The ML.TRIAL_INFO function

This document describes the ML.TRIAL_INFO function, which lets you display information about trials from a model that uses hyperparameter tuning.

You can use this function with models that support hyperparameter tuning. For more information, see End-to-end user journeys for ML models.

Syntax

ML.TRIAL_INFO(MODEL `PROJECT_ID.DATASET.MODEL_NAME`)

Arguments

ML.TRIAL_INFO takes the following arguments:

  • PROJECT_ID: your project ID.
  • DATASET: the BigQuery dataset that contains the model.
  • MODEL_NAME: The name of the model.

Output

ML.TRIAL_INFO returns one row per trial with the following columns:

  • trial_id: an INT64 value that contains the ID assigned to each trial in the approximate order of trial execution. trial_id values start from 1.
  • hyperparameters: a STRUCT value that contains the hyperparameters used in the trial.
  • hparam_tuning_evaluation_metrics: a STRUCT value that contains the evaluation metrics appropriate to the hyperparameter tuning objective specified by the hparam_tuning_objectives argument in the CREATE MODEL statement. Metrics are calculated from the evaluation data. For more information about the datasets used in hyperparameter tuning, see Data split.
  • training_loss: a FLOAT64 value that contains the loss of the trial during the last iteration, calculated using the training data.
  • eval_loss: a FLOAT64 value that contains the loss of the trial during the last iteration, calculated using the evaluation data.

  • status: a STRING value that contains the final status of the trial. Possible values include the following:

    • SUCCEEDED: the trial succeeded.
    • FAILED: the trial failed.
    • INFEASIBLE: the trial was not run due to an invalid combination of hyperparameters.

  • error_message: a STRING value that contains the error message that is returned if the trial didn't succeed. For more information, see Error handling.

  • is_optimal: a BOOL value that indicates whether the trial had the best objective value. If multiple trials are marked as optimal, then the trial with the smallest trial_id value is used as the default trial during model serving.

Example

The following query retrieves information of all trials for the model mydataset.mymodel in your default project:

SELECT
  *
FROM
  ML.TRIAL_INFO(MODEL `mydataset.mymodel`)

What's next