Snowflake ML Model Development¶

Developing models¶

With Container Runtime for ML, available in Notebooks on Container Runtime, you can use popular open-source ML packages vith your Snowflake data, leveraging one or more GPU nodes, within the Snowflake cloud, ensuring security and governance for the entire ML workflow. The included data loading and training APIs are automatically distributed across all available CPUs or GPUs on a node, acelerating model training with large datasets.

For more information, see Getting Started with Snowflake Notebook Container Runtime, which presents a simple ML workflow leveraging the capabilities of the Container Runtime for ML.

Along with the flexibility and power of the Container Runtime for ML, the Snowflake ML Modeling API provides estimators and transformers that have APIs similar to those in the scikit-learn, xgboost, and lightgbm libraries. You can use these APIs to build and train machine learning models that can be used with Snowflake ML Operations such as the Snowpark Model Registry.

Examples¶

Review the following examples to get a sense of the similarities of the Snowflake Modeling API to the machine learning libraries you might be familiar with.

import numpy as np
import pandas as pd
import random
import string

from sklearn.datasets import make_regression
from snowflake.ml.modeling.preprocessing import MinMaxScaler, OrdinalEncoder
from snowflake.ml.modeling.pipeline import Pipeline
from snowflake.snowpark import Session

# Create a session with your preferred method
# session =

NUMERICAL_COLS = ["X1", "X2", "X3"]
CATEGORICAL_COLS = ["C1", "C2", "C3"]
FEATURE_COLS = NUMERICAL_COLS + CATEGORICAL_COLS
CATEGORICAL_OUTPUT_COLS = ["C1_OUT", "C2_OUT", "C3_OUT"]
FEATURE_OUTPUT_COLS = ["X1_FEAT_OUT", "X2_FEAT_OUT", "X3_FEAT_OUT", "C1_FEAT_OUT", "C2_FEAT_OUT", "C3_FEAT_OUT"]

# Create a dataset with numerical and categorical features
X, _ = make_regression(
    n_samples=1000,
    n_features=3,
    noise=0.1,
    random_state=0,
)
X = pd.DataFrame(X, columns=NUMERICAL_COLS)

def generate_random_string(length):
    return "".join(random.choices(string.ascii_uppercase, k=length))

categorical_feature_length = 2
categorical_features = {}
for c in CATEGORICAL_COLS:
    categorical_column = [generate_random_string(categorical_feature_length) for _ in range(X.shape[0])]
    categorical_features[c] = categorical_column

X = X.assign(**categorical_features)

features_df = session.create_dataframe(X)

# Fit a pipeline with OrdinalEncoder and MinMaxScaler on Snowflake
pipeline = Pipeline(
    steps=[
        (
            "OE",
            OrdinalEncoder(
                input_cols=CATEGORICAL_COLS,
                output_cols=CATEGORICAL_OUTPUT_COLS,
            )
        ),
        (
            "MMS",
            MinMaxScaler(
                input_cols=NUMERICAL_COLS + CATEGORICAL_OUTPUT_COLS,
                output_cols=FEATURE_OUTPUT_COLS,
            )
        ),
    ]
)

pipeline.fit(features_df)

# Use the pipeline to transform a dataset.
result = pipeline.transform(features_df)

from snowflake.ml.data.data_connector import DataConnector

# Retrieve data from a snowflake table
table_name = 'LARGE_TABLE_MULTIPLE_GBs'
snowpark_df = session.table(table_name)

# Materialize it into a pandas dataframe using DataConnector
pandas_df = DataConnector.from_dataframe(snowpark_df).to_pandas()

# Materialize it into a pytroch dataset using DataConnector
torch_dataset = data.to_torch_dataset(batch_size=1024)

import pandas as pd
from sklearn.datasets import make_classification

from snowflake.ml.modeling.xgboost import XGBClassifier
from snowflake.ml.utils.connection_params import SnowflakeLoginOptions
from snowflake.snowpark import Session

# Create a session with your preferred method
# session =

FEATURE_COLS = ["X1", "X2", "X3", "X4", "X5", "X6"]
LABEL_COLS = ["Y"]
OUTPUT_COLS = ["PREDICTIONS"]

