Step 5: Train and save a Python model using T-SQL

This article is part of a tutorial, In-database Python analytics for SQL developers.

In this step, you learn how to train a machine learning model using the Python packages scikit-learn and revoscalepy. These Python libraries are already installed with SQL Server Machine Learning Services.

You load the modules and call the necessary functions to create and train the model using a SQL Server stored procedure. The model requires the data features you engineered in earlier lessons. Finally, you save the trained model to a SQL Server table.

Important

There have been several changes in the revoscalepy package, which required small changes in the code for this tutorial. See the changelist at the end of this tutorial.

If you installed Python Services using a prerelease version of SLq Server 2017, we recommend that you upgrade to the latest version.

Split the sample data into training and testing sets

  1. You can use the stored procedure TrainTestSplit to divide the data in the nyctaxi_sample table into two parts: nyctaxi_sample_training and nyctaxi_sample_testing.

    This stored procedure should already be created for you, but you can run the following code to create it:

    CREATE PROCEDURE [dbo].[TrainTestSplit] (@pct int)
    AS
    
    DROP TABLE IF EXISTS dbo.nyctaxi_sample_training
    SELECT * into nyctaxi_sample_training FROM nyctaxi_sample WHERE (ABS(CAST(BINARY_CHECKSUM(medallion,hack_license)  as int)) % 100) < @pct
    
    DROP TABLE IF EXISTS dbo.nyctaxi_sample_testing
    SELECT * into nyctaxi_sample_testing FROM nyctaxi_sample
    WHERE (ABS(CAST(BINARY_CHECKSUM(medallion,hack_license)  as int)) % 100) > @pct
    GO
    
  2. To divide your data using a custom split, run the stored procedure, and type an integer that represents the percentage of data allocated to the training set. For example, the following statement would allocate 60% of data to the training set.

    EXEC TrainTestSplit 60
    GO
    

Build a logistic regression model

After the data has been prepared, you can use it to train a model. You do this by calling a stored procedure that runs some Python code, taking as input the training data table. For this tutorial, you create two models, both binary classification models:

  • The stored procedure TrainTipPredictionModelRxPy creates a tip prediction model using the revoscalepy package.
  • The stored procedure TrainTipPredictionModelSciKitPy creates a tip prediction model using the scikit-learn package.

Each stored procedure uses the input data you provide to create and train a logistic regression model. All Python code is wrapped in the system stored procedure, sp_execute_external_script.

To make it easier to retrain the model on new data, you wrap the call to sp_execute_exernal_script in another stored procedure, and pass in the new training data as a parameter. This section will walk you through that process.

TrainTipPredictionModelSciKitPy

  1. In Management Studio, open a new Query window and run the following statement to create the stored procedure TrainTipPredictionModelSciKitPy. The stored procedure contains a definition of the input data, so you don't need to provide an input query.

    DROP PROCEDURE IF EXISTS TrainTipPredictionModelSciKitPy;
    GO
    
    CREATE PROCEDURE [dbo].[TrainTipPredictionModelSciKitPy] (@trained_model varbinary(max) OUTPUT)
    AS
    BEGIN
      EXEC sp_execute_external_script
      @language = N'Python',
      @script = N'
      import numpy
      import pickle
      from sklearn.linear_model import LogisticRegression
    
      ##Create SciKit-Learn logistic regression model
      X = InputDataSet[["passenger_count", "trip_distance", "trip_time_in_secs", "direct_distance"]]
      y = numpy.ravel(InputDataSet[["tipped"]])
    
      SKLalgo = LogisticRegression()
      logitObj = SKLalgo.fit(X, y)
    
      ##Serialize model
      trained_model = pickle.dumps(logitObj)
      ',
      @input_data_1 = N'
      select tipped, fare_amount, passenger_count, trip_time_in_secs, trip_distance, 
      dbo.fnCalculateDistance(pickup_latitude, pickup_longitude,  dropoff_latitude, dropoff_longitude) as direct_distance
      from nyctaxi_sample_training
      ',
      @input_data_1_name = N'InputDataSet',
      @params = N'@trained_model varbinary(max) OUTPUT',
      @trained_model = @trained_model OUTPUT;
      ;
    END;
    GO
    
