }
  nCrossValidations = {
    mode = "string"
    // For remaining properties, see NCrossValidations objects
  }
  positiveLabel = "string"
  primaryMetric = "string"
  taskType = "Classification"
  testData = {
    description = "string"
    jobInputType = "string"
    mode = "string"
    uri = "string"
  }
  testDataSize = int
  trainingSettings = {
    allowedTrainingAlgorithms = [
      "string"
    ]
    blockedTrainingAlgorithms = [
      "string"
    ]
    enableDnnTraining = bool
    enableModelExplainability = bool
    enableOnnxCompatibleModels = bool
    enableStackEnsemble = bool
    enableVoteEnsemble = bool
    ensembleModelDownloadTimeout = "string"
    stackEnsembleSettings = {
      stackMetaLearnerKWargs = ?
      stackMetaLearnerTrainPercentage = int
      stackMetaLearnerType = "string"
    }
  }
  validationData = {
    description = "string"
    jobInputType = "string"
    mode = "string"
    uri = "string"
  }
  validationDataSize = int
  weightColumnName = "string"
}

針對 預測,請使用:

{
  cvSplitColumnNames = [
    "string"
  ]
  featurizationSettings = {
    blockedTransformers = [
      "string"
    ]
    columnNameAndTypes = {
      {customized property} = "string"
    }
    datasetLanguage = "string"
    enableDnnFeaturization = bool
    mode = "string"
    transformerParams = {
      {customized property} = [
        {
          fields = [
            "string"
          ]
          parameters = ?
        }
      ]
    }
  }
  forecastingSettings = {
    countryOrRegionForHolidays = "string"
    cvStepSize = int
    featureLags = "string"
    forecastHorizon = {
      mode = "string"
      // For remaining properties, see ForecastHorizon objects
    }
    frequency = "string"
    seasonality = {
      mode = "string"
      // For remaining properties, see Seasonality objects
    }
    shortSeriesHandlingConfig = "string"
    targetAggregateFunction = "string"
    targetLags = {
      mode = "string"
      // For remaining properties, see TargetLags objects
    }
    targetRollingWindowSize = {
      mode = "string"
      // For remaining properties, see TargetRollingWindowSize objects
    }
    timeColumnName = "string"
    timeSeriesIdColumnNames = [
      "string"
    ]
    useStl = "string"
  }
  limitSettings = {
    enableEarlyTermination = bool
    exitScore = int
    maxConcurrentTrials = int
    maxCoresPerTrial = int
    maxTrials = int
    timeout = "string"
    trialTimeout = "string"
  }
  nCrossValidations = {
    mode = "string"
    // For remaining properties, see NCrossValidations objects
  }
  primaryMetric = "string"
  taskType = "Forecasting"
  testData = {
    description = "string"
    jobInputType = "string"
    mode = "string"
    uri = "string"
  }
  testDataSize = int
  trainingSettings = {
    allowedTrainingAlgorithms = [
      "string"
    ]
    blockedTrainingAlgorithms = [
      "string"
    ]
    enableDnnTraining = bool
    enableModelExplainability = bool
    enableOnnxCompatibleModels = bool
    enableStackEnsemble = bool
    enableVoteEnsemble = bool
    ensembleModelDownloadTimeout = "string"
    stackEnsembleSettings = {
      stackMetaLearnerKWargs = ?
      stackMetaLearnerTrainPercentage = int
      stackMetaLearnerType = "string"
    }
  }
  validationData = {
    description = "string"
    jobInputType = "string"
    mode = "string"
    uri = "string"
  }
  validationDataSize = int
  weightColumnName = "string"
}

針對 ImageClassification,請使用:

{
  limitSettings = {
    maxConcurrentTrials = int
    maxTrials = int
    timeout = "string"
  }
  modelSettings = {
    advancedSettings = "string"
    amsGradient = bool
    augmentations = "string"
    beta1 = int
    beta2 = int
    checkpointFrequency = int
    checkpointModel = {
      description = "string"
      jobInputType = "string"
      mode = "string"
      uri = "string"
    }
    checkpointRunId = "string"
    distributed = bool
    earlyStopping = bool
    earlyStoppingDelay = int
    earlyStoppingPatience = int
    enableOnnxNormalization = bool
    evaluationFrequency = int
    gradientAccumulationStep = int
    layersToFreeze = int
    learningRate = int
    learningRateScheduler = "string"
    modelName = "string"
    momentum = int
    nesterov = bool
    numberOfEpochs = int
    numberOfWorkers = int
    optimizer = "string"
    randomSeed = int
    stepLRGamma = int
    stepLRStepSize = int
    trainingBatchSize = int
    trainingCropSize = int
    validationBatchSize = int
    validationCropSize = int
    validationResizeSize = int
    warmupCosineLRCycles = int
    warmupCosineLRWarmupEpochs = int
    weightDecay = int
    weightedLoss = int
  }
  primaryMetric = "string"
  searchSpace = [
    {
      amsGradient = "string"
      augmentations = "string"
      beta1 = "string"
      beta2 = "string"
      distributed = "string"
      earlyStopping = "string"
      earlyStoppingDelay = "string"
      earlyStoppingPatience = "string"
      enableOnnxNormalization = "string"
      evaluationFrequency = "string"
      gradientAccumulationStep = "string"
      layersToFreeze = "string"
      learningRate = "string"
      learningRateScheduler = "string"
      modelName = "string"
      momentum = "string"
      nesterov = "string"
      numberOfEpochs = "string"
      numberOfWorkers = "string"
      optimizer = "string"
      randomSeed = "string"
      stepLRGamma = "string"
      stepLRStepSize = "string"
      trainingBatchSize = "string"
      trainingCropSize = "string"
      validationBatchSize = "string"
      validationCropSize = "string"
      validationResizeSize = "string"
      warmupCosineLRCycles = "string"
      warmupCosineLRWarmupEpochs = "string"
      weightDecay = "string"
      weightedLoss = "string"
    }
  ]
  sweepSettings = {
    earlyTermination = {
      delayEvaluation = int
      evaluationInterval = int
      policyType = "string"
      // For remaining properties, see EarlyTerminationPolicy objects
    }
    samplingAlgorithm = "string"
  }
  taskType = "ImageClassification"
  validationData = {
    description = "string"
    jobInputType = "string"
    mode = "string"
    uri = "string"
  }
  validationDataSize = int
}

針對 ImageClassificationMultilabel,請使用:

{
  limitSettings = {
    maxConcurrentTrials = int
    maxTrials = int
    timeout = "string"
  }
  modelSettings = {
    advancedSettings = "string"
    amsGradient = bool
    augmentations = "string"
    beta1 = int
    beta2 = int
    checkpointFrequency = int
    checkpointModel = {
      description = "string"
      jobInputType = "string"
      mode = "string"
      uri = "string"
    }
    checkpointRunId = "string"
    distributed = bool
    earlyStopping = bool
    earlyStoppingDelay = int
    earlyStoppingPatience = int
    enableOnnxNormalization = bool
    evaluationFrequency = int
    gradientAccumulationStep = int
    layersToFreeze = int
    learningRate = int
    learningRateScheduler = "string"
    modelName = "string"
    momentum = int
    nesterov = bool
    numberOfEpochs = int
    numberOfWorkers = int
    optimizer = "string"
    randomSeed = int
    stepLRGamma = int
    stepLRStepSize = int
    trainingBatchSize = int
    trainingCropSize = int
    validationBatchSize = int
    validationCropSize = int
    validationResizeSize = int
    warmupCosineLRCycles = int
    warmupCosineLRWarmupEpochs = int
    weightDecay = int
    weightedLoss = int
  }
  primaryMetric = "string"
  searchSpace = [
    {
      amsGradient = "string"
      augmentations = "string"
      beta1 = "string"
      beta2 = "string"
      distributed = "string"
      earlyStopping = "string"
      earlyStoppingDelay = "string"
      earlyStoppingPatience = "string"
      enableOnnxNormalization = "string"
      evaluationFrequency = "string"
      gradientAccumulationStep = "string"
      layersToFreeze = "string"
      learningRate = "string"
      learningRateScheduler = "string"
      modelName = "string"
      momentum = "string"
      nesterov = "string"
      numberOfEpochs = "string"
      numberOfWorkers = "string"
      optimizer = "string"
      randomSeed = "string"
      stepLRGamma = "string"
      stepLRStepSize = "string"
      trainingBatchSize = "string"
      trainingCropSize = "string"
      validationBatchSize = "string"
      validationCropSize = "string"
      validationResizeSize = "string"
      warmupCosineLRCycles = "string"
      warmupCosineLRWarmupEpochs = "string"
      weightDecay = "string"
      weightedLoss = "string"
    }
  ]
  sweepSettings = {
    earlyTermination = {
      delayEvaluation = int
      evaluationInterval = int
      policyType = "string"
      // For remaining properties, see EarlyTerminationPolicy objects
    }
    samplingAlgorithm = "string"
  }
  taskType = "ImageClassificationMultilabel"
  validationData = {
    description = "string"
    jobInputType = "string"
    mode = "string"
    uri = "string"
  }
  validationDataSize = int
}

針對 ImageInstanceSegmentation,請使用:

{
  limitSettings = {
    maxConcurrentTrials = int
    maxTrials = int
    timeout = "string"
  }
  modelSettings = {
    advancedSettings = "string"
    amsGradient = bool
    augmentations = "string"
    beta1 = int
    beta2 = int
    boxDetectionsPerImage = int
    boxScoreThreshold = int
    checkpointFrequency = int
    checkpointModel = {
      description = "string"
      jobInputType = "string"
      mode = "string"
      uri = "string"
    }
    checkpointRunId = "string"
    distributed = bool
    earlyStopping = bool
    earlyStoppingDelay = int
    earlyStoppingPatience = int
    enableOnnxNormalization = bool
    evaluationFrequency = int
    gradientAccumulationStep = int
    imageSize = int
    layersToFreeze = int
    learningRate = int
    learningRateScheduler = "string"
    maxSize = int
    minSize = int
    modelName = "string"
    modelSize = "string"
    momentum = int
    multiScale = bool
    nesterov = bool
    nmsIouThreshold = int
    numberOfEpochs = int
    numberOfWorkers = int
    optimizer = "string"
    randomSeed = int
    stepLRGamma = int
    stepLRStepSize = int
    tileGridSize = "string"
    tileOverlapRatio = int
    tilePredictionsNmsThreshold = int
    trainingBatchSize = int
    validationBatchSize = int
    validationIouThreshold = int
    validationMetricType = "string"
    warmupCosineLRCycles = int
    warmupCosineLRWarmupEpochs = int
    weightDecay = int
  }
  primaryMetric = "string"
  searchSpace = [
    {
      amsGradient = "string"
      augmentations = "string"
      beta1 = "string"
      beta2 = "string"
      boxDetectionsPerImage = "string"
      boxScoreThreshold = "string"
      distributed = "string"
      earlyStopping = "string"
      earlyStoppingDelay = "string"
      earlyStoppingPatience = "string"
      enableOnnxNormalization = "string"
      evaluationFrequency = "string"
      gradientAccumulationStep = "string"
      imageSize = "string"
      layersToFreeze = "string"
      learningRate = "string"
      learningRateScheduler = "string"
      maxSize = "string"
      minSize = "string"
      modelName = "string"
      modelSize = "string"
      momentum = "string"
      multiScale = "string"
      nesterov = "string"
      nmsIouThreshold = "string"
      numberOfEpochs = "string"
      numberOfWorkers = "string"
      optimizer = "string"
      randomSeed = "string"
      stepLRGamma = "string"
      stepLRStepSize = "string"
      tileGridSize = "string"
      tileOverlapRatio = "string"
      tilePredictionsNmsThreshold = "string"
      trainingBatchSize = "string"
      validationBatchSize = "string"
      validationIouThreshold = "string"
      validationMetricType = "string"
      warmupCosineLRCycles = "string"
      warmupCosineLRWarmupEpochs = "string"
      weightDecay = "string"
    }
  ]
  sweepSettings = {
    earlyTermination = {
      delayEvaluation = int
      evaluationInterval = int
      policyType = "string"
      // For remaining properties, see EarlyTerminationPolicy objects
    }
    samplingAlgorithm = "string"
  }
  taskType = "ImageInstanceSegmentation"
  validationData = {
    description = "string"
    jobInputType = "string"
    mode = "string"
    uri = "string"
  }
  validationDataSize = int
}

