snowflake-ml-python

Snowpark ML

Snowpark ML is a set of tools including SDKs and underlying infrastructure to build and deploy machine learning models. With Snowpark ML, you can pre-process data, train, manage and deploy ML models all within Snowflake, using a single SDK, and benefit from Snowflake’s proven performance, scalability, stability and governance at every stage of the Machine Learning workflow.

Key Components of Snowpark ML

The Snowpark ML Python SDK provides a number of APIs to support each stage of an end-to-end Machine Learning development and deployment process, and includes two key components.

Snowpark ML Development

Snowpark ML Development provides a collection of python APIs enabling efficient ML model development directly in Snowflake:

1. Modeling API (snowflake.ml.modeling) for data preprocessing, feature engineering and model training in Snowflake. This includes the snowflake.ml.modeling.preprocessing module for scalable data transformations on large data sets utilizing the compute resources of underlying Snowpark Optimized High Memory Warehouses, and a large collection of ML model development classes based on sklearn, xgboost, and lightgbm.

2. Framework Connectors: Optimized, secure and performant data provisioning for Pytorch and Tensorflow frameworks in their native data loader formats.

3. FileSet API: FileSet provides a Python fsspec-compliant API for materializing data into a Snowflake internal stage from a query or Snowpark Dataframe along with a number of convenience APIs.

Snowflake MLOps

Snowflake MLOps contains suit of tools and objects to make ML development cycle. It complements the Snowpark ML Development API, and provides end to end development to deployment within Snowflake. Currently, the API consists of:

1. Registry: A python API allows secure deployment and management of models in Snowflake, supporting models trained both inside and outside of Snowflake.
2. Feature Store: A fully integrated solution for defining, managing, storing and discovering ML features derived from your data. The Snowflake Feature Store supports automated, incremental refresh from batch and streaming data sources, so that feature pipelines need be defined only once to be continuously updated with new data.
3. Datasets: Dataset provide an immutable, versioned snapshot of your data suitable for ingestion by your machine learning models.

Getting started

Have your Snowflake account ready

If you don't have a Snowflake account yet, you can sign up for a 30-day free trial account.

Installation

Follow the installation instructions in the Snowflake documentation.

Python versions 3.9 to 3.11 are supported. You can use miniconda or anaconda to create a Conda environment (recommended), or virtualenv to create a virtual environment.

Conda channels

The Snowflake Conda Channel contains the official snowpark ML package releases. The recommended approach is to install snowflake-ml-python this conda channel:

conda install \
  -c https://repo.anaconda.com/pkgs/snowflake \
  --override-channels \
  snowflake-ml-python

See the developer guide for installation instructions.

The latest version of the snowpark-ml-python package is also published in a conda channel in this repository. Package versions in this channel may not yet be present in the official Snowflake conda channel.

Install snowflake-ml-python from this channel with the following (being sure to replace <version_specifier> with the desired version, e.g. 1.0.10):

conda install \
  -c https://raw.githubusercontent.com/snowflakedb/snowflake-ml-python/conda/releases/  \
  -c https://repo.anaconda.com/pkgs/snowflake \
  --override-channels \
  snowflake-ml-python==<version_specifier>

Note that until a snowflake-ml-python package version is available in the official Snowflake conda channel, there may be compatibility issues. Server-side functionality that snowflake-ml-python depends on may not yet be released.

Verifying the package

1. Install cosign. This example is using golang installation: installing-cosign-with-go.

2. Download the file from the repository like pypi.

3. Download the signature files from the release tag.

4. Verify signature on projects signed using Jenkins job:

   cosign verify-blob snowflake_ml_python-1.7.0.tar.gz --key snowflake-ml-python-1.7.0.pub --signature resources.linux.snowflake_ml_python-1.7.0.tar.gz.sig
   
   cosign verify-blob snowflake_ml_python-1.7.0.tar.gz --key snowflake-ml-python-1.7.0.pub --signature resources.linux.snowflake_ml_python-1.7.0
   

NOTE: Version 1.7.0 is used as example here. Please choose the the latest version.

Release History

1.7.2

Bug Fixes

• Model Explainability: Fix issue that explain is enabled for scikit-learn pipeline whose task is UNKNOWN and fails later when invoked.

Behavior Changes

New Features

• Registry: Support asynchronous model inference service creation with the block option in ModelVersion.create_service() set to True by default.

1.7.1 (2024-11-05)

Bug Fixes

• Registry: Null value is now allowed in the dataframe used in model signature inference. Null values will be ignored and others will be used to infer the signature.
• Registry: Pandas Extension DTypes (pandas.StringDType(), pandas.BooleanDType(), etc.) are now supported in model signature inference.
• Registry: Null value is now allowed in the dataframe used to predict.
• Data: Fix missing snowflake.ml.data.* module exports in wheel
• Dataset: Fix missing snowflake.ml.dataset.* module exports in wheel.
• Registry: Fix the issue that tf_keras.Model is not recognized as keras model when logging.

