3.14.2.1 ONNX Model Metadata

For model_metadata, see: Section 3.14.1.3, “Model Metadata”. This includes onnx_inputs_info and onnx_outputs_info.

onnx_inputs_info includes data_types_map. See Section 3.14.1.3, “Model Metadata” for the default value.

onnx_outputs_info includes predictions_name, prediction_probabilities_name, and labels_map.

Use the data_types_map to map the data type of each column to an ONNX model data type. For example, to convert inputs of the type tensor(float) to float64:

data_types_map = {"tensor(float)": "float64"}

HeatWave AutoML first checks the user data_types_map, and then the default data_types_map to check if the data type exists. HeatWave AutoML supports the following numpy data types:

Table 3.1 Supported numpy data types

str_	unicode_	int8	int16	int32	int64	int_	uint16
uint32	uint64	byte	ubyte	short	ushort	intc	uintc
uint	longlong	ulonglong	intp	uintp	float16	float32	float64
half	single	longfloat	double	longdouble	bool_	datetime64	complex_
complex64	complex128	complex256	csingle	cdouble	clongdouble

The use of any other numpy data type will cause an error.

Use predictions_name to determine which of the ONNX model outputs is associated with predictions. Use prediction_probabilities_name to determine which of the ONNX model outputs is associated with prediction probabilities. Use use a labels_map to map prediction probabilities to predictions, known as labels.

For regression tasks, if the ONNX model generates only one output, then predictions_name is optional. If the ONNX model generates more than one output, then predictions_name is required. Do not provide prediction_probabilities_name as this will cause an error.

For classification tasks use predictions_name or prediction_probabilities_name or both. Failure to provide at least one will cause an error. The model explainers SHAP, Fast SHAP and Partial Dependence require prediction_probabilities_name.

Only use a labels_map with classification tasks. A labels_map requires predictions_probabilities_name. The use of a labels_map with any other task, or with predictions_name or without predictions_probabilities_name will cause an error.

An example of a predictions_probabilities_name with a labels_map produces these labels:

predictions_probabilities_name = array([[0.35, 0.50, 0.15],
                                        [0.10, 0.20, 0.70],
                                        [0.90, 0.05, 0.05],
                                        [0.55, 0.05, 0.40]], dtype=float32)
  
labels_map = {0:'Iris-virginica', 1:'Iris-versicolor', 2:'Iris-setosa'}

labels=['Iris-versicolor', 'Iris-setosa', 'Iris-virginica', 'Iris-virginica']

Do not provide predictions_name or prediction_probabilities_name when the task is NULL as this will cause an error.

HeatWave AutoML adds a note for ONNX models that have inputs with four dimensions about the reshaping of data to a suitable shape for an ONNX model. This would typically be for ONNX models that are trained on image data. An example of this note added to the ml_results column:

mysql> CALL sys.ML_PREDICT_TABLE('mlcorpus_v5.mnist_test_temp', @model, 
          'mlcorpus_v5.`mnist_predictions`', NULL);
Query OK, 0 rows affected (20.6296 sec)

mysql> SELECT ml_results FROM mnist_predictions;;
+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| ml_results                                                                                                                                                                                                                                                                                                                          |
+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| {'predictions': {'prediction': 7}, 'Notes': 'Input data is reshaped into (1, 28, 28).', 'probabilities': {0: -552.7100219726562, 1: 138.27000427246094, 2: 2178.510009765625, 3: 2319.860107421875, 4: -3466.5400390625, 5: -1778.3499755859375, 6: -6441.83984375, 7: 8062.9599609375, 8: -1860.2099609375, 9: 1034.239990234375}} |
+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+