針對 ImageObjectDetection,請使用:

{
  limitSettings = {
    maxConcurrentTrials = int
    maxTrials = int
    timeout = "string"
  }
  modelSettings = {
    advancedSettings = "string"
    amsGradient = bool
    augmentations = "string"
    beta1 = int
    beta2 = int
    boxDetectionsPerImage = int
    boxScoreThreshold = int
    checkpointFrequency = int
    checkpointModel = {
      description = "string"
      jobInputType = "string"
      mode = "string"
      uri = "string"
    }
    checkpointRunId = "string"
    distributed = bool
    earlyStopping = bool
    earlyStoppingDelay = int
    earlyStoppingPatience = int
    enableOnnxNormalization = bool
    evaluationFrequency = int
    gradientAccumulationStep = int
    imageSize = int
    layersToFreeze = int
    learningRate = int
    learningRateScheduler = "string"
    maxSize = int
    minSize = int
    modelName = "string"
    modelSize = "string"
    momentum = int
    multiScale = bool
    nesterov = bool
    nmsIouThreshold = int
    numberOfEpochs = int
    numberOfWorkers = int
    optimizer = "string"
    randomSeed = int
    stepLRGamma = int
    stepLRStepSize = int
    tileGridSize = "string"
    tileOverlapRatio = int
    tilePredictionsNmsThreshold = int
    trainingBatchSize = int
    validationBatchSize = int
    validationIouThreshold = int
    validationMetricType = "string"
    warmupCosineLRCycles = int
    warmupCosineLRWarmupEpochs = int
    weightDecay = int
  }
  primaryMetric = "string"
  searchSpace = [
    {
      amsGradient = "string"
      augmentations = "string"
      beta1 = "string"
      beta2 = "string"
      boxDetectionsPerImage = "string"
      boxScoreThreshold = "string"
      distributed = "string"
      earlyStopping = "string"
      earlyStoppingDelay = "string"
      earlyStoppingPatience = "string"
      enableOnnxNormalization = "string"
      evaluationFrequency = "string"
      gradientAccumulationStep = "string"
      imageSize = "string"
      layersToFreeze = "string"
      learningRate = "string"
      learningRateScheduler = "string"
      maxSize = "string"
      minSize = "string"
      modelName = "string"
      modelSize = "string"
      momentum = "string"
      multiScale = "string"
      nesterov = "string"
      nmsIouThreshold = "string"
      numberOfEpochs = "string"
      numberOfWorkers = "string"
      optimizer = "string"
      randomSeed = "string"
      stepLRGamma = "string"
      stepLRStepSize = "string"
      tileGridSize = "string"
      tileOverlapRatio = "string"
      tilePredictionsNmsThreshold = "string"
      trainingBatchSize = "string"
      validationBatchSize = "string"
      validationIouThreshold = "string"
      validationMetricType = "string"
      warmupCosineLRCycles = "string"
      warmupCosineLRWarmupEpochs = "string"
      weightDecay = "string"
    }
  ]
  sweepSettings = {
    earlyTermination = {
      delayEvaluation = int
      evaluationInterval = int
      policyType = "string"
      // For remaining properties, see EarlyTerminationPolicy objects
    }
    samplingAlgorithm = "string"
  }
  taskType = "ImageObjectDetection"
  validationData = {
    description = "string"
    jobInputType = "string"
    mode = "string"
    uri = "string"
  }
  validationDataSize = int
}

針對 回歸,請使用:

{
  cvSplitColumnNames = [
    "string"
  ]
  featurizationSettings = {
    blockedTransformers = [
      "string"
    ]
    columnNameAndTypes = {
      {customized property} = "string"
    }
    datasetLanguage = "string"
    enableDnnFeaturization = bool
    mode = "string"
    transformerParams = {
      {customized property} = [
        {
          fields = [
            "string"
          ]
          parameters = ?
        }
      ]
    }
  }
  limitSettings = {
    enableEarlyTermination = bool
    exitScore = int
    maxConcurrentTrials = int
    maxCoresPerTrial = int
    maxTrials = int
    timeout = "string"
    trialTimeout = "string"
  }
  nCrossValidations = {
    mode = "string"
    // For remaining properties, see NCrossValidations objects
  }
  primaryMetric = "string"
  taskType = "Regression"
  testData = {
    description = "string"
    jobInputType = "string"
    mode = "string"
    uri = "string"
  }
  testDataSize = int
  trainingSettings = {
    allowedTrainingAlgorithms = [
      "string"
    ]
    blockedTrainingAlgorithms = [
      "string"
    ]
    enableDnnTraining = bool
    enableModelExplainability = bool
    enableOnnxCompatibleModels = bool
    enableStackEnsemble = bool
    enableVoteEnsemble = bool
    ensembleModelDownloadTimeout = "string"
    stackEnsembleSettings = {
      stackMetaLearnerKWargs = ?
      stackMetaLearnerTrainPercentage = int
      stackMetaLearnerType = "string"
    }
  }
  validationData = {
    description = "string"
    jobInputType = "string"
    mode = "string"
    uri = "string"
  }
  validationDataSize = int
  weightColumnName = "string"
}

針對 TextClassification,請使用:

{
  featurizationSettings = {
    datasetLanguage = "string"
  }
  limitSettings = {
    maxConcurrentTrials = int
    maxTrials = int
    timeout = "string"
  }
  primaryMetric = "string"
  taskType = "TextClassification"
  validationData = {
    description = "string"
    jobInputType = "string"
    mode = "string"
    uri = "string"
  }
}

針對 TextClassificationMultilabel,請使用:

{
  featurizationSettings = {
    datasetLanguage = "string"
  }
  limitSettings = {
    maxConcurrentTrials = int
    maxTrials = int
    timeout = "string"
  }
  taskType = "TextClassificationMultilabel"
  validationData = {
    description = "string"
    jobInputType = "string"
    mode = "string"
    uri = "string"
  }
}

針對 TextNER,請使用:

{
  featurizationSettings = {
    datasetLanguage = "string"
  }
  limitSettings = {
    maxConcurrentTrials = int
    maxTrials = int
    timeout = "string"
  }
  taskType = "TextNER"
  validationData = {
    description = "string"
    jobInputType = "string"
    mode = "string"
    uri = "string"
  }
}

屬性值

Microsoft.MachineLearningServices/workspaces/jobs

名稱 說明 價值觀
名稱 資源名稱 字串

約束:
模式 = ^[a-zA-Z0-9][a-zA-Z0-9\-_]{0,254}$ (必要)
父項識別碼 此資源為父系之資源的標識碼。 類型資源的標識碼: 工作區
屬性 [必要]實體的其他屬性。 JobBaseProperties (必要)
型別 資源類型 “Microsoft.MachineLearningServices/workspaces/jobs@2024-07-01-preview”

所有節點

名稱 說明 價值觀
nodesValueType [必要]Nodes 值的類型 'All' (必要)

AmlToken

名稱 說明 價值觀
身份類型 [必要]指定身分識別架構的類型。 'AMLToken' (必要)

自動預測地平線

名稱 說明 價值觀
模式 [必要]設定預測地平線值選取模式。 'Auto' (必要)

AutoMLJob

名稱 說明 價值觀
環境ID 作業之環境規格的 ARM 資源識別碼。
如果未提供,這是選擇性的值,AutoML 會在執行作業時,將此預設為 Production AutoML 策展環境版本。
字串
環境變數 作業中包含的環境變數。 AutoMLJobEnvironmentVariables
職位類型 [必要]指定作業的類型。 'AutoML' (必要)
輸出 對應作業中使用的輸出數據系結。 AutoMLJob 輸出
queueSettings 作業的佇列設定 佇列設置
資源 作業的計算資源組態。 JobResourceConfiguration
taskDetails [必要]這代表可以是其中一個數據表/NLP/Image 的案例 AutoMLVertical (必要)

AutoMLJobEnvironmentVariables

名稱 說明 價值觀

AutoMLJob 輸出

名稱 說明 價值觀

AutoMLVertical

名稱 說明 價值觀
logVerbosity (對數詳細程度) 作業的記錄詳細資訊。 “嚴重”
“調試”
“錯誤”
'資訊'
'NotSet'
“警告”
targetColumnName 目標數據行名稱:這是預測值數據行。
也稱為分類工作內容中的標籤數據行名稱。
字串
任務類型 設定為 [分類] 類型為 [ 分類]。 設定為 [預測] 類型為 [ 預測]。 設定為 ImageClassification 類型的 『ImageClassification』。 設定為 ImageClassificationMultilabel 類型的 'ImageClassificationMultilabel'。 針對 ImageInstanceSegmentation 類型,設定為 'ImageInstanceSegmentation'。 針對 ImageObjectDetection 類型,設定為 'ImageObjectDetection'。 將類型回歸設定為 [ 回歸]。 設定為 TextClassification 類型的 『TextClassification』。 針對 TextClassificationMultilabel 類型,設定為 『TextClassificationMultilabel』。 設定為 TextNer 類型的 『TextNER』。 '分類'
「預測」
'ImageClassification' (圖像分類)
'ImageClassificationMultilabel'
'ImageInstanceSegmentation' (圖像實例分段)
'ImageObjectDetection' (圖像對象檢測)
“回歸”
'TextClassification' (文本分類)
'TextClassificationMultilabel'
'TextNER' (必要)
trainingData (訓練數據) [必要]定型數據輸入。 MLTableJobInput (必要)

AutoNCross驗證

名稱 說明 價值觀
模式 [必要]判斷 N-Cross 驗證的模式。 'Auto' (必要)

自動季節性

名稱 說明 價值觀
模式 [必要]季節性模式。 'Auto' (必要)

AutoTargetLags 自動目標滯後

名稱 說明 價值觀
模式 [必要]設定目標延隔模式 - 自動/自定義 'Auto' (必要)

AutoTargetRollingWindowSize (自動目標滾動視窗大小)

名稱 說明 價值觀
模式 [必要]TargetRollingWindowSiz 偵測模式。 'Auto' (必要)

AzureDevOpsWebhook

名稱 說明 價值觀
webhook類型 [必要]指定要傳送回呼的服務類型 'AzureDevOps' (必要)