Behavior Changes

New Features

• Registry: Option to enable_monitoring set to False by default. This will gate access to preview features of Model Monitoring.
• Model Monitoring: show_model_monitors Registry method. This feature is still in Private Preview.
• Registry: Support pd.Series in input and output data.
• Model Monitoring: add_monitor Registry method. This feature is still in Private Preview.
• Model Monitoring: resume and suspend ModelMonitor. This feature is still in Private Preview.
• Model Monitoring: get_monitor Registry method. This feature is still in Private Preview.
• Model Monitoring: delete_monitor Registry method. This feature is still in Private Preview.

1.7.0 (10-22-2024)

Behavior Change

• Generic: Require python >= 3.9.
• Data Connector: Update to_torch_dataset and to_torch_datapipe to add a dimension for scalar data. This allows for more seamless integration with PyTorch DataLoader, which creates batches by stacking inputs of each batch.

Examples:

ds = connector.to_torch_dataset(shuffle=False, batch_size=3)

• Input: "col1": [10, 11, 12]

  • Previous batch: array([10., 11., 12.]) with shape (3,)
  • New batch: array([[10.], [11.], [12.]]) with shape (3, 1)

• Input: "col2": [[0, 100], [1, 110], [2, 200]]

  • Previous batch: array([[ 0, 100], [ 1, 110], [ 2, 200]]) with shape (3,2)
  • New batch: No change

• Model Registry: External access integrations are optional when creating a model inference service in Snowflake >= 8.40.0.

• Model Registry: Deprecate build_external_access_integration with build_external_access_integrations in ModelVersion.create_service().

Bug Fixes

• Registry: Updated log_model API to accept both signature and sample_input_data parameters.
• Feature Store: ExampleHelper uses fully qualified path for table name. change weather features aggregation from 1d to 1h.
• Data Connector: Return numpy array with appropriate object type instead of list for multi-dimensional data from to_torch_dataset and to_torch_datapipe
• Model explainability: Incompatibility between SHAP 0.42.1 and XGB 2.1.1 resolved by using latest SHAP 0.46.0.

New Features

• Registry: Provide pass keyworded variable length of arguments to class ModelContext. Example usage:

mc = custom_model.ModelContext(
    config = 'local_model_dir/config.json',
    m1 = model1
)

class ExamplePipelineModel(custom_model.CustomModel):
    def __init__(self, context: custom_model.ModelContext) -> None:
      super().__init__(context)
      v = open(self.context['config']).read()
      self.bias = json.loads(v)['bias']

    @custom_model.inference_api
    def predict(self, input: pd.DataFrame) -> pd.DataFrame:
      model_output = self.context['m1'].predict(input)
      return pd.DataFrame({'output': model_output + self.bias})

• Model Development: Upgrade scikit-learn in UDTF backend for log_loss metric. As a result, eps argument is now ignored.
• Data Connector: Add the option of passing a None sized batch to to_torch_dataset for better interoperability with PyTorch DataLoader.
• Model Registry: Support pandas.CategoricalDtype
• Registry: It is now possible to pass signatures and sample_input_data at the same time to capture background data from explainablity and data lineage.

1.6.4 (2024-10-17)

Bug Fixes

• Registry: Fix an issue that leads to incident when using ModelVersion.run with service.

1.6.3 (2024-10-07)

• Model Registry (PrPr) has been removed.

Bug Fixes

• Registry: Fix a bug that when package whose name does not follow PEP-508 is provided when logging the model, an unexpected normalization is happening.
• Registry: Fix not a valid remote uri error when logging mlflow models.
• Registry: Fix a bug that ModelVersion.run is called in a nested way.
• Registry: Fix an issue that leads to log_model failure when local package version contains parts other than base version.
• Fix issue where sample_weights were not being applied to search estimators.
• Model explainability: Fix bug which creates explain as a function instead of table function when enabling by default.
• Model explainability: Update lightgbm binary classification to return non-json values, from customer feedback.

New Features

• Data: Improve DataConnector.to_pandas() performance when loading from Snowpark DataFrames.
• Model Registry: Allow users to set a model task while using log_model.
• Feature Store: FeatureView supports ON_CREATE or ON_SCHEDULE initialize mode.

1.6.2 (2024-09-04)

Bug Fixes

• Modeling: Support XGBoost version that is larger than 2.

• Data: Fix multiple epoch iteration over DataConnector.to_torch_datapipe() DataPipes.

