Deploy and make predictions with an ONNX model and SQL machine learning

Important

Azure SQL Edge will be retired on September 30, 2025. For more information and migration options, see the Retirement notice.

Note

Azure SQL Edge no longer supports the ARM64 platform.

In this quickstart, you'll learn how to train a model, convert it to ONNX, deploy it to Azure SQL Edge, and then run native PREDICT on data using the uploaded ONNX model.

This quickstart is based on scikit-learn and uses the Boston Housing dataset.

Before you begin

  • If you're using Azure SQL Edge, and you haven't deployed an Azure SQL Edge module, follow the steps of deploy SQL Edge using the Azure portal.

  • Install Azure Data Studio.

  • Install Python packages needed for this quickstart:

    1. Open New Notebook connected to the Python 3 Kernel.
    2. Select Manage Packages
    3. In the Installed tab, look for the following Python packages in the list of installed packages. If any of these packages aren't installed, select the Add New tab, search for the package, and select Install.
      • scikit-learn
      • numpy
      • onnxmltools
      • onnxruntime
      • pyodbc
      • setuptools
      • skl2onnx
      • sqlalchemy
  • For each script part in the following sections, enter it in a cell in the Azure Data Studio notebook and run the cell.

Train a pipeline

Split the dataset to use features to predict the median value of a house.

import numpy as np
import onnxmltools
import onnxruntime as rt
import pandas as pd
import skl2onnx
import sklearn
import sklearn.datasets

from sklearn.datasets import load_boston
boston = load_boston()
boston

df = pd.DataFrame(data=np.c_[boston['data'], boston['target']], columns=boston['feature_names'].tolist() + ['MEDV'])

target_column = 'MEDV'

# Split the data frame into features and target
x_train = pd.DataFrame(df.drop([target_column], axis = 1))
y_train = pd.DataFrame(df.iloc[:,df.columns.tolist().index(target_column)])

print("\n*** Training dataset x\n")
print(x_train.head())

print("\n*** Training dataset y\n")
print(y_train.head())

Output:

*** Training dataset x

        CRIM    ZN  INDUS  CHAS    NOX     RM   AGE     DIS  RAD    TAX  \
0  0.00632  18.0   2.31   0.0  0.538  6.575  65.2  4.0900  1.0  296.0
1  0.02731   0.0   7.07   0.0  0.469  6.421  78.9  4.9671  2.0  242.0
2  0.02729   0.0   7.07   0.0  0.469  7.185  61.1  4.9671  2.0  242.0
3  0.03237   0.0   2.18   0.0  0.458  6.998  45.8  6.0622  3.0  222.0
4  0.06905   0.0   2.18   0.0  0.458  7.147  54.2  6.0622  3.0  222.0

    PTRATIO       B  LSTAT
0     15.3  396.90   4.98
1     17.8  396.90   9.14
2     17.8  392.83   4.03
3     18.7  394.63   2.94
4     18.7  396.90   5.33

*** Training dataset y

0    24.0
1    21.6
2    34.7
3    33.4
4    36.2
Name: MEDV, dtype: float64

Create a pipeline to train the LinearRegression model. You can also use other regression models.

from sklearn.compose import ColumnTransformer
from sklearn.linear_model import LinearRegression
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import RobustScaler

continuous_transformer = Pipeline(steps=[('scaler', RobustScaler())])

# All columns are numeric - normalize them
preprocessor = ColumnTransformer(
    transformers=[
        ('continuous', continuous_transformer, [i for i in range(len(x_train.columns))])])

model = Pipeline(
    steps=[
        ('preprocessor', preprocessor),
        ('regressor', LinearRegression())])

# Train the model
model.fit(x_train, y_train)

Check the accuracy of the model and then calculate the R2 score and mean squared error.