AzureOpenAiFineTuning

名稱 說明 價值觀
超參數 用於微調 Azure Open AI 模型的 HyperParameters。 AzureOpenAiHyperParameters
模型提供程式 [必要]用來判斷微調類型的列舉。 'AzureOpenAI' (必要)

AzureOpenAiHyperParameters

名稱 說明 價值觀
批次大小 每個批次中的範例數目。 批次大小越大,表示模型參數的更新頻率越低,而變異數越少。 整數 (int)
learningRateMultiplier 的 學習率的縮放比例。 較小的學習率可能有助於避免過度調整。 整數 (int)
nEpochs 訓練模型的週期數目。 Epoch 是指透過訓練資料集的一個完整週期。 整數 (int)

強盜政策 (BanditPolicy)

名稱 說明 價值觀
policyType [必要]原則設定的名稱 “強盜”(必要)
slackAmount 從最佳執行執行中允許的絕對距離。 整數 (int)
slackFactor 的 與最佳執行距離的允許距離比率。 整數 (int)

BayesianSamplingAlgorithm

名稱 說明 價值觀
採樣演算法類型 [必要]用來產生超參數值的演算法,以及組態屬性 “貝氏” (必要)

分類

名稱 說明 價值觀
cvSplitColumnNames 要用於 CVSplit 資料的數據行。 字串[]
featurizationSettings AutoML 作業所需的特徵化輸入。 TableVerticalFeaturizationSettings
limitSettings AutoMLJob 的執行條件約束。 TableVerticalLimitSettings
nCrossValidations 要套用至定型數據集的交叉驗證折疊數目
未提供驗證數據集時。
NCross驗證
積極標籤 二進位計量計算的正標籤。 字串
primaryMetric 工作的主要計量。 '精確度'
'AUCWeighted'
'AveragePrecisionScoreWeighted'
'NormMacroRecall' (規範巨集調用)
'PrecisionScoreWeighted'
任務類型 [必要]AutoMLJob 的工作類型。 '分類' (必要)
測試數據 測試數據輸入。 MLTableJobInput
testDataSize 必須為驗證目的而保留的測試數據集分數。
介於 (0.0 、 1.0) 之間的值
未提供驗證數據集時套用。
整數 (int)
trainingSettings AutoML 作業定型階段的輸入。 分類訓練設置
validationData (驗證資料) 驗證數據輸入。 MLTableJobInput
validationDataSize 需要為驗證目的預留的定型數據集分數。
介於 (0.0 、 1.0) 之間的值
未提供驗證數據集時套用。
整數 (int)
權重列名稱 範例權數數據行的名稱。 自動化 ML 支援加權數據行做為輸入,導致數據中的數據列向上或向下加權。 字串

分類訓練設置

名稱 說明 價值觀
allowedTrainingAlgorithms 分類工作的允許模型。 包含任何的字串數組:
'伯努利樸素貝葉斯'
“決策樹”
'ExtremeRandomTrees'
'梯度提升'
'KNN'
'LightGBM'
'線性 SVM'
'LogisticRegression' (邏輯回歸)
'多項式樸素貝葉斯'
'隨機森林'
'新幣'
“SVM”
'XGBoostClassifier'
blockedTrainingAlgorithms 分類工作的封鎖模型。 包含任何的字串數組:
'伯努利樸素貝葉斯'
“決策樹”
'ExtremeRandomTrees'
'梯度提升'
'KNN'
'LightGBM'
'線性 SVM'
'LogisticRegression' (邏輯回歸)
'多項式樸素貝葉斯'
'隨機森林'
'新幣'
“SVM”
'XGBoostClassifier'
enableDnnTraining 啟用 DNN 模型的建議。 布爾 (bool)
enableModelExplainability 在最佳模型上開啟可解釋性的旗標。 布爾 (bool)
enableOnnxCompatibleModels 啟用 onnx 相容模型的旗標。 布爾 (bool)
enableStackEnsemble 啟用堆疊合奏執行。 布爾 (bool)
enableVoteEnsemble 啟用投票合奏執行。 布爾 (bool)
ensembleModelDownloadTimeout 在 VotingEnsemble 和 StackEnsemble 模型產生期間,會下載先前子回合中的多個安裝模型。
如果需要更多時間,請使用高於 300 秒的值來設定此參數。
字串
stackEnsembleSettings 堆疊合奏執行的堆疊合奏設定。 StackEnsembleSettings

ColumnTransformer (列變換器)

名稱 說明 價值觀
領域 要套用轉換器邏輯的欄位。 字串[]
參數 要傳遞至轉換器的不同屬性。
輸入必須是 JSON 格式的索引鍵、值組字典。
任何

CommandJob

名稱 說明 價值觀
代碼Id 程式代碼資產的 ARM 資源識別碼。 字串
指令 [必要]在作業啟動時執行的命令。 例如 “python train.py” 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)
分配 作業的散發組態。 如果設定,這應該是其中一個 Mpi、Tensorflow、PyTorch 或 null。 分發配置
環境ID [必要]作業之環境規格的 ARM 資源識別碼。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)
環境變數 作業中包含的環境變數。 CommandJobEnvironmentVariables
輸入 對應作業中使用的輸入數據系結。 CommandJobInputs 命令作業輸入
職位類型 [必要]指定作業的類型。 'Command' (必要)
限制 命令作業限制。 CommandJobLimits 命令作業限制
輸出 對應作業中使用的輸出數據系結。 CommandJobOutputs 命令作業輸出
queueSettings 作業的佇列設定 佇列設置
資源 作業的計算資源組態。 JobResourceConfiguration

CommandJobEnvironmentVariables

名稱 說明 價值觀

CommandJobInputs 命令作業輸入

名稱 說明 價值觀

CommandJobLimits 命令作業限制

名稱 說明 價值觀
jobLimits類型 [必要]JobLimit 類型。 “命令”
'掃掠' (必要)
超時 ISO 8601 格式的最大執行持續時間,之後作業將會取消。 僅支援有效位數為秒數的持續時間。 字串

CommandJobOutputs 命令作業輸出

名稱 說明 價值觀

CustomForecastHorizon

名稱 說明 價值觀
模式 [必要]設定預測地平線值選取模式。 'Custom' (必要)
價值 [必要]預測地平線值。 int (必要)

CustomModelFineTuning 微調

名稱 說明 價值觀
超參數 微調自定義模型的 HyperParameters。 CustomModelFineTuningHyperParameters
模型提供程式 [必要]用來判斷微調類型的列舉。 'Custom' (必要)

CustomModelFineTuningHyperParameters

名稱 說明 價值觀

CustomModelJobInput (自定義模型作業輸入)

名稱 說明 價值觀
jobInputType (作業輸入類型) [必要]指定作業的類型。 'custom_model' (必要)
模式 輸入資產傳遞模式。 “直接”
'下載'
'EvalDownload'
'EvalMount'
'ReadOnlyMount'
“ReadWriteMount”
統一資源識別碼 (URI) [必要]輸入資產 URI。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)

CustomModelJobOutput

名稱 說明 價值觀
jobOutputType [必要]指定作業的類型。 'custom_model' (必要)
模式 輸出資產傳遞模式。 “直接”
“ReadWriteMount”
'上傳'
統一資源識別碼 (URI) 輸出資產 URI。 字串

CustomNCross驗證

名稱 說明 價值觀
模式 [必要]判斷 N-Cross 驗證的模式。 'Custom' (必要)
價值 [必要]N-Cross 驗證值。 int (必要)

CustomSeasonality (自定義季節性)

名稱 說明 價值觀
模式 [必要]季節性模式。 'Custom' (必要)
價值 [必要]季節性值。 int (必要)

CustomTargetLags (自定義目標滯後)

名稱 說明 價值觀
模式 [必要]設定目標延隔模式 - 自動/自定義 'Custom' (必要)
價值觀 [必要]設定目標延遲值。 int[] (必要)

CustomTargetRollingWindowSize

名稱 說明 價值觀
模式 [必要]TargetRollingWindowSiz 偵測模式。 'Custom' (必要)
價值 [必要]TargetRollingWindowSize 值。 int (必要)

分發配置

名稱 說明 價值觀
distributionType 設定為 Mpi 類型的 『Mpi』。 將類型 PyTorch 設定為 『PyTorch』。 將 設定為 TensorFlow 類型的 『TensorFlow』。 'Mpi'
“PyTorch”
'TensorFlow' (必要)

提早終止政策

名稱 說明 價值觀
delay評估 延遲第一次評估的間隔數目。 整數 (int)
evaluationInterval 原則評估之間的間隔(執行次數)。 整數 (int)
policyType 針對 BanditPolicy 類型,設定為 『Bandit』。 針對 MedianStoppingPolicy 類型,設定為 'MedianStoppingPolicy'。 針對 TruncationSelectionPolicy 類型,設定為 『TruncationSelection』 “強盜”
'MedianStopping' (中位數停止)
'TruncationSelection' (必要)

微調作業

名稱 說明 價值觀
微調細節 [必要] FineTuningVertical (必要)
職位類型 [必要]指定作業的類型。 'FineTuning' (必要)
輸出 [必要] FineTuningJobOutputs (必要)
queueSettings 作業的佇列設定 佇列設置
資源 作業的實例類型和其他資源 工作資源

FineTuningJobOutputs(微調作業輸出)

名稱 說明 價值觀

微調垂直

名稱 說明 價值觀
型號 [必要]用於微調的輸入模型。 MLFlowModelJobInput (必要)
模型提供程式 針對 AzureOpen AiFineTuning 類型,設定為 『AzureOpenAI』。 設定為 CustomModelFineTuning 類型的 [自定義]。 'AzureOpenAI'
'Custom' (必要)
任務類型 [必要]微調工作類型。 'ChatCompletion'
'ImageClassification' (圖像分類)
'ImageInstanceSegmentation' (圖像實例分段)
'ImageObjectDetection' (圖像對象檢測)
“問答”
'TextClassification' (文本分類)
'TextCompletion'
'TextSummarization' (文本摘要)
'TextTranslation'
'TokenClassification'
'VideoMultiObjectTracking' (必要)
trainingData (訓練數據) [必要]定型數據以進行微調。 JobInput (必要)
validationData (驗證資料) 微調的驗證數據。 作業輸入

預測地平線

名稱 說明 價值觀
模式 針對 AutoForecastHorizon 類型,設定為 'Auto'。 針對 CustomForecastHorizon 類型,設定為 'Custom'。 '自動'
'Custom' (必要)

預測

名稱 說明 價值觀
cvSplitColumnNames 要用於 CVSplit 資料的數據行。 字串[]
featurizationSettings AutoML 作業所需的特徵化輸入。 TableVerticalFeaturizationSettings
forecasting設置 預測工作特定輸入。 ForecastingSettings
limitSettings AutoMLJob 的執行條件約束。 TableVerticalLimitSettings
nCrossValidations 要套用至定型數據集的交叉驗證折疊數目
未提供驗證數據集時。
NCross驗證
primaryMetric 預測工作的主要計量。 'NormalizedMeanAbsoluteError'
'NormalizedRootMeanSquaredError'
'R2Score'
'SpearmanCorrelation'
任務類型 [必要]AutoMLJob 的工作類型。 “預測” (必要)
測試數據 測試數據輸入。 MLTableJobInput
testDataSize 必須為驗證目的而保留的測試數據集分數。
介於 (0.0 、 1.0) 之間的值
未提供驗證數據集時套用。
整數 (int)
trainingSettings AutoML 作業定型階段的輸入。 ForecastingTrainingSettings
validationData (驗證資料) 驗證數據輸入。 MLTableJobInput
validationDataSize 需要為驗證目的預留的定型數據集分數。
介於 (0.0 、 1.0) 之間的值
未提供驗證數據集時套用。
整數 (int)
權重列名稱 範例權數數據行的名稱。 自動化 ML 支援加權數據行做為輸入,導致數據中的數據列向上或向下加權。 字串