• Generic: Fix a bug that when an invalid name is provided to argument where fully qualified name is expected, it will be parsed wrongly. Now it raises an exception correctly.

• Model Explainability: Handle explanations for multiclass XGBoost classification models

• Model Explainability: Workarounds and better error handling for XGB>2.1.0 not working with SHAP==0.42.1

New Features

• Data: Add top-level exports for DataConnector and DataSource to snowflake.ml.data.
• Data: Add native batching support via batch_size and drop_last_batch arguments to DataConnector.to_torch_dataset()
• Feature Store: update_feature_view() supports taking feature view object as argument.

1.6.1 (2024-08-12)

Bug Fixes

• Feature Store: Support large metadata blob when generating dataset
• Feature Store: Added a hidden knob in FeatureView as kargs for setting customized refresh_mode
• Registry: Fix an error message in Model Version run when function_name is not mentioned and model has multiple target methods.
• Cortex inference: snowflake.cortex.Complete now only uses the REST API for streaming and the use_rest_api_experimental is no longer needed.
• Feature Store: Add a new API: FeatureView.list_columns() which list all column information.
• Data: Fix DataFrame ingestion with ArrowIngestor.

New Features

• Enable set_params to set the parameters of the underlying sklearn estimator, if the snowflake-ml model has been fit.
• Data: Add snowflake.ml.data.ingestor_utils module with utility functions helpful for DataIngestor implementations.
• Data: Add new to_torch_dataset() connector to DataConnector to replace deprecated DataPipe.
• Registry: Option to enable_explainability set to True by default for XGBoost, LightGBM and CatBoost as PuPr feature.
• Registry: Option to enable_explainability when registering SHAP supported sklearn models.

1.6.0 (2024-07-29)

Bug Fixes

• Modeling: SimpleImputer can impute integer columns with integer values.
• Registry: Fix an issue when providing a pandas Dataframe whose index is not starting from 0 as the input to the ModelVersion.run.

New Features

• Feature Store: Add overloads to APIs accept both object and name/version. Impacted APIs include read_feature_view(), refresh_feature_view(), get_refresh_history(), resume_feature_view(), suspend_feature_view(), delete_feature_view().
• Feature Store: Add docstring inline examples for all public APIs.
• Feature Store: Add new utility class ExampleHelper to help with load source data to simplify public notebooks.
• Registry: Option to enable_explainability when registering XGBoost models as a pre-PuPr feature.
• Feature Store: add new API update_entity().
• Registry: Option to enable_explainability when registering Catboost models as a pre-PuPr feature.
• Feature Store: Add new argument warehouse to FeatureView constructor to overwrite the default warehouse. Also add a new column 'warehouse' to the output of list_feature_views().
• Registry: Add support for logging model from a model version.
• Modeling: Distributed Hyperparameter Optimization now announce GA refresh version. The latest memory efficient version will not have the 10GB training limitation for dataset any more. To turn off, please run from snowflake.ml.modeling._internal.snowpark_implementations import ( distributed_hpo_trainer, ) distributed_hpo_trainer.ENABLE_EFFICIENT_MEMORY_USAGE = False
• Registry: Option to enable_explainability when registering LightGBM models as a pre-PuPr feature.
• Data: Add new snowflake.ml.data preview module which contains data reading utilities like DataConnector
  • DataConnector provides efficient connectors from Snowpark DataFrame and Snowpark ML Dataset to external frameworks like PyTorch, TensorFlow, and Pandas. Create DataConnector instances using the classmethod constructors DataConnector.from_dataset() and DataConnector.from_dataframe().
• Data: Add new DataConnector.from_sources() classmethod constructor for constructing from DataSource objects.
• Data: Add new ingestor_class arg to DataConnector classmethod constructors for easier DataIngestor injection.
• Dataset: DatasetReader now subclasses new DataConnector class.
  • Add optional limit arg to DatasetReader.to_pandas()

Behavior Changes

• Feature Store: change some positional parameters to keyword arguments in following APIs:
  • Entity(): desc.
  • FeatureView(): timestamp_col, refresh_freq, desc.
  • FeatureStore(): creation_mode.
  • update_entity(): desc.
  • register_feature_view(): block, overwrite.
  • list_feature_views(): entity_name, feature_view_name.
  • get_refresh_history(): verbose.
  • retrieve_feature_values(): spine_timestamp_col, exclude_columns, include_feature_view_timestamp_col.
  • generate_training_set(): save_as, spine_timestamp_col, spine_label_cols, exclude_columns, include_feature_view_timestamp_col.
  • generate_dataset(): version, spine_timestamp_col, spine_label_cols, exclude_columns, include_feature_view_timestamp_col, desc, output_type.