# Set up data.
X, y = make_classification(
    n_samples=40000,
    n_features=6,
    n_informative=4,
    n_redundant=1,
    random_state=0,
    shuffle=True,
)

X = pd.DataFrame(X, columns=FEATURE_COLS)
y = pd.DataFrame(y, columns=LABEL_COLS)

features_pandas = pd.concat([X, y], axis=1)
features_df = session.create_dataframe(features_pandas)

# Train an XGBoost model on snowflake.
xgboost_model = XGBClassifier(
    input_cols=FEATURE_COLS,
    label_cols=LABEL_COLS,
    output_cols=OUTPUT_COLS
)

xgboost_model.fit(features_df)

# Use the model to make predictions.
predictions = xgboost_model.predict(features_df)
predictions[OUTPUT_COLS].show()

DATA_FILE_PATH = "~/Downloads/magic+gamma+telescope/magic04.data"

# Setup
from snowflake.ml.utils.connection_params import SnowflakeLoginOptions
from snowflake.snowpark import Session
import posixpath
import os

##
# Note: Create session https://docs.snowflake.com/en/developer-guide/snowpark/reference/python/latest/api/snowflake.snowpark.Session
##
session = Session.builder.configs(SnowflakeLoginOptions()).create()

session.sql("""
CREATE OR REPLACE TABLE Gamma_Telescope_Data(
    F_LENGTH FLOAT,
    F_WIDTH FLOAT,
    F_SIZE FLOAT,
    F_CONC FLOAT,
    F_CONC1 FLOAT,
    F_ASYM FLOAT,
    F_M3_LONG FLOAT,
    F_M3_TRANS FLOAT,
    F_ALPHA FLOAT,
    F_DIST FLOAT,
    CLASS VARCHAR(10))
""").collect()
session.sql("CREATE OR REPLACE STAGE SNOWPARK_ML_TEST_DATA_STAGE").collect()
session.file.put(
    DATA_FILE_PATH,
    "SNOWPARK_ML_TEST_DATA_STAGE/magic04.data",
    auto_compress=False,
    overwrite=True,
)

session.sql("""
COPY INTO Gamma_Telescope_Data FROM @SNOWPARK_ML_TEST_DATA_STAGE/magic04.data
FILE_FORMAT = (TYPE = 'CSV' field_optionally_enclosed_by='"',SKIP_HEADER = 0);
""").collect()

session.sql("select * from Gamma_Telescope_Data limit 5").collect()

from snowflake.ml.utils.connection_params import SnowflakeLoginOptions
from snowflake.snowpark import Session, DataFrame

from snowflake.ml.modeling.preprocessing import StandardScaler
from snowflake.ml.modeling.impute import SimpleImputer
from snowflake.ml.modeling.pipeline import Pipeline
from snowflake.ml.modeling.xgboost import XGBClassifier

from snowflake.ml.modeling.metrics import accuracy_score

##
# Note: Create session https://docs.snowflake.com/en/developer-guide/snowpark/reference/python/latest/api/snowflake.snowpark.Session
##
session = Session.builder.configs(SnowflakeLoginOptions()).create()

# Step 1: Create train and test dataframes
all_data = session.sql("select *, IFF(CLASS = 'g', 1.0, 0.0) as LABEL from Gamma_Telescope_Data").drop("CLASS")
train_data, test_data = all_data.random_split(weights=[0.9, 0.1], seed=0)

# Step 2: Construct training pipeline with preprocessing and modeling steps
FEATURE_COLS = [c for c in train_data.columns if c != "LABEL"]
LABEL_COLS = ["LABEL"]

pipeline = Pipeline(steps = [
    ("impute", SimpleImputer(input_cols=FEATURE_COLS, output_cols=FEATURE_COLS)),
    ("scaler", StandardScaler(input_cols=FEATURE_COLS, output_cols=FEATURE_COLS)),
    ("model", XGBClassifier(input_cols=FEATURE_COLS, label_cols=LABEL_COLS))
])