ForecastingSettings

名稱 說明 價值觀
國家/地區或地區為假期 用於預測工作的假日國家或地區。
這些應該是 ISO 3166 兩個字母的國家/地區代碼,例如 'US' 或 'GB'。
字串
cvStepSize 一個 CV 折疊的原點時間與下一個折疊之間的期間數。 為了
例如,如果 CVStepSize = 3 表示每日數據,則每個折疊的原點時間將會是
相隔三天。
整數 (int)
featureLags 使用 『auto』 或 null 產生數值特徵延遲的旗標。 '自動'
“無”
forecastHorizon (預測地平線) 所需的最大預測範圍,以時間序列頻率單位為單位。 預測地平線
頻率 預測時,此參數代表預期預測的期間,例如每日、每周、每年等。預測頻率預設為數據集頻率。 字串
季節性 將時間序列季節性設定為數列頻率的整數倍數。
如果季節性設定為 『auto』,則會推斷它。
季節性
shortSeriesHandlingConfig 定義 AutoML 是否應該處理短時間序列的參數。 '自動'
“掉落”
“無”
'墊'
targetAggregateFunction 要用來匯總時間序列目標數據行以符合使用者指定頻率的函式。
如果 TargetAggregateFunction 已設定,亦即不是 『None』,但未設定 freq 參數,則會引發錯誤。 可能的目標聚合函數包括:「sum」、“max”、“min” 和 “mean”。
“最大”
“卑鄙”
'敏'
“無”
'總和'
目標拉格 要從目標數據行延遲的過去期間數。 目標滯後
targetRollingWindowSize 用來建立目標數據行滾動視窗平均值的過去期間數目。 TargetRollingWindowSize
timeColumnName (時間列名稱) 時間數據行的名稱。 當預測以指定用於建置時間序列並推斷其頻率的輸入數據中的 datetime 數據行時,需要此參數。 字串
timeSeriesIdColumnNames 用來分組時間範圍的數據行名稱。 它可以用來建立多個數列。
如果未定義粒紋,則會假設數據集為一個時間序列。 此參數與工作類型預測搭配使用。
字串[]
useStl 設定時間序列目標數據行的 STL 分解。 “無”
“賽季”
'季節趨勢'

ForecastingTrainingSettings

名稱 說明 價值觀
allowedTrainingAlgorithms 用於預測工作的允許模型。 包含任何的字串數組:
“Arimax”
'AutoArima'
“平均”
“決策樹”
“彈性網”
'指數平滑'
'ExtremeRandomTrees'
'梯度提升'
'KNN'
“套索拉斯”
'LightGBM'
天真'
'先知'
'隨機森林'
'季節性平均值'
'季節性天真'
'新幣'
'TCNForecaster'
'XGBoostRegressor'
blockedTrainingAlgorithms 用於預測工作的封鎖模型。 包含任何的字串數組:
“Arimax”
'AutoArima'
“平均”
“決策樹”
“彈性網”
'指數平滑'
'ExtremeRandomTrees'
'梯度提升'
'KNN'
“套索拉斯”
'LightGBM'
天真'
'先知'
'隨機森林'
'季節性平均值'
'季節性天真'
'新幣'
'TCNForecaster'
'XGBoostRegressor'
enableDnnTraining 啟用 DNN 模型的建議。 布爾 (bool)
enableModelExplainability 在最佳模型上開啟可解釋性的旗標。 布爾 (bool)
enableOnnxCompatibleModels 啟用 onnx 相容模型的旗標。 布爾 (bool)
enableStackEnsemble 啟用堆疊合奏執行。 布爾 (bool)
enableVoteEnsemble 啟用投票合奏執行。 布爾 (bool)
ensembleModelDownloadTimeout 在 VotingEnsemble 和 StackEnsemble 模型產生期間,會下載先前子回合中的多個安裝模型。
如果需要更多時間,請使用高於 300 秒的值來設定此參數。
字串
stackEnsembleSettings 堆疊合奏執行的堆疊合奏設定。 StackEnsembleSettings

GridSampling演算法

名稱 說明 價值觀
採樣演算法類型 [必要]用來產生超參數值的演算法,以及組態屬性 'Grid' (必要)

身份配置

名稱 說明 價值觀
身份類型 將 設定為 AmlToken 類型的 『AMLToken』。 針對 ManagedIdentity 類型設定為 『Managed』。 針對 UserIdentity 類型,設定為 ' UserIdentity' 'AMLToken'
“託管”
'UserIdentity' (必要)

圖像分類

名稱 說明 價值觀
limitSettings [必要]限制 AutoML 作業的設定。 ImageLimitSettings (必要)
modelSettings 用於定型模型的設定。 ImageModelSettings分類
primaryMetric 要針對這項工作進行優化的主要計量。 '精確度'
'AUCWeighted'
'AveragePrecisionScoreWeighted'
'NormMacroRecall' (規範巨集調用)
'PrecisionScoreWeighted'
搜索空間 搜尋空間以取樣模型及其超參數的不同組合。 ImageModelDistributionSettingsClassification[]
掃描設置 模型掃掠和超參數掃掠相關設定。 ImageSweepSettings (圖像掃描設定)
任務類型 [必要]AutoMLJob 的工作類型。 'ImageClassification' (必要)
validationData (驗證資料) 驗證數據輸入。 MLTableJobInput
validationDataSize 需要為驗證目的預留的定型數據集分數。
介於 (0.0 、 1.0) 之間的值
未提供驗證數據集時套用。
整數 (int)

ImageClassificationMultilabel

名稱 說明 價值觀
limitSettings [必要]限制 AutoML 作業的設定。 ImageLimitSettings (必要)
modelSettings 用於定型模型的設定。 ImageModelSettings分類
primaryMetric 要針對這項工作進行優化的主要計量。 '精確度'
'AUCWeighted'
'AveragePrecisionScoreWeighted'
'欠條'
'NormMacroRecall' (規範巨集調用)
'PrecisionScoreWeighted'
搜索空間 搜尋空間以取樣模型及其超參數的不同組合。 ImageModelDistributionSettingsClassification[]
掃描設置 模型掃掠和超參數掃掠相關設定。 ImageSweepSettings (圖像掃描設定)
任務類型 [必要]AutoMLJob 的工作類型。 'ImageClassificationMultilabel' (必要)
validationData (驗證資料) 驗證數據輸入。 MLTableJobInput
validationDataSize 需要為驗證目的預留的定型數據集分數。
介於 (0.0 、 1.0) 之間的值
未提供驗證數據集時套用。
整數 (int)

ImageInstanceSegmentation

名稱 說明 價值觀
limitSettings [必要]限制 AutoML 作業的設定。 ImageLimitSettings (必要)
modelSettings 用於定型模型的設定。 ImageModelSettingsObjectDetection
primaryMetric 要針對這項工作進行優化的主要計量。 'MeanAveragePrecision'
搜索空間 搜尋空間以取樣模型及其超參數的不同組合。 ImageModelDistributionSettingsObjectDetection[]
掃描設置 模型掃掠和超參數掃掠相關設定。 ImageSweepSettings (圖像掃描設定)
任務類型 [必要]AutoMLJob 的工作類型。 'ImageInstanceSegmentation' (必要)
validationData (驗證資料) 驗證數據輸入。 MLTableJobInput
validationDataSize 需要為驗證目的預留的定型數據集分數。
介於 (0.0 、 1.0) 之間的值
未提供驗證數據集時套用。
整數 (int)

ImageLimitSettings

名稱 說明 價值觀
maxConcurrentTrials 並行 AutoML 反覆項目的數目上限。 整數 (int)
maxTrials AutoML 反覆項目的數目上限。 整數 (int)
超時 AutoML 作業逾時。 字串

ImageModelDistributionSettings分類

名稱 說明 價值觀
amsGradient 當優化器為 『adam』 或 'adamw' 時啟用AMSGrad。 字串
擴增 使用擴增的設定。 字串
貝塔1 優化器為 'adam' 或 'adamw' 時的 'beta1' 值。 必須是範圍 [0, 1] 內的浮點數。 字串
貝塔2 優化器為 'adam' 或 'adamw' 時的 'beta2' 值。 必須是範圍 [0, 1] 內的浮點數。 字串
分散式 是否要使用散發器定型。 字串
earlyStopping (早期停止) 在定型期間啟用早期停止邏輯。 字串
earlyStoppingDelay 在主要計量改進之前等待的最小 Epoch 或驗證評估數目
會追蹤早期停止。 必須是正整數。
字串
早點停止耐心 在之前沒有主要計量改進的最小 Epoch 或驗證評估數目
執行已停止。 必須是正整數。
字串
enableOnnxNormalization 匯出 ONNX 模型時啟用正規化。 字串
evaluationFrequency 評估驗證數據集以取得計量分數的頻率。 必須是正整數。 字串
梯度累積步驟 漸層累積表示執行未設定數目的 「GradAccumulationStep」 步驟
更新模型權數,同時累積這些步驟的漸層,然後使用
用來計算權數更新的累積漸層。 必須是正整數。
字串
layersToFreeze (圖層凍結) 要凍結模型的圖層數目。 必須是正整數。
例如,傳遞 2 作為 'seresnext' 值表示
凍結 layer0 和 layer1。 如需支援模型的完整清單,以及圖層凍結的詳細數據,請
請參閱:/azure/machine-learning/how-to-auto-train-image-models。
字串
learningRate 初始學習速率。 必須是範圍 [0, 1] 內的浮點數。 字串
learningRateScheduler 的 學習率排程器的類型。 必須是 『warmup_cosine』 或 'step'。 字串
型號名稱 要用於定型的模型名稱。
如需可用模型的詳細資訊,請流覽官方檔:
/azure/machine-learning/how-to-auto-train-image-models。
字串
動量 優化器為 『zn』 時的動量值。 必須是範圍 [0, 1] 內的浮點數。 字串
涅斯特羅夫 在優化器為 『zn』 時啟用 nesterov。 字串
紀元數 定型 Epoch 的數目。 必須是正整數。 字串
工人數 數據載入器背景工作角色的數目。 必須是非負整數。 字串
優化器 優化工具的類型。 必須是 『zn』、『adam』或 『adamw』。 字串
隨機種子 使用決定性定型時要使用的隨機種子。 字串
stepLR伽瑪 學習速率排程器為 『step』 時的 gamma 值。 必須是範圍 [0, 1] 內的浮點數。 字串
stepLRStepSize 學習速率排程器為 『step』 時的步驟大小值。 必須是正整數。 字串
trainingBatchSize 定型批次大小。 必須是正整數。 字串
trainingCropSize 影像裁剪大小,這是訓練數據集之類神經網路的輸入。 必須是正整數。 字串
validationBatchSize 驗證批次大小。 必須是正整數。 字串
validationCropSize 影像裁剪大小,這是用於驗證數據集之類神經網路的輸入。 必須是正整數。 字串
validationResizeSize 在裁剪驗證數據集之前要重設大小的影像大小。 必須是正整數。 字串
warmupCosineLRCycles 學習速率排程器為「warmup_cosine」時,餘弦迴圈的值。 必須是範圍 [0, 1] 內的浮點數。 字串
warmupCosineLRWarmupEpochs 學習速率排程器為「warmup_cosine」時,熱身 Epoch 的值。 必須是正整數。 字串
weightDecay (權重衰減) 優化器為 'zn'、'adam' 或 'adamw' 時的權數衰變值。 必須是範圍[0, 1] 中的浮點數。 字串
weightedLoss 加權減肥。 接受的值是 0,沒有加權損失。
1 用於與 sqrt 減肥。(class_weights)。 2 用於減肥與class_weights。 必須是 0 或 1 或 2。
字串

