The ML.TRAINING_INFO function

This document describes the ML.TRAINING_INFO function, which lets you see information about the training iterations of a model.

You can run ML.TRAINING_INFO while the CREATE MODEL statement for the target model is running, or you can wait until after the CREATE MODEL statement completes. If you run ML.TRAINING_INFO before the first training iteration of the CREATE MODEL statement completes, the query returns a Not found error.

Syntax

ML.TRAINING_INFO(MODEL `project_id.dataset.model`)

Arguments

ML.TRAINING_INFO takes the following arguments:

• project_id: Your project ID.
• dataset: The BigQuery dataset that contains the model.
• model: The name of the model.

Output

ML.TRAINING_INFO returns the following columns:

• training_run: an INT64 value that contains the training run identifier for the model. The value in this column is 0 for a newly created model. If you retrain the model using the warm_start argument of the CREATE MODEL statement, this value is incremented.
• iteration: an INT64 value that contains the iteration number of the training run. The value for the first iteration is 0. This value is incremented for each additional training run.

• loss: a FLOAT64 value that contains the loss metric calculated after an iteration on the training data:

  • For logistic regression models, this is log loss.
  • For linear regression models, this is mean squared error.
  • For multiclass logistic regressions, this is cross-entropy log loss.
  • For explicit matrix factorization models this is mean squared error calculated over the seen input ratings.
  • For implicit matrix factorization models, the loss is calculated using the following formula:
  $$Loss = \sum_{u, i} c_{ui}(p_{ui} - x^T_uy_i)^2 + \lambda(\sum_u||x_u||^2 + \sum_i||y_i||^2)$$

  For more information about what the variables mean, see Feedback types.

• eval_loss: a FLOAT64 value that contains the loss metric calculated on the holdout data. For k-means models, ML.TRAINING_INFO doesn't return an eval_loss column. If the DATA_SPLIT_METHOD argument is NO_SPLIT, then all entries in the eval_loss column are NULL.

• learning_rate: a FLOAT64 value that contains the learning rate in this iteration.

• duration_ms: an INT64 value that contains how long the iteration took, in milliseconds.

• cluster_info: an ARRAY<STRUCT> value that contains the fields centroid_id, cluster_radius, and cluster_size. ML.TRAINING_INFO computes cluster_radius and cluster_size with standardized features. Only returned for k-means models.

Note: For linear and logistic regression models, the learning_rate value can increase over the course of training if the LEARN_RATE_STRATEGY argument isn't set to CONSTANT. This increase is due to the fact that when the LEARN_RATE_STRATEGY is set to LINE_SEARCH, four learning rates are generated from LS_INIT_LEARN_RATE. For example, if LS_INIT_LEARN_RATE is 0.1, then the first iteration of training compares the respective loss from a set of four different models trained from setting the learn rate to 0.2, 0.1, 0.05, and 0.025. If LEARN_RATE=0.2 generates the model with the smallest loss, then the next iteration generates four models with the learning rate set to 0.4, 0.2, 0.1, 0.05 from the previous best fit model, and this process continues until the model converges.

Permissions

You must have the bigquery.models.create and bigquery.models.getData Identity and Access Management (IAM) permissions in order to run ML.TRAINING_INFO.

Limitations

ML.TRAINING_INFO is subject to the following limitations:

• ML.TRAINING_INFO doesn't support imported TensorFlow models.
• For time series models, ML.TRAINING_INFO only returns three columns: training_run, iteration, and duration_ms. It doesn't expose the training information per iteration, or per time series if multiple time series are forecasted at once. The duration_ms is the total time cost for the entire process.

Example

The following example retrieves training information from the model mydataset.mymodel in your default project:

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

What's next

• For information about model evaluation, see BigQuery ML model evaluation overview.
• For information about the supported SQL statements and functions for each model type, see End-to-end user journey for each model.