# Step 3: Train
pipeline.fit(train_data)

# Step 4: Eval
predict_on_training_data = pipeline.predict(train_data)
training_accuracy = accuracy_score(df=predict_on_training_data, y_true_col_names=["LABEL"], y_pred_col_names=["OUTPUT_LABEL"])

predict_on_test_data = pipeline.predict(test_data)
eval_accuracy = accuracy_score(df=predict_on_test_data, y_true_col_names=["LABEL"], y_pred_col_names=["OUTPUT_LABEL"])

print(f"Training accuracy: {training_accuracy} \nEval accuracy: {eval_accuracy}")

from snowflake.snowpark import Session, DataFrame
from snowflake.ml.utils.connection_params import SnowflakeLoginOptions

from sklearn.datasets import make_classification
from snowflake.snowpark import Session, DataFrame
from snowflake.ml.modeling.xgboost import XGBClassifier
from snowflake.ml.modeling.model_selection.grid_search_cv import GridSearchCV

FEATURE_COLS = ["X1", "X2", "X3", "X4", "X5", "X6"]
LABEL_COLS = ["Y"]
OUTPUT_COLS = ["PREDICTIONS"]

# Create a session using your favorite login option.
# In this example we use a session builder with `SnowflakeLoginOptions`.
session = Session.builder.configs(SnowflakeLoginOptions()).create()

# Set up data.
def set_up_data(session: Session, n_samples: int) -> DataFrame:
    X, y = make_classification(
        n_samples=n_samples,
        n_features=6,
        n_informative=2,
        n_redundant=0,
        random_state=0,
        shuffle=True,
    )

    X = pd.DataFrame(X, columns=FEATURE_COLS)
    y = pd.DataFrame(y, columns=LABEL_COLS)

    features_pandas = pd.concat([X, y], axis=1)
    features_pandas.head()

    features_df = session.create_dataframe(features_pandas)
    return features_df

features_df = set_up_data(session, 10**4)

# Create a warehouse to use for the tuning job.
session.sql(
    """
CREATE or replace warehouse HYPERPARAM_WH
    WITH WAREHOUSE_SIZE = 'X-SMALL'
    WAREHOUSE_TYPE = 'Standard'
    AUTO_SUSPEND = 60
    AUTO_RESUME = TRUE
    INITIALLY_SUSPENDED = FALSE;"""
).collect()
session.use_warehouse("HYPERPARAM_WH")

# Tune an XGB Classifier model using sklearn GridSearchCV.
DISTRIBUTIONS = dict(
    n_estimators=[10, 50],
    learning_rate=[0.01, 0.1, 0.2],
)
estimator = XGBClassifier()
grid_search_cv = GridSearchCV(estimator=estimator, param_grid=DISTRIBUTIONS, input_cols=FEATURE_COLS, label_cols=LABEL_COLS, output_cols=OUTPUT_COLS)

grid_search_cv.fit(features_df)

# Use the best model to make predictions.
predictions = grid_search_cv.predict(features_df)
predictions[OUTPUT_COLS].show()

# Retrieve sklearn model, and print the best score
sklearn_grid_search_cv = grid_search_cv.to_sklearn()
print(sklearn_grid_search_cv.best_score_)

large_features_df = set_up_data(session, 10**6)

# Scale up the warehouse for a faster fit. This takes 2m15s to run on an L warehouse versus 4m5s on a XS warehouse.
session.sql(f"ALTER WAREHOUSE {session.get_current_warehouse()} SET WAREHOUSE_SIZE='LARGE'").collect()

grid_search_cv.fit(large_features_df)
print(grid_search_cv.to_sklearn().best_score_)

Snowflake Modeling Classes¶

All Snowflake modeling and preprocessing classes are in the snowflake.ml.modeling namespace. The snowflake-ml-python modules have the same name as the corresponding modules from the sklearn namespace. For example, the module corresponding to sklearn.calibration is snowflake.ml.modeling.calibration. The xgboost and lightgbm modules correspond to snowflake.ml.modeling.xgboost and snowflake.ml.modeling.lightgbm, respectively.

The modeling API provides wrappers for underlying scikit-learn, xgboost, and lightgbm classes, the majority of which are executed as stored procedures (running on a single warehouse node) in the virtual warehouse. Not all of the classes from scikit-learn are supported. See the Python API Reference for a list of the classes currently available.