ImageModelDistributionSettingsObjectDetection

名稱 說明 價值觀
amsGradient 當優化器為 『adam』 或 'adamw' 時啟用AMSGrad。 字串
擴增 使用擴增的設定。 字串
貝塔1 優化器為 'adam' 或 'adamw' 時的 'beta1' 值。 必須是範圍 [0, 1] 內的浮點數。 字串
貝塔2 優化器為 'adam' 或 'adamw' 時的 'beta2' 值。 必須是範圍 [0, 1] 內的浮點數。 字串
boxDetectionsPerImage (每影像檢測) 所有類別的每個影像偵測數目上限。 必須是正整數。
注意:'yolov5' 演算法不支援此設定。
字串
boxScoreThreshold (盒分數閾值) 在推斷期間,只傳回大於分類分數的建議
BoxScoreThreshold。 必須是範圍[0, 1] 中的浮點數。
字串
分散式 是否要使用散發器定型。 字串
earlyStopping (早期停止) 在定型期間啟用早期停止邏輯。 字串
earlyStoppingDelay 在主要計量改進之前等待的最小 Epoch 或驗證評估數目
會追蹤早期停止。 必須是正整數。
字串
早點停止耐心 在之前沒有主要計量改進的最小 Epoch 或驗證評估數目
執行已停止。 必須是正整數。
字串
enableOnnxNormalization 匯出 ONNX 模型時啟用正規化。 字串
evaluationFrequency 評估驗證數據集以取得計量分數的頻率。 必須是正整數。 字串
梯度累積步驟 漸層累積表示執行未設定數目的 「GradAccumulationStep」 步驟
更新模型權數,同時累積這些步驟的漸層,然後使用
用來計算權數更新的累積漸層。 必須是正整數。
字串
影像大小 定型和驗證的影像大小。 必須是正整數。
注意:如果大小太大,定型回合可能會進入 CUDA OOM。
注意:只有 'yolov5' 演算法才支援此設定。
字串
layersToFreeze (圖層凍結) 要凍結模型的圖層數目。 必須是正整數。
例如,傳遞 2 作為 'seresnext' 值表示
凍結 layer0 和 layer1。 如需支援模型的完整清單,以及圖層凍結的詳細數據,請
請參閱:/azure/machine-learning/how-to-auto-train-image-models。
字串
learningRate 初始學習速率。 必須是範圍 [0, 1] 內的浮點數。 字串
learningRateScheduler 的 學習率排程器的類型。 必須是 『warmup_cosine』 或 'step'。 字串
最大尺寸 在將影像饋送至骨幹之前,要重新調整的影像大小上限。
必須是正整數。 注意:如果大小太大,定型執行可能會進入 CUDA OOM。
注意:'yolov5' 演算法不支援此設定。
字串
最小大小 將影像饋送至骨幹之前,要重新調整的影像大小下限。
必須是正整數。 注意:如果大小太大,定型執行可能會進入 CUDA OOM。
注意:'yolov5' 演算法不支援此設定。
字串
型號名稱 要用於定型的模型名稱。
如需可用模型的詳細資訊,請流覽官方檔:
/azure/machine-learning/how-to-auto-train-image-models。
字串
modelSize 模型大小。 必須是 'small'、'medium'、'large' 或 'xlarge'。
注意:如果模型大小太大,定型執行可能會進入 CUDA OOM。
注意:只有 'yolov5' 演算法才支援此設定。
字串
動量 優化器為 『zn』 時的動量值。 必須是範圍 [0, 1] 內的浮點數。 字串
多尺度 依 +/- 50%來啟用多重縮放影像大小。
注意:如果沒有足夠的 GPU 記憶體,定型執行可能會進入 CUDA OOM。
注意:只有 'yolov5' 演算法才支援此設定。
字串
涅斯特羅夫 在優化器為 『zn』 時啟用 nesterov。 字串
nmsIouThreshold 在NMS後置處理中的推斷期間使用的IOU臨界值。 必須是範圍 [0, 1] 中的 float。 字串
紀元數 定型 Epoch 的數目。 必須是正整數。 字串
工人數 數據載入器背景工作角色的數目。 必須是非負整數。 字串
優化器 優化工具的類型。 必須是 『zn』、『adam』或 『adamw』。 字串
隨機種子 使用決定性定型時要使用的隨機種子。 字串
stepLR伽瑪 學習速率排程器為 『step』 時的 gamma 值。 必須是範圍 [0, 1] 內的浮點數。 字串
stepLRStepSize 學習速率排程器為 『step』 時的步驟大小值。 必須是正整數。 字串
tileGridSize (平鋪網格大小) 用來貼上每個影像的網格線大小。 注意:TileGridSize 不得為
無可啟用小型物件偵測邏輯。 包含兩個整數的字串,其格式為一個。
注意:'yolov5' 演算法不支援此設定。
字串
tileOverlapRatio (平鋪重疊比率) 每個維度中相鄰磚之間的重疊比例。 必須是範圍 [0, 1] 中的浮點數。
注意:'yolov5' 演算法不支援此設定。
字串
tilePredictionsNmsThreshold 用來執行NMS的IOU閾值,同時合併磚和影像的預測。
用於驗證/推斷。 必須是範圍 [0, 1] 中的 float。
注意:'yolov5' 演算法不支援此設定。
NMS:非最大歸併
字串
trainingBatchSize 定型批次大小。 必須是正整數。 字串
validationBatchSize 驗證批次大小。 必須是正整數。 字串
validationIouThreshold 計算驗證計量時要使用的IOU閾值。 必須是範圍 [0, 1] 中的 float。 字串
validationMetricType 用於驗證計量的計量計算方法。 必須是 『none』、『coco』、'voc'或 'coco_voc'。 字串
warmupCosineLRCycles 學習速率排程器為「warmup_cosine」時,餘弦迴圈的值。 必須是範圍 [0, 1] 內的浮點數。 字串
warmupCosineLRWarmupEpochs 學習速率排程器為「warmup_cosine」時,熱身 Epoch 的值。 必須是正整數。 字串
weightDecay (權重衰減) 優化器為 'zn'、'adam' 或 'adamw' 時的權數衰變值。 必須是範圍[0, 1] 中的浮點數。 字串

ImageModelSettings分類

名稱 說明 價值觀
高級設置 進階案例的設定。 字串
amsGradient 當優化器為 『adam』 或 'adamw' 時啟用AMSGrad。 布爾 (bool)
擴增 使用擴增的設定。 字串
貝塔1 優化器為 'adam' 或 'adamw' 時的 'beta1' 值。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
貝塔2 優化器為 'adam' 或 'adamw' 時的 'beta2' 值。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
檢查點頻率 儲存模型檢查點的頻率。 必須是正整數。 整數 (int)
checkpoint模型 累加訓練的預先定型檢查點模型。 MLFlowModelJobInput
檢查點RunId 先前執行的標識碼,其具有累加式定型的預先定型檢查點。 字串
分散式 是否要使用分散式定型。 布爾 (bool)
earlyStopping (早期停止) 在定型期間啟用早期停止邏輯。 布爾 (bool)
earlyStoppingDelay 在主要計量改進之前等待的最小 Epoch 或驗證評估數目
會追蹤早期停止。 必須是正整數。
整數 (int)
早點停止耐心 在之前沒有主要計量改進的最小 Epoch 或驗證評估數目
執行已停止。 必須是正整數。
整數 (int)
enableOnnxNormalization 匯出 ONNX 模型時啟用正規化。 布爾 (bool)
evaluationFrequency 評估驗證數據集以取得計量分數的頻率。 必須是正整數。 整數 (int)
梯度累積步驟 漸層累積表示執行未設定數目的 「GradAccumulationStep」 步驟
更新模型權數,同時累積這些步驟的漸層,然後使用
用來計算權數更新的累積漸層。 必須是正整數。
整數 (int)
layersToFreeze (圖層凍結) 要凍結模型的圖層數目。 必須是正整數。
例如,傳遞 2 作為 'seresnext' 值表示
凍結 layer0 和 layer1。 如需支援模型的完整清單,以及圖層凍結的詳細數據,請
請參閱:/azure/machine-learning/how-to-auto-train-image-models。
整數 (int)
learningRate 初始學習速率。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
learningRateScheduler 的 學習率排程器的類型。 必須是 『warmup_cosine』 或 'step'。 “無”
“步驟”
'WarmupCosine'
型號名稱 要用於定型的模型名稱。
如需可用模型的詳細資訊,請流覽官方檔:
/azure/machine-learning/how-to-auto-train-image-models。
字串
動量 優化器為 『zn』 時的動量值。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
涅斯特羅夫 在優化器為 『zn』 時啟用 nesterov。 布爾 (bool)
紀元數 定型 Epoch 的數目。 必須是正整數。 整數 (int)
工人數 數據載入器背景工作角色的數目。 必須是非負整數。 整數 (int)
優化器 優化工具的類型。 “亞當”
“亞當”
“無”
'新元'
隨機種子 使用決定性定型時要使用的隨機種子。 整數 (int)
stepLR伽瑪 學習速率排程器為 『step』 時的 gamma 值。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
stepLRStepSize 學習速率排程器為 『step』 時的步驟大小值。 必須是正整數。 整數 (int)
trainingBatchSize 定型批次大小。 必須是正整數。 整數 (int)
trainingCropSize 影像裁剪大小,這是訓練數據集之類神經網路的輸入。 必須是正整數。 整數 (int)
validationBatchSize 驗證批次大小。 必須是正整數。 整數 (int)
validationCropSize 影像裁剪大小,這是用於驗證數據集之類神經網路的輸入。 必須是正整數。 整數 (int)
validationResizeSize 在裁剪驗證數據集之前要重設大小的影像大小。 必須是正整數。 整數 (int)
warmupCosineLRCycles 學習速率排程器為「warmup_cosine」時,餘弦迴圈的值。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
warmupCosineLRWarmupEpochs 學習速率排程器為「warmup_cosine」時,熱身 Epoch 的值。 必須是正整數。 整數 (int)
weightDecay (權重衰減) 優化器為 'zn'、'adam' 或 'adamw' 時的權數衰變值。 必須是範圍[0, 1] 中的浮點數。 整數 (int)
weightedLoss 加權減肥。 接受的值是 0,沒有加權損失。
1 用於與 sqrt 減肥。(class_weights)。 2 用於減肥與class_weights。 必須是 0 或 1 或 2。
整數 (int)