1.5.4 (2024-07-11)

Bug Fixes

• Model Registry (PrPr): Fix 401 Unauthorized issue when deploying model to SPCS.
• Feature Store: Downgrades exceptions to warnings for few property setters in feature view. Now you can set desc, refresh_freq and warehouse for draft feature views.
• Modeling: Fix an issue with calling OrdinalEncoder with categories as a dictionary and a pandas DataFrame
• Modeling: Fix an issue with calling OneHotEncoder with categories as a dictionary and a pandas DataFrame

New Features

• Registry: Allow overriding device_map and device when loading huggingface pipeline models.
• Registry: Add set_alias method to ModelVersion instance to set an alias to model version.
• Registry: Add unset_alias method to ModelVersion instance to unset an alias to model version.
• Registry: Add partitioned_inference_api allowing users to create partitioned inference functions in registered models. Enable model inference methods with table functions with vectorized process methods in registered models.
• Feature Store: add 3 more columns: refresh_freq, refresh_mode and scheduling_state to the result of list_feature_views().
• Feature Store: update_feature_view() supports updating description.
• Feature Store: add new API refresh_feature_view().
• Feature Store: add new API get_refresh_history().
• Feature Store: Add generate_training_set() API for generating table-backed feature snapshots.
• Feature Store: Add DeprecationWarning for generate_dataset(..., output_type="table").
• Feature Store: update_feature_view() supports updating description.
• Feature Store: add new API refresh_feature_view().
• Feature Store: add new API get_refresh_history().
• Model Development: OrdinalEncoder supports a list of array-likes for categories argument.
• Model Development: OneHotEncoder supports a list of array-likes for categories argument.

1.5.3 (06-17-2024)

Bug Fixes

• Modeling: Fix an issue causing lineage information to be missing for Pipeline, GridSearchCV , SimpleImputer, and RandomizedSearchCV
• Registry: Fix an issue that leads to incorrect result when using pandas Dataframe with over 100, 000 rows as the input of ModelVersion.run method in Stored Procedure.

New Features

• Registry: Add support for TIMESTAMP_NTZ model signature data type, allowing timestamp input and output.
• Dataset: Add DatasetVersion.label_cols and DatasetVersion.exclude_cols properties.

1.5.2 (06-10-2024)

Bug Fixes

• Registry: Fix an issue that leads to unable to log model in store procedure.
• Modeling: Quick fix import snowflake.ml.modeling.parameters.enable_anonymous_sproc cannot be imported due to package dependency error.

Behavior Changes

New Features

1.5.1 (05-22-2024)

Bug Fixes

• Dataset: Fix snowflake.connector.errors.DataError: Query Result did not match expected number of rows when accessing DatasetVersion properties when case insensitive SHOW VERSIONS IN DATASET check matches multiple version names.
• Dataset: Fix bug in SnowFS bulk file read when used with DuckDB
• Registry: Fixed a bug when loading old models.
• Lineage: Fix Dataset source lineage propagation through snowpark.DataFrame transformations

Behavior Changes

• Feature Store: convert clear() into a private function. Also make it deletes feature views and entities only.
• Feature Store: Use NULL as default value for timestamp tag value.

New Features

• Feature Store: Added new snowflake.ml.feature_store.setup_feature_store() API to assist Feature Store RBAC setup.
• Feature Store: Add output_type argument to FeatureStore.generate_dataset() to allow generating data snapshots as Datasets or Tables.
• Registry: log_model, get_model, delete_model now supports fully qualified name.
• Modeling: Supports anonymous stored procedure during fit calls so that modeling would not require sufficient permissions to operate on schema. Please call import snowflake.ml.modeling.parameters.enable_anonymous_sproc # noqa: F401

1.5.0 (05-01-2024)

Bug Fixes

• Registry: Fix invalid parameter 'SHOW_MODEL_DETAILS_IN_SHOW_VERSIONS_IN_MODEL' error.

Behavior Changes

• Model Development: The behavior of fit_transform for all estimators is changed. Firstly, it will cover all the estimator that contains this function, secondly, the output would be the union of pandas DataFrame and snowpark DataFrame.

Model Registry (PrPr)

snowflake.ml.registry.artifact and related snowflake.ml.model_registry.ModelRegistry APIs have been removed.

• Removed snowflake.ml.registry.artifact module.
• Removed ModelRegistry.log_artifact(), ModelRegistry.list_artifacts(), ModelRegistry.get_artifact()
• Removed artifacts argument from ModelRegistry.log_model()

Dataset (PrPr)

snowflake.ml.dataset.Dataset has been redesigned to be backed by Snowflake Dataset entities.