Some classes (including preprocessing and metrics classes) support distributed execution and may provide significant performance benefits compared to running the same operations locally. For more information, see Distributed Preprocessing and Distributed Hyperparameter Optimization. The table below lists the specific classes that support distributed execution.

General API Differences¶

Snowflake modeling classes includes data preprocessing, transformation, and prediction algorithms based on scikit-learn, xgboost, and lightgbm. The Snowpark Python classes are replacements for the corresponding classes from the original packages, with similar signatures. However, these APIs are designed to work with Snowpark DataFrames instead of NumPy arrays.

Although the API is similar to scikit-learn, there are some key differences. This section explains how to call the __init__ (constructor), fit, and predict methods for the Snowflake estimator and transformer classes.

• The constructor of all Snowflake model classes accepts five additional parameters (input_cols, output_cols, sample_weight_col, label_cols, and drop_input_cols) in addition to the parameters accepted by the equivalent classes in scikit-learn, xgboost, or lightgbm. These are strings or sequences of strings that specify the names of the input columns, output columns, sample weight column, and label columns in a Snowpark or Pandas DataFrame. If some of the datasets you use have different names, you can change these names after instantiation using one of the provided setter methods, such as set_input_cols.

• Because you specify column names when instantiating the class (or afterward, using setter methods) the fit and predict methods accept a single DataFrame instead of separate arrays for inputs, weights, and labels. The provided column names are used to access the appropriate column from the DataFrame in fit or predict. See fit and predict.

• By default, the transform and predict methods return a DataFrame containing all of the columns from the DataFrame passed to the method, with the output from the prediction stored in additional columns. You can transform in place by specifying output column names that match the input column names, or drop the input columns by passing drop_input_cols = True.) The scikit-learn, xgboost, and lightgbm equivalents return arrays containing only the results.

• Snowpark Python transformers do not have a fit_transform method. However, as with scikit-learn, parameter validation is only performed in the fit method, so you should call fit at some point before transform, even when the transformer does not do any fitting. fit returns the transformer, so the method calls may be chained; for example, Binarizer(threshold=0.5).fit(df).transform(df).

• Snowflake transformers do not currently have an inverse_transform method. In many use cases, this method is unnecessary because the input columns are retained in the output dataframe by default.

You can convert any Snowfalke modeling object to the corresponding scikit-learn, xgboost, or lightgbm object, allowing you to use all the methods and attributes of the underlying type. See Retrieving the Underlying Model.

from snowflake.ml.modeling.preprocessing import MinMaxScaler
from snowflake.snowpark import Session

# Snowflake identifiers are not case sensitive by default.
# These column names will be automatically updated to ["COLUMN_1", "COLUMN_2", "COLUMN_3"] by the Snowpark DataFrame.
schema = ["column_1", "column_2", "column_3"]
df = session.create_dataframe([[1, 2, 3]], schema = schema)
df.show()

--------------------------------------
|"COLUMN_1"  |"COLUMN_2"  |"COLUMN_3"|
--------------------------------------
|1           |2          |3          |
--------------------------------------

# Identify the column names using the Snowflake identifier.
input_cols = ["COLUMN_1", "COLUMN_2", "COLUMN_3"]
mms = MinMaxScaler(input_cols=input_cols)
mms.fit(df)

# To maintain lower case column names, include a double quote within the string.
schema = ['"column_1"', '"column_2"', '"column_3"']
df = session.create_dataframe([[1, 2, 3]], schema = schema)
df.show()

----------------------------------------
|'"column_1"'|'"column_2"'|'"column_3"'|
----------------------------------------
|1           |2           |3           |
----------------------------------------

# Since no conversion took place, the schema labels can be used as the column identifiers.
mms = MinMaxScaler(input_cols=schema)
mms.fit(df)

from sklearn.tree import DecisionTreeClassifier

model = DecisionTreeClassifier()

model.fit(
    X=df[feature_column_names], y=df[label_column_names], sample_weight=df[weight_column_name]
)

model.fit(df)

prediction_results = model.predict(X=df[feature_column_names])

prediction_results = model.predict(df)[output_column_names]

Distributed Training and Inference with SPCS¶

When running in a Snowflake Notebook on Snowpark Container Services (SPCS), model training and inference for these modeling classes are executed on the underlying compute cluster, not in a warehouse, and are transparently distributed across all nodes in the cluster to employ all available compute capability.