ImageModelSettingsObjectDetection

名稱 說明 價值觀
高級設置 進階案例的設定。 字串
amsGradient 當優化器為 『adam』 或 'adamw' 時啟用AMSGrad。 布爾 (bool)
擴增 使用擴增的設定。 字串
貝塔1 優化器為 'adam' 或 'adamw' 時的 'beta1' 值。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
貝塔2 優化器為 'adam' 或 'adamw' 時的 'beta2' 值。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
boxDetectionsPerImage (每影像檢測) 所有類別的每個影像偵測數目上限。 必須是正整數。
注意:'yolov5' 演算法不支援此設定。
整數 (int)
boxScoreThreshold (盒分數閾值) 在推斷期間,只傳回大於分類分數的建議
BoxScoreThreshold。 必須是範圍[0, 1] 中的浮點數。
整數 (int)
檢查點頻率 儲存模型檢查點的頻率。 必須是正整數。 整數 (int)
checkpoint模型 累加訓練的預先定型檢查點模型。 MLFlowModelJobInput
檢查點RunId 先前執行的標識碼,其具有累加式定型的預先定型檢查點。 字串
分散式 是否要使用分散式定型。 布爾 (bool)
earlyStopping (早期停止) 在定型期間啟用早期停止邏輯。 布爾 (bool)
earlyStoppingDelay 在主要計量改進之前等待的最小 Epoch 或驗證評估數目
會追蹤早期停止。 必須是正整數。
整數 (int)
早點停止耐心 在之前沒有主要計量改進的最小 Epoch 或驗證評估數目
執行已停止。 必須是正整數。
整數 (int)
enableOnnxNormalization 匯出 ONNX 模型時啟用正規化。 布爾 (bool)
evaluationFrequency 評估驗證數據集以取得計量分數的頻率。 必須是正整數。 整數 (int)
梯度累積步驟 漸層累積表示執行未設定數目的 「GradAccumulationStep」 步驟
更新模型權數,同時累積這些步驟的漸層,然後使用
用來計算權數更新的累積漸層。 必須是正整數。
整數 (int)
影像大小 定型和驗證的影像大小。 必須是正整數。
注意:如果大小太大,定型回合可能會進入 CUDA OOM。
注意:只有 'yolov5' 演算法才支援此設定。
整數 (int)
layersToFreeze (圖層凍結) 要凍結模型的圖層數目。 必須是正整數。
例如,傳遞 2 作為 'seresnext' 值表示
凍結 layer0 和 layer1。 如需支援模型的完整清單,以及圖層凍結的詳細數據,請
請參閱:/azure/machine-learning/how-to-auto-train-image-models。
整數 (int)
learningRate 初始學習速率。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
learningRateScheduler 的 學習率排程器的類型。 必須是 『warmup_cosine』 或 'step'。 “無”
“步驟”
'WarmupCosine'
最大尺寸 在將影像饋送至骨幹之前,要重新調整的影像大小上限。
必須是正整數。 注意:如果大小太大,定型執行可能會進入 CUDA OOM。
注意:'yolov5' 演算法不支援此設定。
整數 (int)
最小大小 將影像饋送至骨幹之前,要重新調整的影像大小下限。
必須是正整數。 注意:如果大小太大,定型執行可能會進入 CUDA OOM。
注意:'yolov5' 演算法不支援此設定。
整數 (int)
型號名稱 要用於定型的模型名稱。
如需可用模型的詳細資訊,請流覽官方檔:
/azure/machine-learning/how-to-auto-train-image-models。
字串
modelSize 模型大小。 必須是 'small'、'medium'、'large' 或 'xlarge'。
注意:如果模型大小太大,定型執行可能會進入 CUDA OOM。
注意:只有 'yolov5' 演算法才支援此設定。
“超大號”
“大”
'中等'
“無”
“小”
動量 優化器為 『zn』 時的動量值。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
多尺度 依 +/- 50%來啟用多重縮放影像大小。
注意:如果沒有足夠的 GPU 記憶體,定型執行可能會進入 CUDA OOM。
注意:只有 'yolov5' 演算法才支援此設定。
布爾 (bool)
涅斯特羅夫 在優化器為 『zn』 時啟用 nesterov。 布爾 (bool)
nmsIouThreshold 在NMS後置處理中的推斷期間使用的IOU臨界值。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
紀元數 定型 Epoch 的數目。 必須是正整數。 整數 (int)
工人數 數據載入器背景工作角色的數目。 必須是非負整數。 整數 (int)
優化器 優化工具的類型。 “亞當”
“亞當”
“無”
'新元'
隨機種子 使用決定性定型時要使用的隨機種子。 整數 (int)
stepLR伽瑪 學習速率排程器為 『step』 時的 gamma 值。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
stepLRStepSize 學習速率排程器為 『step』 時的步驟大小值。 必須是正整數。 整數 (int)
tileGridSize (平鋪網格大小) 用來貼上每個影像的網格線大小。 注意:TileGridSize 不得為
無可啟用小型物件偵測邏輯。 包含兩個整數的字串,其格式為一個。
注意:'yolov5' 演算法不支援此設定。
字串
tileOverlapRatio (平鋪重疊比率) 每個維度中相鄰磚之間的重疊比例。 必須是範圍 [0, 1] 中的浮點數。
注意:'yolov5' 演算法不支援此設定。
整數 (int)
tilePredictionsNmsThreshold 用來執行NMS的IOU閾值,同時合併磚和影像的預測。
用於驗證/推斷。 必須是範圍 [0, 1] 中的 float。
注意:'yolov5' 演算法不支援此設定。
整數 (int)
trainingBatchSize 定型批次大小。 必須是正整數。 整數 (int)
validationBatchSize 驗證批次大小。 必須是正整數。 整數 (int)
validationIouThreshold 計算驗證計量時要使用的IOU閾值。 必須是範圍 [0, 1] 中的 float。 整數 (int)
validationMetricType 用於驗證計量的計量計算方法。 《尋夢環遊記》
“可哥沃克”
“無”
'Voc'
warmupCosineLRCycles 學習速率排程器為「warmup_cosine」時,餘弦迴圈的值。 必須是範圍 [0, 1] 內的浮點數。 整數 (int)
warmupCosineLRWarmupEpochs 學習速率排程器為「warmup_cosine」時,熱身 Epoch 的值。 必須是正整數。 整數 (int)
weightDecay (權重衰減) 優化器為 'zn'、'adam' 或 'adamw' 時的權數衰變值。 必須是範圍[0, 1] 中的浮點數。 整數 (int)

ImageObjectDetection

名稱 說明 價值觀
limitSettings [必要]限制 AutoML 作業的設定。 ImageLimitSettings (必要)
modelSettings 用於定型模型的設定。 ImageModelSettingsObjectDetection
primaryMetric 要針對這項工作進行優化的主要計量。 'MeanAveragePrecision'
搜索空間 搜尋空間以取樣模型及其超參數的不同組合。 ImageModelDistributionSettingsObjectDetection[]
掃描設置 模型掃掠和超參數掃掠相關設定。 ImageSweepSettings (圖像掃描設定)
任務類型 [必要]AutoMLJob 的工作類型。 'ImageObjectDetection' (必要)
validationData (驗證資料) 驗證數據輸入。 MLTableJobInput
validationDataSize 需要為驗證目的預留的定型數據集分數。
介於 (0.0 、 1.0) 之間的值
未提供驗證數據集時套用。
整數 (int)

ImageSweepSettings (圖像掃描設定)

名稱 說明 價值觀
early終止 早期終止原則的類型。 提前終止政策
sampling演算法 [必要]超參數取樣演算法的類型。 “貝氏”
'網格'
'Random' (必要)

JobBaseProperties

名稱 說明 價值觀
元件ID 元件資源的 ARM 資源識別碼。 字串
計算ID 計算資源的 ARM 資源識別碼。 字串
說明 資產描述文字。 字串
顯示名稱 工作的顯示名稱。 字串
實驗名稱 作業所屬的實驗名稱。 如果未設定,作業會放在 「默認」實驗中。 字串
身分識別 身分識別設定。 如果設定,這應該是其中一個 AmlToken、ManagedIdentity、UserIdentity 或 null。
如果為 null,則預設為 AmlToken。
身份配置
isArchived 已存檔 資產是否已封存? 布爾 (bool)
職位類型 設定為 AutoMLJob 類型的 『AutoML』。 設定為 CommandJob 類型的 『Command』。 設定為 FineTuningJob 類型的 『FineTuning』。 設定為 PipelineJob 類型的 'Pipeline'。 針對 SparkJob 類型設定為 『Spark』。 設定為 掃掠Job 類型的 [掃掠]。 'AutoML'
“命令”
“微調”
“管道”
“火花”
'掃掠' (必要)
notification設置 作業的通知設定 通知設置
屬性 資產屬性字典。 ResourceBaseProperties
服務 JobEndpoints 的清單。
針對本機作業,作業端點會有 FileStreamObject 的端點值。
JobBase服務
標記 標記字典。 標記可以新增、移除和更新。 ResourceBaseTags

JobBase服務

名稱 說明 價值觀

作業輸入

名稱 說明 價值觀
說明 輸入的描述。 字串
jobInputType (作業輸入類型) 設定為 CustomModelJobInput 類型的 『custom_model』。 設定為 常值JobInput 類型的 『literal』。 設定為 MLFlowModelJobInput 類型的 『mlflow_model』。 設定為 MLTableJobInput 類型的 'mltable'。 設定為 TritonModelJobInput 類型的 『triton_model』。 針對 UriFileJobInput類型設定為 『uri_file』。 針對 UriFolderJobInput 類型設定為 『uri_folder』。 “custom_model”
'字面值'
“mlflow_model”
'mltable'
“triton_model”
“uri_file”
'uri_folder' (必要)

JobOutput (作業輸出)

名稱 說明 價值觀
說明 輸出的描述。 字串
jobOutputType 設定為 CustomModelJobOutput 類型的 『custom_model』。 設定為 MLFlowModelJobOutput 類型的 『mlflow_model』。 設定為 MLTableJobOutput 類型的 'mltable'。 將 設定為 TritonModelJobOutput 類型的 『triton_model』。 針對 UriFileJobOutput類型設定為 『uri_file』。 針對 UriFolderJobOutput類型設定為 『uri_folder』。 “custom_model”
“mlflow_model”
'mltable'
“triton_model”
“uri_file”
'uri_folder' (必要)

JobResourceConfiguration

名稱 說明 價值觀
dockerArgs 傳遞至 Docker run 命令的額外自變數。 這會覆寫系統或本節中已設定的任何參數。 此參數僅支援 Azure ML 計算類型。 字串
實例數量 計算目標所使用的實例或節點選擇性數目。 整數 (int)
instanceType 計算目標所支持的選擇性 VM 類型。 字串
屬性 其他屬性包。 ResourceConfigurationProperties
shmSize Docker 容器的共用記憶體區塊大小。 這應該是 (number)(unit) 的格式,其中數位大於 0,單位可以是 b(位元組)、k(KB)、m(MB)、或 g(GB)。 字串

約束:
圖案 = \d+[bBkKmMgG]

工作資源

名稱 說明 價值觀
instanceTypes 要從中選擇的實例類型清單。 字串[]