# Score the model
from sklearn.metrics import r2_score, mean_squared_error
y_pred = model.predict(x_train)
sklearn_r2_score = r2_score(y_train, y_pred)
sklearn_mse = mean_squared_error(y_train, y_pred)
print('*** Scikit-learn r2 score: {}'.format(sklearn_r2_score))
print('*** Scikit-learn MSE: {}'.format(sklearn_mse))

Output:

*** Scikit-learn r2 score: 0.7406426641094094
*** Scikit-learn MSE: 21.894831181729206

Convert the model to ONNX

Convert the data types to the supported SQL data types. This conversion is required for other dataframes as well.

from skl2onnx.common.data_types import FloatTensorType, Int64TensorType, DoubleTensorType

def convert_dataframe_schema(df, drop=None, batch_axis=False):
    inputs = []
    nrows = None if batch_axis else 1
    for k, v in zip(df.columns, df.dtypes):
        if drop is not None and k in drop:
            continue
        if v == 'int64':
            t = Int64TensorType([nrows, 1])
        elif v == 'float32':
            t = FloatTensorType([nrows, 1])
        elif v == 'float64':
            t = DoubleTensorType([nrows, 1])
        else:
            raise Exception("Bad type")
        inputs.append((k, t))
    return inputs

Using skl2onnx, convert the LinearRegression model to the ONNX format and save it locally.

# Convert the scikit model to onnx format
onnx_model = skl2onnx.convert_sklearn(model, 'Boston Data', convert_dataframe_schema(x_train), final_types=[('variable1',FloatTensorType([1,1]))])
# Save the onnx model locally
onnx_model_path = 'boston1.model.onnx'
onnxmltools.utils.save_model(onnx_model, onnx_model_path)

Note

You may need to set the target_opset parameter for the skl2onnx.convert_sklearn function if there's a mismatch between ONNX runtime version in SQL Edge and skl2onnx packge. For more information, see the SQL Edge Release notes to get the ONNX runtime version corresponding for the release, and pick the target_opset for the ONNX runtime based on the ONNX backward compatibility matrix.

Test the ONNX model

After converting the model to ONNX format, score the model to show little to no degradation in performance.

Note

ONNX Runtime uses floats instead of doubles so small discrepancies are possible.

import onnxruntime as rt
sess = rt.InferenceSession(onnx_model_path)

y_pred = np.full(shape=(len(x_train)), fill_value=np.nan)

for i in range(len(x_train)):
    inputs = {}
    for j in range(len(x_train.columns)):
        inputs[x_train.columns[j]] = np.full(shape=(1,1), fill_value=x_train.iloc[i,j])

    sess_pred = sess.run(None, inputs)
    y_pred[i] = sess_pred[0][0][0]

onnx_r2_score = r2_score(y_train, y_pred)
onnx_mse = mean_squared_error(y_train, y_pred)

print()
print('*** Onnx r2 score: {}'.format(onnx_r2_score))
print('*** Onnx MSE: {}\n'.format(onnx_mse))
print('R2 Scores are equal' if sklearn_r2_score == onnx_r2_score else 'Difference in R2 scores: {}'.format(abs(sklearn_r2_score - onnx_r2_score)))
print('MSE are equal' if sklearn_mse == onnx_mse else 'Difference in MSE scores: {}'.format(abs(sklearn_mse - onnx_mse)))
print()

Output:

*** Onnx r2 score: 0.7406426691136831
*** Onnx MSE: 21.894830759270633

R2 Scores are equal
MSE are equal

Insert the ONNX model

Store the model in Azure SQL Edge, in a models table in a database onnx. In the connection string, specify the server address, username, and password.

import pyodbc

server = '' # SQL Server IP address
username = '' # SQL Server username
password = '' # SQL Server password

# Connect to the master DB to create the new onnx database
connection_string = "Driver={ODBC Driver 17 for SQL Server};Server=" + server + ";Database=master;UID=" + username + ";PWD=" + password + ";"

conn = pyodbc.connect(connection_string, autocommit=True)
cursor = conn.cursor()

database = 'onnx'
query = 'DROP DATABASE IF EXISTS ' + database
cursor.execute(query)
conn.commit()