• New Datasets can be created with Dataset.create() and existing Datasets may be loaded with Dataset.load().
• Datasets now maintain an immutable selected_version state. The Dataset.create_version() and Dataset.load_version() APIs return new Dataset objects with the requested selected_version state.
• Added dataset.create_from_dataframe() and dataset.load_dataset() convenience APIs as a shortcut to creating and loading Datasets with a pre-selected version.
• Dataset.materialized_table and Dataset.snapshot_table no longer exist with Dataset.fully_qualified_name as the closest equivalent.
• Dataset.df no longer exists. Instead, use DatasetReader.read.to_snowpark_dataframe().
• Dataset.owner has been moved to Dataset.selected_version.owner
• Dataset.desc has been moved to DatasetVersion.selected_version.comment
• Dataset.timestamp_col, Dataset.label_cols, Dataset.feature_store_metadata, and Dataset.schema_version have been removed.

Feature Store (PrPr)

• FeatureStore.generate_dataset argument list has been changed to match the new snowflake.ml.dataset.Dataset definition

  • materialized_table has been removed and replaced with name and version.
  • name moved to first positional argument
  • save_mode has been removed as merge behavior is no longer supported. The new behavior is always errorifexists.

• Change feature view version type from str to FeatureViewVersion. It is a restricted string literal.

• Remove as_dataframe arg from FeatureStore.list_feature_views(), now always returns result as DataFrame.

• Combines few metadata tags into a new tag: SNOWML_FEATURE_VIEW_METADATA. This will make previously created feature views not readable by new SDK.

New Features

• Registry: Add export method to ModelVersion instance to export model files.
• Registry: Add load method to ModelVersion instance to load the underlying object from the model.
• Registry: Add Model.rename method to Model instance to rename or move a model.

Dataset (PrPr)

• Added Snowpark DataFrame integration using Dataset.read.to_snowpark_dataframe()
• Added Pandas DataFrame integration using Dataset.read.to_pandas()
• Added PyTorch and TensorFlow integrations using Dataset.read.to_torch_datapipe() and Dataset.read.to_tf_dataset() respectively.
• Added fsspec style file integration using Dataset.read.files() and Dataset.read.filesystem()

Feature Store

• use new tag_reference_internal to speed up metadata lookup.

1.4.1 (2024-04-18)

New Features

• Registry: Add support for catboost model (catboost.CatBoostClassifier, catboost.CatBoostRegressor).
• Registry: Add support for lightgbm model (lightgbm.Booster, lightgbm.LightGBMClassifier, lightgbm.LightGBMRegressor).

Bug Fixes

• Registry: Fix a bug that leads to relax_version option is not working.

Behavior changes

• Feature Store: update_feature_view takes refresh_freq and warehouse as argument.

1.4.0 (2024-04-08)

Bug Fixes

• Registry: Fix a bug when multiple models are being called from the same query, models other than the first one will have incorrect result. This fix only works for newly logged model.
• Modeling: When registering a model, only method(s) that is mentioned in save_model would be added to model signature in SnowML models.
• Modeling: Fix a bug that when n_jobs is not 1, model cannot execute methods such as predict, predict_log_proba, and other batch inference methods. The n_jobs would automatically set to 1 because vectorized udf currently doesn't support joblib parallel backend.
• Modeling: Fix a bug that batch inference methods cannot infer the datatype when the first row of data contains NULL.
• Modeling: Matches Distributed HPO output column names with the snowflake identifier.
• Modeling: Relax package versions for all Distributed HPO methods if the installed version is not available in the Snowflake conda channel
• Modeling: Add sklearn as required dependency for LightGBM package.

Behavior Changes

• Registry: apply method is no longer by default logged when logging a xgboost model. If that is required, it could be specified manually when logging the model by log_model(..., options={"target_methods": ["apply", ...]}).
• Feature Store: register_entity returns an entity object.
• Feature Store: register_feature_view block=true becomes default.

New Features

• Registry: Add support for sentence-transformers model (sentence_transformers.SentenceTransformer).
• Registry: Now version name is no longer required when logging a model. If not provided, a random human readable ID will be generated.

1.3.1 (2024-03-21)

New Features

• FileSet: snowflake.ml.fileset.sfcfs.SFFileSystem can now be used in UDFs and stored procedures.

1.3.0 (2024-03-12)

Bug Fixes

• Registry: Fix a bug that leads to module in code_paths when log_model cannot be correctly imported.
• Registry: Fix incorrect error message when validating input Snowpark DataFrame with array feature.
• Model Registry: Fix an issue when deploying a model to SPCS that some files do not have proper permission.
• Model Development: Relax package versions for all inference methods if the installed version is not available in the Snowflake conda channel

Behavior Changes

• Registry: When running the method of a model, the value range based input validation to avoid input from overflowing is now optional rather than enforced, this should improve the performance and should not lead to problem for most kinds of model. If you want to enable this check as previous, specify strict_input_validation=True when calling run.
• Registry: By default relax_version=True when logging a model instead of using the specific local dependency versions. This improves dependency versioning by using versions available in Snowflake. To switch back to the previous behavior and use specific local dependency versions, specify relax_version=False when calling log_model.
• Model Development: The behavior of fit_predict for all estimators is changed. Firstly, it will cover all the estimator that contains this function, secondly, the output would be the union of pandas DataFrame and snowpark DataFrame.