工作服務

名稱 說明 價值觀
端點 端點的 URL。 字串
jobServiceType 端點類型。 字串
節點 使用者想要啟動服務的節點。
如果未將 Nodes 設定或設定為 null,則服務只會在領導者節點上啟動。
節點
傳輸埠 端點的埠。 整數 (int)
屬性 在端點上設定的其他屬性。 JobServiceProperties

JobServiceProperties

名稱 說明 價值觀

LiteralJobInput (文字作業輸入)

名稱 說明 價值觀
jobInputType (作業輸入類型) [必要]指定作業的類型。 'literal' (必要)
價值 [必要]輸入的常值。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)

ManagedIdentity (託管身份)

名稱 說明 價值觀
用戶端ID 依用戶端識別元指定使用者指派的身分識別。 若為系統指派,請勿設定此欄位。 字串

約束:
最小長度 = 36
最大長度 = 36
圖案 = ^[0-9a-fA-F]{8}-([0-9a-fA-F]{4}-){3}[0-9a-fA-F]{12}$
身份類型 [必要]指定身分識別架構的類型。 'Managed' (必要)
物件ID 依物件識別元指定使用者指派的身分識別。 若為系統指派,請勿設定此欄位。 字串

約束:
最小長度 = 36
最大長度 = 36
圖案 = ^[0-9a-fA-F]{8}-([0-9a-fA-F]{4}-){3}[0-9a-fA-F]{12}$
資源ID 依 ARM 資源識別碼指定使用者指派的身分識別。 若為系統指派,請勿設定此欄位。 字串

中位數停止政策

名稱 說明 價值觀
policyType [必要]原則設定的名稱 'MedianStopping' (必要)

MLFlowModelJobInput

名稱 說明 價值觀
說明 輸入的描述。 字串
jobInputType (作業輸入類型) [必要]指定作業的類型。 “custom_model”
'字面值'
“mlflow_model”
'mltable'
“triton_model”
“uri_file”
'uri_folder' (必要)
模式 輸入資產傳遞模式。 “直接”
'下載'
'EvalDownload'
'EvalMount'
'ReadOnlyMount'
“ReadWriteMount”
統一資源識別碼 (URI) [必要]輸入資產 URI。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)

MLFlowModelJobInput

名稱 說明 價值觀
jobInputType (作業輸入類型) [必要]指定作業的類型。 'mlflow_model' (必要)
模式 輸入資產傳遞模式。 “直接”
'下載'
'EvalDownload'
'EvalMount'
'ReadOnlyMount'
“ReadWriteMount”
統一資源識別碼 (URI) [必要]輸入資產 URI。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)

MLFlowModelJobOutput

名稱 說明 價值觀
jobOutputType [必要]指定作業的類型。 'mlflow_model' (必要)
模式 輸出資產傳遞模式。 “直接”
“ReadWriteMount”
'上傳'
統一資源識別碼 (URI) 輸出資產 URI。 字串

MLTableJobInput

名稱 說明 價值觀
說明 輸入的描述。 字串
jobInputType (作業輸入類型) [必要]指定作業的類型。 “custom_model”
'字面值'
“mlflow_model”
'mltable'
“triton_model”
“uri_file”
'uri_folder' (必要)
模式 輸入資產傳遞模式。 “直接”
'下載'
'EvalDownload'
'EvalMount'
'ReadOnlyMount'
“ReadWriteMount”
統一資源識別碼 (URI) [必要]輸入資產 URI。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)

MLTableJobInput

名稱 說明 價值觀
jobInputType (作業輸入類型) [必要]指定作業的類型。 'mltable' (必要)
模式 輸入資產傳遞模式。 “直接”
'下載'
'EvalDownload'
'EvalMount'
'ReadOnlyMount'
“ReadWriteMount”
統一資源識別碼 (URI) [必要]輸入資產 URI。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)

MLTableJobOutput

名稱 說明 價值觀
jobOutputType [必要]指定作業的類型。 'mltable' (必要)
模式 輸出資產傳遞模式。 “直接”
“ReadWriteMount”
'上傳'
統一資源識別碼 (URI) 輸出資產 URI。 字串

MPI

名稱 說明 價值觀
distributionType [必要]指定散發架構的類型。 'Mpi' (必要)
processCountPerInstance 每個 MPI 節點的進程數目。 整數 (int)

NCross驗證

名稱 說明 價值觀
模式 針對 AutoNCrossValidations 類型,設定為 『Auto』。 設定為 CustomNCrossValidations 類型的 [自定義]。 '自動'
'Custom' (必要)

NlpVerticalFeaturizationSettings

名稱 說明 價值觀
datasetLanguage 數據集語言,適用於文字數據。 字串

NlpVerticalLimitSettings

名稱 說明 價值觀
maxConcurrentTrials 並行 AutoML 反覆專案上限。 整數 (int)
maxTrials AutoML 反覆項目的數目。 整數 (int)
超時 AutoML 作業逾時。 字串

節點

名稱 說明 價值觀
nodesValueType 針對 AllNodes 類型,設定為 'All'。 'All' (必要)

通知設置

名稱 說明 價值觀
電子郵件 在指定的通知類型上傳送電子郵件通知給使用者 包含任何的字串數組:
'JobCancelled' (工作已取消)
'JobCompleted'
“作業失敗”
電子郵件 這是電子郵件收件者清單,其總限製為499個字元,且具有逗號分隔符 字串[]
webhook 將 Webhook 回呼傳送至服務。 機碼是 Webhook 的使用者提供名稱。 NotificationSettingWebhook

NotificationSettingWebhook

名稱 說明 價值觀

目的

名稱 說明 價值觀
目標 [必要]定義超參數微調支援的計量目標 “最大化”
'最小化' (必要)
primaryMetric [必要]要優化之計量的名稱。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)

管道作業

名稱 說明 價值觀
輸入 管線作業的輸入。 管道作業輸入
工作 作業會建構管線作業。 PipelineJobJob 作業
職位類型 [必要]指定作業的類型。 'Pipeline' (必要)
輸出 管線作業的輸出 PipelineJobOutputs
設定 管線設定,適用於 ContinueRunOnStepFailure 等專案。 任何
源作業 ID 來源作業的 ARM 資源識別碼。 字串

管道作業輸入

名稱 說明 價值觀

PipelineJobJob 作業

名稱 說明 價值觀

PipelineJobOutputs

名稱 說明 價值觀

PyTorch

名稱 說明 價值觀
distributionType [必要]指定散發架構的類型。 'PyTorch' (必要)
processCountPerInstance 每個節點的進程數目。 整數 (int)

佇列設置

名稱 說明 價值觀
jobTier 控制計算作業層 “基本”
'Null'
“高級”
'Spot'
“標準”

RandomSamplingAlgorithm 演算法

名稱 說明 價值觀
規則 隨機演算法的特定類型 '隨機'
“索博爾”
採樣演算法類型 [必要]用來產生超參數值的演算法,以及組態屬性 'Random' (必要)
種子 要作為隨機數產生種子的選擇性整數 整數 (int)

迴歸

名稱 說明 價值觀
cvSplitColumnNames 要用於 CVSplit 資料的數據行。 字串[]
featurizationSettings AutoML 作業所需的特徵化輸入。 TableVerticalFeaturizationSettings
limitSettings AutoMLJob 的執行條件約束。 TableVerticalLimitSettings
nCrossValidations 要套用至定型數據集的交叉驗證折疊數目
未提供驗證數據集時。
NCross驗證
primaryMetric 回歸工作的主要計量。 'NormalizedMeanAbsoluteError'
'NormalizedRootMeanSquaredError'
'R2Score'
'SpearmanCorrelation'
任務類型 [必要]AutoMLJob 的工作類型。 '回歸' (必要)
測試數據 測試數據輸入。 MLTableJobInput
testDataSize 必須為驗證目的而保留的測試數據集分數。
介於 (0.0 、 1.0) 之間的值
未提供驗證數據集時套用。
整數 (int)
trainingSettings AutoML 作業定型階段的輸入。 回歸訓練設置
validationData (驗證資料) 驗證數據輸入。 MLTableJobInput
validationDataSize 需要為驗證目的預留的定型數據集分數。
介於 (0.0 、 1.0) 之間的值
未提供驗證數據集時套用。
整數 (int)
權重列名稱 範例權數數據行的名稱。 自動化 ML 支援加權數據行做為輸入,導致數據中的數據列向上或向下加權。 字串

回歸訓練設置

名稱 說明 價值觀
allowedTrainingAlgorithms 回歸工作的允許模型。 包含任何的字串數組:
“決策樹”
“彈性網”
'ExtremeRandomTrees'
'梯度提升'
'KNN'
“套索拉斯”
'LightGBM'
'隨機森林'
'新幣'
'XGBoostRegressor'
blockedTrainingAlgorithms 回歸工作的封鎖模型。 包含任何的字串數組:
“決策樹”
“彈性網”
'ExtremeRandomTrees'
'梯度提升'
'KNN'
“套索拉斯”
'LightGBM'
'隨機森林'
'新幣'
'XGBoostRegressor'
enableDnnTraining 啟用 DNN 模型的建議。 布爾 (bool)
enableModelExplainability 在最佳模型上開啟可解釋性的旗標。 布爾 (bool)
enableOnnxCompatibleModels 啟用 onnx 相容模型的旗標。 布爾 (bool)
enableStackEnsemble 啟用堆疊合奏執行。 布爾 (bool)
enableVoteEnsemble 啟用投票合奏執行。 布爾 (bool)
ensembleModelDownloadTimeout 在 VotingEnsemble 和 StackEnsemble 模型產生期間,會下載先前子回合中的多個安裝模型。
如果需要更多時間,請使用高於 300 秒的值來設定此參數。
字串
stackEnsembleSettings 堆疊合奏執行的堆疊合奏設定。 StackEnsembleSettings

ResourceBaseProperties

名稱 說明 價值觀

ResourceBaseTags

名稱 說明 價值觀

ResourceConfigurationProperties

名稱 說明 價值觀

採樣演算法

名稱 說明 價值觀
採樣演算法類型 針對 BayesianSamplingAlgorithm類型設定為 『Bayesian』。 設定為 GridSamplingAlgorithm 類型的 『Grid』。 設定為 RandomSamplingAlgorithm 類型的 『Random』。 “貝氏”
'網格'
'Random' (必要)

季節性

名稱 說明 價值觀
模式 針對 AutoSeasonality 類型,設定為 [自動]。 針對 CustomSeasonality 類型,設定為 [自定義]。 '自動'
'Custom' (必要)

SparkJob

名稱 說明 價值觀
檔案 封存作業中使用的檔案。 字串[]
參數 作業的自變數。 字串
代碼Id [必要] 程式代碼資產的arm識別碼。 字串 (必要)
會議 Spark 設定的屬性。 SparkJobConf 函數
輸入 [必要]在作業啟動時要執行的專案。 SparkJobEntry (必要)
環境ID 作業之環境規格的 ARM 資源識別碼。 字串
環境變數 作業中包含的環境變數。 SparkJobEnvironment變數
檔案 作業中使用的檔案。 字串[]
輸入 對應作業中使用的輸入數據系結。 SparkJobInputs
罐子 作業中使用的 Jar 檔案。 字串[]
職位類型 [必要]指定作業的類型。 'Spark' (必要)
輸出 對應作業中使用的輸出數據系結。 SparkJob輸出
py檔 作業中使用的 Python 檔案。 字串[]
queueSettings 作業的佇列設定 佇列設置
資源 作業的計算資源組態。 SparkResourceConfiguration