Preprocessing and metrics operations are pushed down to the warehouse. Many preprocessing classes support distributed execution when run in the warehouse; see Distributed Preprocessing.

Distributed Preprocessing¶

Many Snowflake data preprocessing and transformation functions are implemented using Snowflake’s distributed execution engine, which yields significant performance benefit compared to single-node execution (that is, stored procedures). To find out which functions support distributed execution, see Snowflake Modeling Classes.

The chart below shows illustrative performance numbers on large public datasets, running in a medium Snowpark-optimized warehouse, comparing scikit-learn running in stored procedures to Snowflake’s distributed implementations. In mary scenarios, your code can run 25 to 50 times faster when using Snowflake modeling classes.

Distributed Hyperparameter Optimization¶

Hyperparameter tuning is an integral part of the data science workflow. The Snowflake API provides distributed implementations of the scikit-learn GridSearchCV and RandomizedSearchCV APIs to enable efficient hyperparameter tuning on both single-node and multiple-node warehouses.

import snowflake.ml.modeling.parameters.disable_distributed_hpo

The smallest Snowflake virtual warehouse (XS) or Snowpark-optimized warehouse (M) has one node. Each successively larger size doubles the number of nodes.

For single-node (XS) warehouses, the full capacity of the node is utilized by default using scikit-learn’s joblib multiprocessing framework.

For multiple-node warehouses, the fit operations within your cross-validation tuning job are distributed across the nodes. No code changes are required to scale up. Estimator fits are executed in parallel across all available cores on all nodes in the warehouse.

As an illustration, consider the California housing dataset provided with the scikit-learn library. The data includes 20,640 rows of data with the following information:

• MedInc: Median income in the block group

• HouseAge: Median house age in the block group

• AveRooms: Ave number of rooms per household

• AveBedrms: Average number of bedrooms per household

• Population: The block group population

• AveOccup: Average number of household members

• Latitude and Longitude

The target of the dataset is the median income, expressed in hundreds of thousands of dollars.

In this example, we do grid search cross-validation on a random forest regressor to find the best hyperparameter combination to predict the median income.

from snowflake.ml.modeling.ensemble.random_forest_regressor import RandomForestRegressor
from snowflake.ml.modeling.model_selection.grid_search_cv import GridSearchCV
from sklearn import datasets

def load_housing_data() -> DataFrame:
    input_df_pandas = datasets.fetch_california_housing(as_frame=True).frame
    # Set the columns to be upper case for consistency with Snowflake identifiers.
    input_df_pandas.columns = [c.upper() for c in input_df_pandas.columns]
    input_df = session.create_dataframe(input_df_pandas)

    return input_df

input_df = load_housing_data()

# Use all the columns besides the median value as the features
input_cols = [c for c in input_df.columns if not c.startswith("MEDHOUSEVAL")]
# Set the target median value as the only label columns
label_cols = [c for c in input_df.columns if c.startswith("MEDHOUSEVAL")]


DISTRIBUTIONS = dict(
            max_depth=[80, 90, 100, 110],
            min_samples_leaf=[1,3,10],
            min_samples_split=[1.0, 3,10],
            n_estimators=[100,200,400]
        )
estimator = RandomForestRegressor()
n_folds = 5

clf = GridSearchCV(estimator=estimator, param_grid=DISTRIBUTIONS, cv=n_folds, input_cols=input_cols, label_cols=label_col)
clf.fit(input_df)

This example runs in just over 7 minutes on a Medium (single node) Snowpark-optimized warehouse, and takes just 3 minutes to run on an X-Large warehouse.

Deploying and Running Your Model¶

The result of training a model is a Python model object. You can use the trained model to make predictions by calling the model’s predict method. This creates a temporary user-defined function to run the model in your Snowflake virtual warehouse. This function is automatically deleted at the end of your Snowflake session (for example, when your script ends or when you close your notebook).