New Features

• FileSet: snowflake.ml.fileset.sfcfs.SFFileSystem can now be serialized with pickle.

1.2.3 (2024-02-26)

Bug Fixes

• Registry: Now when providing Decimal Type column to a DOUBLE or FLOAT feature will not error out but auto cast with warnings.
• Registry: Improve the error message when specifying currently unsupported pip_requirements argument.
• Model Development: Fix precision_recall_fscore_support incorrect results when average="samples".
• Model Registry: Fix an issue that leads to description, metrics or tags are not correctly returned in newly created Model Registry (PrPr) due to Snowflake BCR 2024_01

Behavior Changes

• Feature Store: FeatureStore.suspend_feature_view and FeatureStore.resume_feature_view doesn't mutate input feature view argument any more. The updated status only reflected in the returned feature view object.

New Features

• Model Development: support score_samples method for all the classes, including Pipeline, GridSearchCV, RandomizedSearchCV, PCA, IsolationForest, ...
• Registry: Support deleting a version of a model.

1.2.2 (2024-02-13)

New Features

• Model Registry: Support providing external access integrations when deploying a model to SPCS. This will help and be required to make sure the deploying process work as long as SPCS will by default deny all network connections. The following endpoints must be allowed to make deployment work: docker.com:80, docker.com:443, anaconda.com:80, anaconda.com:443, anaconda.org:80, anaconda.org:443, pypi.org:80, pypi.org:443. If you are using snowflake.ml.model.models.huggingface_pipeline.HuggingFacePipelineModel object, the following endpoints are required to be allowed: huggingface.com:80, huggingface.com:443, huggingface.co:80, huggingface.co:443.

1.2.1 (2024-01-25)

New Features

• Model Development: Infers output column data type for transformers when possible.
• Registry: relax_version option is available in the options argument when logging the model.

1.2.0 (2024-01-11)

Bug Fixes

• Model Registry: Fix "XGBoost version not compiled with GPU support" error when running CPU inference against open-source XGBoost models deployed to SPCS.
• Model Registry: Fix model deployment to SPCS on Windows machines.

New Features

• Model Development: Introduced XGBoost external memory training feature. This feature enables training XGBoost models on large datasets that don't fit into memory.
• Registry: New Registry class named snowflake.ml.registry.Registry providing similar APIs as the old one but works with new MODEL object in Snowflake SQL. Also, we are providingsnowflake.ml.model.Model and snowflake.ml.model.ModelVersion to represent a model and a specific version of a model.
• Model Development: Add support for fit_predict method in AgglomerativeClustering, DBSCAN, and OPTICS classes;
• Model Development: Add support for fit_transform method in MDS, SpectralEmbedding and TSNE class.

Additional Notes

• Model Registry: The snowflake.ml.registry.model_registry.ModelRegistry has been deprecated starting from version 1.2.0. It will stay in the Private Preview phase. For future implementations, kindly utilize snowflake.ml.registry.Registry, except when specifically required. The old model registry will be removed once all its primary functionalities are fully integrated into the new registry.

1.1.2 (2023-12-18)

Bug Fixes

• Generic: Fix the issue that stack trace is hidden by telemetry unexpectedly.
• Model Development: Execute model signature inference without materializing full dataframe in memory.
• Model Registry: Fix occasional 'snowflake-ml-python library does not exist' error when deploying to SPCS.

Behavior Changes

• Model Registry: When calling predict with Snowpark DataFrame, both inferred or normalized column names are accepted.
• Model Registry: When logging a Snowpark ML Modeling Model, sample input data or manually provided signature will be ignored since they are not necessary.

New Features

• Model Development: SQL implementation of binary precision_score metric.

1.1.1 (2023-12-05)

Bug Fixes

• Model Registry: The predict target method on registered models is now compatible with unsupervised estimators.
• Model Development: Fix confusion_matrix incorrect results when the row number cannot be divided by the batch size.

New Features

• Introduced passthrough_col param in Modeling API. This new param is helpful in scenarios requiring automatic input_cols inference, but need to avoid using specific columns, like index columns, during training or inference.