SparkJobConf 函數

名稱 說明 價值觀

SparkJobEntry

名稱 說明 價值觀
sparkJobEntryType 針對 SparkJobPythonEntry 類型,設定為 'SparkJobPythonEntry'。 針對 SparkJobScalaEntry 類型,設定為 'SparkJobScalaEntry'。 “SparkJobPythonEntry”
'SparkJobScalaEntry' (必要)

SparkJobEnvironment變數

名稱 說明 價值觀

SparkJobInputs

名稱 說明 價值觀

SparkJob輸出

名稱 說明 價值觀

SparkJobPython條目

名稱 說明 價值觀
檔案 [必要]作業進入點的相對 Python 檔案路徑。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)
sparkJobEntryType [必要]作業進入點的類型。 'SparkJobPythonEntry' (必要)

SparkJobScalaEntry

名稱 說明 價值觀
類別名稱 [必要]做為進入點的 Scala 類別名稱。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)
sparkJobEntryType [必要]作業進入點的類型。 'SparkJobScalaEntry' (必要)

SparkResourceConfiguration

名稱 說明 價值觀
instanceType 計算目標所支持的選擇性 VM 類型。 字串
runtimeVersion (運行時版本) 用於作業的 Spark 執行時間版本。 字串

StackEnsembleSettings

名稱 說明 價值觀
stackMetaLearnerKWargs 要傳遞至中繼學習模組初始化表達式的選擇性參數。 任何
stackMetaLearnerTrainPercentage 指定要保留用於訓練中繼學習工具的定型集比例(選擇定型和定型類型的訓練類型)。 預設值為 0.2。 整數 (int)
stackMetaLearnerType 中繼學習器是針對個別異質模型輸出定型的模型。 “彈性網”
'ElasticNetCV'
'LightGBMClassifier'
'LightGBMRegressor'
'線性回歸'
'LogisticRegression' (邏輯回歸)
'LogisticRegressionCV'
“無”

SweepJob (掃描作業)

名稱 說明 價值觀
early終止 早期終止原則可在完成之前取消執行效能不佳 提前終止政策
輸入 對應作業中使用的輸入數據系結。 SweepJobInputs
職位類型 [必要]指定作業的類型。 '掃掠' (必要)
限制 掃掠作業限制。 SweepJobLimits (掃描作業限制)
目的 [必要]優化目標。 目標 (必要)
輸出 對應作業中使用的輸出數據系結。 SweepJobOutputs (掃描作業輸出)
queueSettings 作業的佇列設定 佇列設置
sampling演算法 [必要]超參數取樣演算法 SamplingAlgorithm (必要)
搜索空間 [必要]包含每個參數及其散發的字典。 字典索引鍵是參數的名稱 任何 (必要)
試用 [必要]試用版元件定義。 TrialComponent (必要)

SweepJobInputs

名稱 說明 價值觀

SweepJobLimits (掃描作業限制)

名稱 說明 價值觀
jobLimits類型 [必要]JobLimit 類型。 “命令”
'掃掠' (必要)
maxConcurrentTrials 掃掠作業最大並行試用版。 整數 (int)
maxTotalTrials 掃掠作業最大總試用版。 整數 (int)
超時 ISO 8601 格式的最大執行持續時間,之後作業將會取消。 僅支援有效位數為秒數的持續時間。 字串
trialTimeout (試用超時) 掃掠作業試用版逾時值。 字串

SweepJobOutputs (掃描作業輸出)

名稱 說明 價值觀

TableVerticalFeaturizationSettings

名稱 說明 價值觀
阻塞變壓器 這些轉換器不得用於特徵化。 包含任何的字串數組:
'CatTargetEncoder' (貓目標編碼器)
'CountVectorizer'
'HashOneHotEncoder'
'標籤編碼器'
“天真貝葉斯”
'OneHotEncoder' (一熱編碼器)
'TextTargetEncoder'
'TfIdf'
'WoETargetEncoder' (WoETarget編碼器)
'WordEmbedding' (單詞嵌入)
columnNameAnd類型 數據行名稱及其類型字典(int、float、string、datetime 等)。 TableVerticalFeaturizationSettingsColumnNameAndTypes
datasetLanguage 數據集語言,適用於文字數據。 字串
enableDnnFeaturization 判斷是否要使用 Dnn 型特徵化程式進行數據特徵化。 布爾 (bool)
模式 特徵化模式 - 使用者可以保留預設的「自動」模式,而 AutoML 會負責在特徵化階段中轉換數據的必要轉換。
如果選取 [關閉],則不會完成特徵化。
如果選取 [自定義],則使用者可以指定其他輸入,以自定義特徵化的完成方式。
'自動'
“自定義”
“關閉”
transformer參數 用戶可以指定要搭配套用的其他轉換器,以及要套用它的數據行,以及轉換建構函式的參數。 TableVerticalFeaturizationSettingsTransformerParams

TableVerticalFeaturizationSettingsColumnNameAndTypes

名稱 說明 價值觀

TableVerticalFeaturizationSettingsTransformerParams

名稱 說明 價值觀

TableVerticalLimitSettings

名稱 說明 價值觀
enableEarlyTermination 啟用提早終止,判斷如果過去 20 個反覆項目沒有分數改善,AutoMLJob 是否會提早終止。 布爾 (bool)
退出評分 AutoML 作業的結束分數。 整數 (int)
maxConcurrentTrials 並行反覆運算數上限。 整數 (int)
maxCoresPerTrial 每個反覆專案的核心數上限。 整數 (int)
maxTrials 反覆項目的數目。 整數 (int)
超時 AutoML 作業逾時。 字串
trialTimeout (試用超時) 反覆專案逾時。 字串

目標滯後

名稱 說明 價值觀
模式 針對 AutoTargetLags 類型設定為 『Auto』。 設定為 CustomTargetLags 類型的 [自定義]。 '自動'
'Custom' (必要)

TargetRollingWindowSize

名稱 說明 價值觀
模式 針對 AutoTargetRollingWindowSize 類型,設定為 『Auto』。 針對 CustomTargetRollingWindowSize 類型設定為 『Custom』。 '自動'
'Custom' (必要)

TensorFlow

名稱 說明 價值觀
distributionType [必要]指定散發架構的類型。 'TensorFlow' (必要)
parameterServerCount 參數伺服器工作的數目。 整數 (int)
workerCount 工人計數 工人人數。 如果未指定,則會預設為實例計數。 整數 (int)

文本分類

名稱 說明 價值觀
featurizationSettings AutoML 作業所需的特徵化輸入。 NlpVerticalFeaturizationSettings
limitSettings AutoMLJob 的執行條件約束。 NlpVerticalLimitSettings
primaryMetric Text-Classification 工作的主要計量。 '精確度'
'AUCWeighted'
'AveragePrecisionScoreWeighted'
'NormMacroRecall' (規範巨集調用)
'PrecisionScoreWeighted'
任務類型 [必要]AutoMLJob 的工作類型。 'TextClassification' (必要)
validationData (驗證資料) 驗證數據輸入。 MLTableJobInput

TextClassificationMultilabel

名稱 說明 價值觀
featurizationSettings AutoML 作業所需的特徵化輸入。 NlpVerticalFeaturizationSettings
limitSettings AutoMLJob 的執行條件約束。 NlpVerticalLimitSettings
任務類型 [必要]AutoMLJob 的工作類型。 'TextClassificationMultilabel' (必要)
validationData (驗證資料) 驗證數據輸入。 MLTableJobInput

TextNer (文字產生器)

名稱 說明 價值觀
featurizationSettings AutoML 作業所需的特徵化輸入。 NlpVerticalFeaturizationSettings
limitSettings AutoMLJob 的執行條件約束。 NlpVerticalLimitSettings
任務類型 [必要]AutoMLJob 的工作類型。 'TextNER' (必要)
validationData (驗證資料) 驗證數據輸入。 MLTableJobInput

TrialComponent (試用元件)

名稱 說明 價值觀
代碼Id 程式代碼資產的 ARM 資源識別碼。 字串
指令 [必要]在作業啟動時執行的命令。 例如 “python train.py” 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)
分配 作業的散發組態。 如果設定,這應該是其中一個 Mpi、Tensorflow、PyTorch 或 null。 分發配置
環境ID [必要]作業之環境規格的 ARM 資源識別碼。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)
環境變數 作業中包含的環境變數。 TrialComponentEnvironmentVariables
資源 作業的計算資源組態。 JobResourceConfiguration

TrialComponentEnvironmentVariables

名稱 說明 價值觀

TritonModelJobInput

名稱 說明 價值觀
jobInputType (作業輸入類型) [必要]指定作業的類型。 'triton_model' (必要)
模式 輸入資產傳遞模式。 “直接”
'下載'
'EvalDownload'
'EvalMount'
'ReadOnlyMount'
“ReadWriteMount”
統一資源識別碼 (URI) [必要]輸入資產 URI。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)

TritonModelJobOutput

名稱 說明 價值觀
jobOutputType [必要]指定作業的類型。 'triton_model' (必要)
模式 輸出資產傳遞模式。 “直接”
“ReadWriteMount”
'上傳'
統一資源識別碼 (URI) 輸出資產 URI。 字串

截斷選擇策略 (TruncationSelectionPolicy)

名稱 說明 價值觀
policyType [必要]原則設定的名稱 'TruncationSelection' (必要)
截斷百分比 要在每個評估間隔取消的執行百分比。 整數 (int)

UriFileJobInput

名稱 說明 價值觀
jobInputType (作業輸入類型) [必要]指定作業的類型。 'uri_file' (必要)
模式 輸入資產傳遞模式。 “直接”
'下載'
'EvalDownload'
'EvalMount'
'ReadOnlyMount'
“ReadWriteMount”
統一資源識別碼 (URI) [必要]輸入資產 URI。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)

UriFileJobOutput

名稱 說明 價值觀
jobOutputType [必要]指定作業的類型。 'uri_file' (必要)
模式 輸出資產傳遞模式。 “直接”
“ReadWriteMount”
'上傳'
統一資源識別碼 (URI) 輸出資產 URI。 字串

UriFolderJobInput

名稱 說明 價值觀
jobInputType (作業輸入類型) [必要]指定作業的類型。 'uri_folder' (必要)
模式 輸入資產傳遞模式。 “直接”
'下載'
'EvalDownload'
'EvalMount'
'ReadOnlyMount'
“ReadWriteMount”
統一資源識別碼 (URI) [必要]輸入資產 URI。 字串

約束:
最小長度 = 1
模式 = [a-zA-Z0-9_] (必要)

UriFolderJobOutput

名稱 說明 價值觀
jobOutputType [必要]指定作業的類型。 'uri_folder' (必要)
模式 輸出資產傳遞模式。 “直接”
“ReadWriteMount”
'上傳'
統一資源識別碼 (URI) 輸出資產 URI。 字串

使用者身份

名稱 說明 價值觀
身份類型 [必要]指定身分識別架構的類型。 'UserIdentity' (必要)

Webhook

名稱 說明 價值觀
事件類型 在指定的通知事件上傳送回呼 字串
webhook類型 針對 AzureDevOpsWebhook 類型,設定為 『AzureDevOps』。 'AzureDevOps' (必要)