  2. Run the following SQL statements to insert the trained model into table nyc_taxi_models.

    DECLARE @model VARBINARY(MAX);
    EXEC TrainTipPredictionModelSciKitPy @model OUTPUT;
    INSERT INTO nyc_taxi_models (name, model) VALUES('SciKit_model', @model);
    

    Processing of the data and fitting the model might take a couple of mins. Messages that would be piped to Python's stdout stream are displayed in the Messages window of Management Studio. For example:

    STDOUT message(s) from external script: C:\Program Files\Microsoft SQL Server\MSSQL14.MSSQLSERVER\PYTHON_SERVICES\lib\site-packages\revoscalepy

  3. Open the table nyc_taxi_models. You can see that one new row has been added, which contains the serialized model in the column model.

    linear_model 0x800363736B6C6561726E2E6C696E6561....

TrainTipPredictionModelRxPy

This stored procedure uses the new revoscalepy package, which is a new package for Python. It contains objects, transformation, and algorithms similar to those provided for the R language's RevoScaleR package.

By using revoscalepy, you can create remote compute contexts, move data between compute contexts, transform data, and train predictive models using popular algorithms such as logistic and linear regression, decision trees, and more. For more information, see what is revoscalepy?

  1. In Management Studio, open a new Query window and run the following statement to create the stored procedure TrainTipPredictionModelRxPy. Because the stored procedure already includes a definition of the input data, you don't need to provide an input query.

    DROP PROCEDURE IF EXISTS TrainTipPredictionModelRxPy;
    GO
    
    CREATE PROCEDURE [dbo].[TrainTipPredictionModelRxPy] (@trained_model varbinary(max) OUTPUT)
    AS
    BEGIN
    EXEC sp_execute_external_script 
      @language = N'Python',
      @script = N'
    import numpy
    import pickle
    from revoscalepy.functions.RxLogit import rx_logit
    
    ## Create a logistic regression model using rx_logit function from revoscalepy package
    logitObj = rx_logit("tipped ~ passenger_count + trip_distance + trip_time_in_secs + direct_distance", data = InputDataSet);
    
    ## Serialize model
    trained_model = pickle.dumps(logitObj)
    ',
    @input_data_1 = N'
    select tipped, fare_amount, passenger_count, trip_time_in_secs, trip_distance, 
    dbo.fnCalculateDistance(pickup_latitude, pickup_longitude,  dropoff_latitude, dropoff_longitude) as direct_distance
    from nyctaxi_sample_training
    ',
    @input_data_1_name = N'InputDataSet',
    @params = N'@trained_model varbinary(max) OUTPUT',
    @trained_model = @trained_model OUTPUT;
    ;
    END;
    GO
    

    This stored procedure performs the following steps as part of model training:

    • The SELECT query applies the custom scalar function fnCalculateDistance to calculate the direct distance between the pick-up and drop-off locations. The results of the query are stored in the default Python input variable, InputDataset.
    • The binary variable tipped is used as the label or outcome column, and the model is fit using these feature columns: passenger_count, trip_distance, trip_time_in_secs, and direct_distance.
    • The trained model is serialized and stored in the Python variable logitObj. By adding the T-SQL keyword OUTPUT, you can add the variable as an output of the stored procedure. In the next step, that variable is used to insert the binary code of the model into a database table nyc_taxi_models. This mechanism makes it easy to store and re-use models.
  2. Run the stored procedure as follows to insert the trained revoscalepy model into the table _nyc_taxi_models.

    DECLARE @model VARBINARY(MAX);
    EXEC TrainTipPredictionModelRxPy @model OUTPUT;
    
    INSERT INTO nyc_taxi_models (name, model) VALUES('revoscalepy_model', @model);
    

    Processing of the data and fitting the model might take a while. Messages that would be piped to Python's stdout stream are displayed in the Messages window of Management Studio. For example:

    STDOUT message(s) from external script: C:\Program Files\Microsoft SQL Server\MSSQL14.MSSQLSERVER\PYTHON_SERVICES\lib\site-packages\revoscalepy

  3. Open the table nyc_taxi_models. You can see that one new row has been added, which contains the serialized model in the column model.

    rx_model 0x8003637265766F7363616c....

In the next step, you use the trained models to create predictions.

Next step

Step 6: Operationalize the Python model using SQL Server

Previous step

Step 4: Create data features using T-SQL