1.1.0 (2023-12-01)

Bug Fixes

• Model Registry: Fix panda dataframe input not handling first row properly.
• Model Development: OrdinalEncoder and LabelEncoder output_columns do not need to be valid snowflake identifiers. They would previously be excluded if the normalized name did not match the name specified in output_columns.

New Features

• Model Registry: Add support for invoking public endpoint on SPCS service, by providing a "enable_ingress" SPCS deployment option.
• Model Development: Add support for distributed HPO - GridSearchCV and RandomizedSearchCV execution will be distributed on multi-node warehouses.

1.0.12 (2023-11-13)

Bug Fixes

• Model Registry: Fix regression issue that container logging is not shown during model deployment to SPCS.
• Model Development: Enhance the column capacity of OrdinalEncoder.
• Model Registry: Fix unbound batch_size error when deploying a model other than Hugging Face Pipeline and LLM with GPU on SPCS.

Behavior Changes

• Model Registry: Raise early error when deploying to SPCS with db/schema that starts with underscore.
• Model Registry: conda-forge channel is now automatically added to channel lists when deploying to SPCS.
• Model Registry: relax_version will not strip all version specifier, instead it will relax ==x.y.z specifier to >=x.y,<(x+1).
• Model Registry: Python with different patchlevel but the same major and minor will not result a warning when loading the model via Model Registry and would be considered to use when deploying to SPCS.
• Model Registry: When logging a snowflake.ml.model.models.huggingface_pipeline.HuggingFacePipelineModel object, versions of local installed libraries won't be picked as dependencies of models, instead it will pick up some pre- defined dependencies to improve user experience.

New Features

• Model Registry: Enable best-effort SPCS job/service log streaming when logging level is set to INFO.

1.0.11 (2023-10-27)

New Features

• Model Registry: Add log_artifact() public method.
• Model Development: Add support for kneighbors.

Behavior Changes

• Model Registry: Change log_model() argument from TrainingDataset to List of Artifact.
• Model Registry: Change get_training_dataset() to get_artifact().

Bug Fixes

• Model Development: Fix support for XGBoost and LightGBM models using SKLearn Grid Search and Randomized Search model selectors.
• Model Development: DecimalType is now supported as a DataType.
• Model Development: Fix metrics compatibility with Snowpark Dataframes that use Snowflake identifiers
• Model Registry: Resolve 'delete_deployment' not deleting the SPCS service in certain cases.

1.0.10 (2023-10-13)

Behavior Changes

• Model Development: precision_score, recall_score, f1_score, fbeta_score, precision_recall_fscore_support, mean_absolute_error, mean_squared_error, and mean_absolute_percentage_error metric calculations are now distributed.
• Model Registry: deploy will now return Deployment for deployment information.

New Features

• Model Registry: When the model signature is auto-inferred, it will be printed to the log for reference.
• Model Registry: For SPCS deployment, Deployment details will contains image_name, service_spec and service_function_sql.

Bug Fixes

• Model Development: Fix an issue that leading to UTF-8 decoding errors when using modeling modules on Windows.
• Model Development: Fix an issue that alias definitions cause SnowparkSQLUnexpectedAliasException in inference.
• Model Registry: Fix an issue that signature inference could be incorrect when using Snowpark DataFrame as sample input.
• Model Registry: Fix too strict data type validation when predicting. Now, for example, if you have a INT8 type feature in the signature, if providing a INT64 dataframe but all values are within the range, it would not fail.

1.0.9 (2023-09-28)

Behavior Changes

• Model Development: log_loss metric calculation is now distributed.

Bug Fixes

• Model Registry: Fix an issue that building images fails with specific docker setup.
• Model Registry: Fix an issue that unable to embed local ML library when the library is imported by zipimport.
• Model Registry: Fix out-of-date doc about platform argument in the deploy function.
• Model Registry: Fix an issue that unable to deploy a GPU-trained PyTorch model to a platform where GPU is not available.

1.0.8 (2023-09-15)

Bug Fixes

• Model Development: Ordinal encoder can be used with mixed input column types.
• Model Development: Fix an issue when the sklearn default value is np.nan.
• Model Registry: Fix an issue that incorrect docker executable is used when building images.
• Model Registry: Fix an issue that specifying token argument when using snowflake.ml.model.models.huggingface_pipeline.HuggingFacePipelineModel with transformers < 4.32.0 is not effective.
• Model Registry: Fix an issue that incorrect system function call is used when deploying to SPCS.
• Model Registry: Fix an issue when using a transformers.pipeline that does not have a tokenizer.
• Model Registry: Fix incorrectly-inferred image repository name during model deployment to SPCS.
• Model Registry: Fix GPU resource retention issue caused by failed or stuck previous deployments in SPCS.

1.0.7 (2023-09-05)

Bug Fixes

• Model Development & Model Registry: Fix an error related to pandas.io.json.json_normalize.
• Allow disabling telemetry.

