Running the ML_MODEL_IMPORT
routine imports a pre-trained model in ONNX (Open Neural
Network Exchange) format into the model catalog. After import,
all the HeatWave AutoML routines can be used with the ONNX model.
ML_MODEL_IMPORT
is available as
of MySQL 8.0.31.
Models in ONNX format (.onnx
) cannot be
loaded directly into a MySQL table. They require string
serialization and conversion to Base64 binary encoding. Before
running ML_MODEL_IMPORT
, follow
the instructions in Section 3.10.2, “Importing ONNX Models” to
carry out the required pre-processing and then load the model
into a temporary table for import to MySQL HeatWave.
mysql> CALL sys.ML_MODEL_IMPORT (model_object, model_metadata,
model_handle_variable);
ML_MODEL_IMPORT
parameters:
model_object
: The preprocessed ONNX model object, which must be string serialized and BASE64 encoded. Follow the instructions in Section 3.10.2, “Importing ONNX Models” to do this.-
model_metadata
: An optional JSON object containing key-value pairs with model metadata. UseNULL
if you do not want to specify model metadata. For ONNX models, you can add the following optional model metadata:-
training_score
: any real number including zero The cross-validation score achieved for the model by training.
-
task
: string The task type for the model -
classification
,regression
, orforecasting
.-
n_rows
: integer greater than 0 The number of rows in the training table.
-
n_columns
: integer greater than 0 The number of columns in the training table.
-
algorithm_name
: string The name of the chosen algorithm.
-
training_time
: float greater than 0 The time in seconds taken to train the model.
-
target_column_name
: string The name of the target column in the training table.
-
train_table_name
: string The name of the training table.
-
column_names
: JSON object A JSON object containing a list of the column names in the training table.
-
notes
: string Your own notes on the model.
-
model_handle
: The model handle for the ONNX model. The model is stored in the model catalog under this name and accessed using it. Specify a model handle that does not already exist in the model catalog.