3.14.1.3 Model Metadata

Metadata for the model. It is a column in the model catalog, see Section 3.14.1.1, “The Model Catalog Table”, and a parameter in ML_MODEL_IMPORT. The default value for model_metadata is NULL.

model_metadata has several fields that replace deprecated columns in the model catalog, and fields that support ONNX model import, see: Section 3.14.2, “ONNX Model Import”.

model_metadata contains the following metadata as key-value pairs in JSON format:

• task: string

  The task type specified in the ML_TRAIN query. The default is classification when used with ML_MODEL_IMPORT.

• build_timestamp: number

  A timestamp indicating when the model was created, in UNIX epoch time. A model is created when the ML_TRAIN routine finishes executing.

• target_column_name: string

  The name of the column in the training table that was specified as the target column.

• train_table_name: string

  The name of the input table specified in the ML_TRAIN query.

• column_names: JSON array

  The feature columns used to train the model.

• model_explanation: JSON object literal

  The model explanation generated during training. See Section 3.14.7, “Model Explanations”.

• notes: string

  The notes specified in the ML_TRAIN query. It also records any error messages that occur during model training.

• format: string

  The model serialization format. HWMLv1.0 for a HeatWave AutoML model or ONNX for a ONNX model. The default is ONNX when used with ML_MODEL_IMPORT.

• status: string

  The status of the model. The default is Ready when used with ML_MODEL_IMPORT.

  • Creating

    The model is still being created.

  • Ready

    The model is trained and active.

  • Error

    Either training was canceled or an error occurred during training. Any error message appears in the notes column. The error message also appears in model_metadata notes.

• model_quality: string

  The quality of the model object. Either low or high.

• training_time: number

  The time in seconds taken to train the model.

• algorithm_name: string

  The name of the chosen algorithm.

• training_score: number

  The cross-validation score achieved for the model by training.

• n_rows: number

  The number of rows in the training table.

• n_columns: number

  The number of columns in the training table.

• n_selected_rows: number

  The number of rows selected by adaptive sampling.

• n_selected_columns: number

  The number of columns selected by feature selection.

• optimization_metric: string

  The optimization metric used for training.

• selected_column_names: JSON array

  The names of the columns selected by feature selection.

• contamination: number

  The contamination factor for a model.

• options: JSON object literal

  The options specified in the ML_TRAIN query.

• training_params: JSON object literal

  Internal task dependent parameters used during ML_TRAIN.

• onnx_inputs_info: JSON object literal

  Information about the format of the ONNX model inputs. This only applies to ONNX models. See Section 3.14.2, “ONNX Model Import”.

  Do not provide onnx_inputs_info if the model is not ONNX format. This will cause an error.

  • data_types_map: JSON object literal

    This maps the data type of each column to an ONNX model data type. The default value is:

    JSON_OBJECT("tensor(int64)": "int64", "tensor(float)": "float32", "tensor(string)": "str_")

• onnx_outputs_info: JSON object literal

  Information about the format of the ONNX model outputs. This only applies to ONNX models. See Section 3.14.2, “ONNX Model Import”.

  Do not provide onnx_outputs_info if the model is not ONNX format, or if task is NULL. This will cause an error.

  • predictions_name: string

    This name determines which of the ONNX model outputs is associated with predictions.

  • prediction_probabilities_name: string

    This name determines which of the ONNX model outputs is associated with prediction probabilities.

  • labels_map: JSON object literal

    This maps prediction probabilities to predictions, known as labels.

• training_drift_metric: JSON object literal

  Contains data drift information about the training data, see: Section 3.14.11, “Data Drift Detection”. This only applies to classification and regression models.

  • mean: number

    The mean value of drift metrics of all the training data. ≥ 0.

  • variance: number

    The variance value of drift metrics of all the training data. ≥ 0.

  Both mean and variance should be low. To avoid divide by zero, the lowest value for both is 1e-10.

• chunks: number

  The total number of chunks that the model has been split into. This was added in MySQL 9.0.0.