# Create onnx database
query = 'CREATE DATABASE ' + database
cursor.execute(query)
conn.commit()

# Connect to onnx database
db_connection_string = "Driver={ODBC Driver 17 for SQL Server};Server=" + server + ";Database=" + database + ";UID=" + username + ";PWD=" + password + ";"

conn = pyodbc.connect(db_connection_string, autocommit=True)
cursor = conn.cursor()

table_name = 'models'

# Drop the table if it exists
query = f'drop table if exists {table_name}'
cursor.execute(query)
conn.commit()

# Create the model table
query = f'create table {table_name} ( ' \
    f'[id] [int] IDENTITY(1,1) NOT NULL, ' \
    f'[data] [varbinary](max) NULL, ' \
    f'[description] varchar(1000))'
cursor.execute(query)
conn.commit()

# Insert the ONNX model into the models table
query = f"insert into {table_name} ([description], [data]) values ('Onnx Model',?)"

model_bits = onnx_model.SerializeToString()

insert_params  = (pyodbc.Binary(model_bits))
cursor.execute(query, insert_params)
conn.commit()

Load the data

Load the data into SQL.

First, create two tables, features and target, to store subsets of the Boston housing dataset.

  • Features contains all data being used to predict the target, median value.
  • Target contains the median value for each record in the dataset.
import sqlalchemy
from sqlalchemy import create_engine
import urllib

db_connection_string = "Driver={ODBC Driver 17 for SQL Server};Server=" + server + ";Database=" + database + ";UID=" + username + ";PWD=" + password + ";"

conn = pyodbc.connect(db_connection_string)
cursor = conn.cursor()

features_table_name = 'features'

# Drop the table if it exists
query = f'drop table if exists {features_table_name}'
cursor.execute(query)
conn.commit()

# Create the features table
query = \
    f'create table {features_table_name} ( ' \
    f'    [CRIM] float, ' \
    f'    [ZN] float, ' \
    f'    [INDUS] float, ' \
    f'    [CHAS] float, ' \
    f'    [NOX] float, ' \
    f'    [RM] float, ' \
    f'    [AGE] float, ' \
    f'    [DIS] float, ' \
    f'    [RAD] float, ' \
    f'    [TAX] float, ' \
    f'    [PTRATIO] float, ' \
    f'    [B] float, ' \
    f'    [LSTAT] float, ' \
    f'    [id] int)'

cursor.execute(query)
conn.commit()

target_table_name = 'target'

# Create the target table
query = \
    f'create table {target_table_name} ( ' \
    f'    [MEDV] float, ' \
    f'    [id] int)'

x_train['id'] = range(1, len(x_train)+1)
y_train['id'] = range(1, len(y_train)+1)

print(x_train.head())
print(y_train.head())

Finally, use sqlalchemy to insert the x_train and y_train pandas dataframes into the tables features and target, respectively.

db_connection_string = 'mssql+pyodbc://' + username + ':' + password + '@' + server + '/' + database + '?driver=ODBC+Driver+17+for+SQL+Server'
sql_engine = sqlalchemy.create_engine(db_connection_string)
x_train.to_sql(features_table_name, sql_engine, if_exists='append', index=False)
y_train.to_sql(target_table_name, sql_engine, if_exists='append', index=False)

Now you can view the data in the database.

Run PREDICT using the ONNX model

With the model in SQL, run native PREDICT on the data using the uploaded ONNX model.

Note

Change the notebook kernel to SQL to run the remaining cell.

USE onnx

DECLARE @model VARBINARY(max) = (
        SELECT DATA
        FROM dbo.models
        WHERE id = 1
        );

WITH predict_input
AS (
    SELECT TOP (1000) [id],
        CRIM,
        ZN,
        INDUS,
        CHAS,
        NOX,
        RM,
        AGE,
        DIS,
        RAD,
        TAX,
        PTRATIO,
        B,
        LSTAT
    FROM [dbo].[features]
    )
SELECT predict_input.id,
    p.variable1 AS MEDV
FROM PREDICT(MODEL = @model, DATA = predict_input, RUNTIME = ONNX) WITH (variable1 FLOAT) AS p;