Create, train and use a Snowflake ML model in an app

This topic provides an example of how to train a Snowflake ML model within a Snowflake Native App using the scikit-learn Python package. The example in this topic can be used to train models on data in the consumer or provider accounts.

Create a versioned schema to hold the stored procedures

Within the setup script create a versioned schema that contains the stored procedure as shown in the following example:

  1. Create a versioned schema for the stored procedure

    CREATE OR ALTER VERSIONED SCHEMA core;
GRANT USAGE ON SCHEMA core TO APPLICATION ROLE app_public;
    Copy

Create a stored procedure to create and train a model

  1. Create a stored procedure for the Python function you are using to train a model as shown in the following example:

CREATE OR REPLACE PROCEDURE core.py_log_model(db STRING, schema STRING, mname STRING, mvname STRING)
RETURNS STRING
LANGUAGE python
RUNTIME_VERSION = 3.11
HANDLER = 'log_model'
PACKAGES = ('snowflake-snowpark-python','scikit-learn', 'snowflake-ml-python >=1.6.2', 'pandas', 'numpy')
AS '
  -- <body of the stored procedure>
';
Copy

This example creates a stored procedure named py_log_model and declares the Python packages required to train a model using scikit-learn:

  • snowflake-snowpark-python

  • scikit-learn

  • snowflake-ml-python

  • pandas

  • numpy

  • xgboost

After creating a stored procedure, add the following code to the body of the stored procedure:

  1. Add Python code to the body of the stored procedure

import _snowflake
from snowflake.ml.registry import Registry
import pandas as pd
import numpy as np
from sklearn import datasets
from snowflake.ml.modeling.xgboost import XGBClassifier

def log_model(sp_session, mname, mvname):
    reg = Registry(session=sp_session, schema_name=''stateful_schema'')

    iris = datasets.load_iris()
    df = pd.DataFrame(data=np.c_[iris["data"], iris["target"]], columns=iris["feature_names"] + ["target"])
    df.columns = [s.replace(" (CM)", "").replace(" ", "") for s in df.columns.str.upper()]
    input_cols = ["SEPALLENGTH", "SEPALWIDTH", "PETALLENGTH", "PETALWIDTH"]
    label_cols = "TARGET"
    output_cols = "PREDICTED_TARGET"

    clf_xgb = XGBClassifier(
        input_cols=input_cols, output_cols=output_cols, label_cols=label_cols, drop_input_cols=True
    )
    clf_xgb.fit(df)
    model_ref = reg.log_model(
        clf_xgb,
        model_name=f"{mname}",
        version_name=f"{mvname}",
        options={"enable_explainability": False},
    )
    return "success"
Copy

The log_model function performs the following:

  • Uses pandas and numpy to create a DataFrame to serve a the training data for the model.

  • Creates an instance of the XGBoost to serve as the training algorithm for the data.

  • Calls the fit() function of XGBoost to create a model and train it on the dataset.

  • Calls the log_model() function of Snowflake Model Registry to add the model to the model registry.

Note

Models created by an app must be stored in a model registry. Apps cannot access models that are stored on a stage.

  1. Optional: To allow consumers to run the stored procedure to train the model, grant the USAGE privilege on the stored procedure:

    GRANT USAGE ON PROCEDURE core.py_log_model(STRING, STRING) TO APPLICATION ROLE app_public;
    Copy

Create a stored procedure to run a model

  1. Create a stored procedure for the Python function you use to call the model.

  CREATE OR REPLACE PROCEDURE core.py_call_predict(mname STRING, mvname STRING)
  RETURNS TABLE()
  LANGUAGE python
  RUNTIME_VERSION = 3.11
  HANDLER = 'run_model'
  PACKAGES = ('snowflake-snowpark-python','scikit-learn', 'snowflake-ml-python>=1.6.2', 'pandas', 'xgboost')
  AS
'
-- <body of the stored procedure>
';
Copy

  1. Add the Python code you use to call the model

import _snowflake
from snowflake.ml.registry import Registry
import pandas as pd
from sklearn import datasets

def run_model(sp_session, mname, mvname):
  iris = datasets.load_iris()
  df = pd.DataFrame(data=iris["data"], columns=iris["feature_names"])
  df.columns = [s.replace(" (CM)", "").replace(" ", "") for s in df.columns.str.upper()]

  reg = Registry(session=sp_session, schema_name="stateful_schema")
  mv = reg.get_model(mname).version(mvname)

  pred = mv.run(df.head(10), function_name="predict")
  return sp_session.create_dataframe(pred)
Copy

The run_model function does the following:

  • Runs the load_iris() function to load the iris machine learning dataset.

  • Uses pandas to create a DataFrame based on the iris data set.

  • Runs the get_model() function to get the model registry.

  • Runs the predict function on the model.

  • Returns the result.

  1. Optional: To allow consumers to run the stored procedure to train the model, grant the USAGE privilege on the stored procedure:

    GRANT USAGE ON PROCEDURE core.py_call_predict(STRING, STRING) TO APPLICATION ROLE app_public;
    Copy

Run the stored procedures

If the app grants the USAGE privilege on these stored procedures to an application role, consumers can call the stored procedures to train and run the models as shown in the following examples:

CALL my_app.core.py_log_model('md1', 'V1');
Copy

This command calls the py_log_model stored procedure to train the model.

CALL my_app.core.py_call_predict('md1', 'V1');
Copy

This command calls the py_call_predict stored procedure to call the predict function on the model.

Language: English