To keep the user-defined function after your session ends, you can create it manually. See the Quickstart on the topic for further information.

The Snowflake model registry also supports persistent models and makes finding and deploying them easier. See Snowflake Model Registry.

Partitioned Custom Models¶

The model registry also supports a special type of custom model where fit and inference are executed in parallel for a set of partitions. This can be a performant way to create many models at once from one dataset and execute inference immediately. Please see Snowflake Model Registry: Partitioned Custom Models for more details.

Pipeline for Multiple Transformations¶

With scikit-learn, it is common to run a series of transformations using a pipeline. scikit-learn pipelines do not work with Snowflake classes, so a Snowflake version of sklearn.pipeline.Pipeline is provided for running a series of transformations. This class is in the snowflake.ml.modeling.pipeline package, and it works the same as the scikit-learn version.

Retrieving the Underlying Model¶

Snowflake ML models can be “unwrapped,” that is, converted to the underlying third-party model types, with the following methods (depending on the library):

• to_sklearn

• to_xgboost

• to_lightgbm

All attributes and methods of the underlying model can then be accessed and run locally against the estimator. For example, in the GridSearchCV example, we convert the grid search estimator to a scikit-learn object in order to retrieve the best score.

best_score = grid_search_cv.to_sklearn().best_score_

Known Limitations¶

• Snowflake estimators and transformers do not currently support sparse inputs or sparse responses. If you have sparse data, convert it to a dense format before passing it to Snowflake’s estimators or transformers.

• The snowflake-ml-python package does not currently support matrix data types. Any operation on estimators and transformers that would produce a matrix as a result fails.

• The order of rows in result data is not guaranteed to match the order of rows in input data.

• Snowflake ML does not yet support pandas on Snowflake DataFrames. Convert the Pandas on Snowflake dataframe to a Snowpark dataframe to use it with the Snowflake modeling classes. The following example converts a DataFrame we have read from a Snowflake table:

  import modin.pandas as pd
  import snowflake.snowpark.modin.plugin
  from snowflake.ml.modeling.xgboost import XGBClassifier
  
  snowpark_pandas_df: modin.pandas.DataFrame = read_snowflake('MY_TABLE')
  # converting to Snowpark DataFrame adds an index column
  index_label_name = "_INDEX"
  snowpark_df = snowpark_pandas_df.to_snowpark(index=True, index_label=index_label_name)
  snowpark_df.show()
  

  Copy

  The resulting Snowpark DataFrame is as follows:

  --------------------------------------------------
  |"COLUMN_1"  |"COLUMN_2" |"TARGET"   |  "_INDEX" |
  --------------------------------------------------
  |1           |2          |3          |1          |
  --------------------------------------------------
  

  Copy

  The DataFrame can then be used to train the an XGBoost classifier as follows:

  # Identify the column names using the Snowflake identifier
  input_cols = ["COLUMN_1", "COLUMN_2", "COLUMN_3"]
  # Pass through the _INDEX column rather than using it for training
  xgb_clf = XGBClassifier(input_cols=input_cols, passthrough_cols=index_label_name, label_cols="TARGET")
  xgb_clf.fit(snowpark_df)
  

  Copy

Troubleshooting¶

import logging, sys

logging.basicConfig(stream=sys.stdout, level=logging.DEBUG)

def fix_values(F, column):
    return F.upper(F.regexp_replace(F.col(column), "[^a-zA-Z0-9]+", "_"))

Further Reading¶

See the documentation of the original libraries for complete information on their functionality.

• Scikit-Learn

• XGBoost

• LightGBM

Acknowledgement¶

Some parts of this document are derived from the Scikit-learn documentation, which is licensed under the BSD-3 “New” or “Revised” license and Copyright © 2007-2023 The scikit-learn developers. All rights reserved.

Some parts of this document are derived from the XGboost documentation, which is covered by Apache License 2.0, January 2004 and Copyright © 2019. All rights reserved.

Some parts of this document are derived from the LightGBM documentation, which is MIT-licensed and Copyright © Microsoft Corp. All rights reserved.