Automl SDK 代码具有来自 bigquery 的文件位置，但在预测时出现问题

Posted 2023-03-29

【中文标题】Automl SDK 代码具有来自 bigquery 的文件位置，但在预测时出现问题

【英文标题】：Automl SDK code with file location from bigquery but having issue while predicting

【发布时间】：2021-10-09 12:59:30

【问题描述】：

我正在使用 AI Platform 中的 GCP automl sdk 模块创建模型（使用 bigquery 表作为训练和预测的输入）并使用 batch_prediction 进行预测。问题是代码运行良好，但预测的输出表为空，错误表包含来自预测数据帧和附加列的所有系列，说明错误代码 3，错误是“时间序列没有要预测的值。时间序列已经从预测中排除。”。

我用于模型训练的代码：

job = aiplatform.AutoMLForecastingTrainingJob(
    display_name='train-sdk-automl_tst1',
    optimization_objective='minimize-mae',    
    column_transformations=[
        "timestamp": "column_name": "Date",
        "numeric": "column_name": "Price",
        "numeric": "column_name": "Grammage",
        "numeric": "column_name": "apparentTemperatureMax",
        "numeric": "column_name": "apparentTemperatureMin",
        "numeric": "column_name": "Consumer_promo",
        "numeric": "column_name": "Promo_Value",
        "numeric": "column_name": "Trade_Promotion",
        "numeric": "column_name": "Holiday",
        "numeric": "column_name": "Sales",
    ]
)

# This will take around an hour to run
my_model = job.run(
    dataset=ds,
    target_column='Sales',
    time_column='Date',
    time_series_identifier_column='SKU',
    available_at_forecast_columns=['Date', 'Price','Grammage'
                                   ,'apparentTemperatureMax','apparentTemperatureMin','Consumer_promo',
                                   "Promo_Value","Trade_Promotion","Holiday"],
    unavailable_at_forecast_columns=['Sales'],
    forecast_horizon=21.0,
    data_granularity_unit='week',
    data_granularity_count=1,
    weight_column=None,
    budget_milli_node_hours=1000,
    model_display_name='sdk_tsting_bq-forecast-model', 
    predefined_split_column_name=None
)

预测代码：

BATCH_PREDICT_SOURCE = 'bq://acn-intelligent-supply-chain.scoa_ml_forecast_tool.test_data_sdk1'
BATCH_PREDICT_DESTINATION_PREFIX = 'bq://acn-intelligent-supply-chain.scoa_ml_forecast_tool' 
my_model.batch_predict(
   bigquery_source=BATCH_PREDICT_SOURCE,
   instances_format='bigquery',
   bigquery_destination_prefix = BATCH_PREDICT_DESTINATION_PREFIX,
   predictions_format='bigquery',
   job_display_name='predict_sdk_tst')

请提出这里可能出现的问题。

【参考方案1】：

我认为用于批量预测的 BQ 表需要在目标（例如“销售”）列中具有 NULL 值。如果您还没有这样做，请尝试一下，看看是否有帮助。