1.0.6 (2023-09-01)

New Features

• Model Registry: add create_if_not_exists parameter in constructor.
• Model Registry: Added get_or_create_model_registry API.
• Model Registry: Added support for using GPU inference when deploying XGBoost (xgboost.XGBModel and xgboost.Booster ), PyTorch (torch.nn.Module and torch.jit.ScriptModule) and TensorFlow (tensorflow.Module and tensorflow.keras.Model) models to Snowpark Container Services.
• Model Registry: When inferring model signature, Sequence of built-in types, Sequence of numpy.ndarray, Sequence of torch.Tensor, Sequence of tensorflow.Tensor and Sequence of tensorflow.Tensor can be used instead of only List of them.
• Model Registry: Added get_training_dataset API.
• Model Development: Size of metrics result can exceed previous 8MB limit.
• Model Registry: Added support save/load/deploy HuggingFace pipeline object (transformers.Pipeline) and our wrapper (snowflake.ml.model.models.huggingface_pipeline.HuggingFacePipelineModel) to it. Using the wrapper to specify configurations and the model for the pipeline will be loaded dynamically when deploying. Currently, following tasks are supported to log without manually specifying model signatures:
  • "conversational"
  • "fill-mask"
  • "question-answering"
  • "summarization"
  • "table-question-answering"
  • "text2text-generation"
  • "text-classification" (alias "sentiment-analysis" available)
  • "text-generation"
  • "token-classification" (alias "ner" available)
  • "translation"
  • "translation_xx_to_yy"
  • "zero-shot-classification"

Bug Fixes

• Model Development: Fixed a bug when using simple imputer with numpy >= 1.25.
• Model Development: Fixed a bug when inferring the type of label columns.

Behavior Changes

• Model Registry: log_model() now return a ModelReference object instead of a model ID.
• Model Registry: When deploying a model with 1 target method only, the target_method argument can be omitted.
• Model Registry: When using the snowflake-ml-python with version newer than what is available in Snowflake Anaconda Channel, embed_local_ml_library option will be set as True automatically if not.
• Model Registry: When deploying a model to Snowpark Container Services and using GPU, the default value of num_workers will be 1.
• Model Registry: keep_order and output_with_input_features in the deploy options have been removed. Now the behavior is controlled by the type of the input when calling model.predict(). If the input is a pandas.DataFrame, the behavior will be the same as keep_order=True and output_with_input_features=False before. If the input is a snowpark.DataFrame, the behavior will be the same as keep_order=False and output_with_input_features=True before.
• Model Registry: When logging and deploying PyTorch (torch.nn.Module and torch.jit.ScriptModule) and TensorFlow (tensorflow.Module and tensorflow.keras.Model) models, we no longer accept models whose input is a list of tensor and output is a list of tensors. Instead, now we accept models whose input is 1 or more tensors as positional arguments, and output is a tensor or a tuple of tensors. The input and output dataframe when predicting keep the same as before, that is every column is an array feature and contains a tensor.

1.0.5 (2023-08-17)

New Features

• Model Registry: Added support save/load/deploy xgboost Booster model.
• Model Registry: Added support to get the model name and the model version from model references.

Bug Fixes

• Model Registry: Restore the db/schema back to the session after create_model_registry().
• Model Registry: Fixed an issue that the UDF name created when deploying a model is not identical to what is provided and cannot be correctly dropped when deployment getting dropped.
• connection_params.SnowflakeLoginOptions(): Added support for private_key_path.

1.0.4 (2023-07-28)

New Features

• Model Registry: Added support save/load/deploy Tensorflow models (tensorflow.Module).
• Model Registry: Added support save/load/deploy MLFlow PyFunc models (mlflow.pyfunc.PyFuncModel).
• Model Development: Input dataframes can now be joined against data loaded from staged files.
• Model Development: Added support for non-English languages.

Bug Fixes

• Model Registry: Fix an issue that model dependencies are incorrectly reported as unresolvable on certain platforms.

1.0.3 (2023-07-14)

Behavior Changes

• Model Registry: When predicting a model whose output is a list of NumPy ndarray, the output would not be flattened, instead, every ndarray will act as a feature(column) in the output.

New Features

• Model Registry: Added support save/load/deploy PyTorch models (torch.nn.Module and torch.jit.ScriptModule).

Bug Fixes

• Model Registry: Fix an issue that when database or schema name provided to create_model_registry contains special characters, the model registry cannot be created.
• Model Registry: Fix an issue that get_model_description returns with additional quotes.
• Model Registry: Fix incorrect error message when attempting to remove a unset tag of a model.
• Model Registry: Fix a typo in the default deployment table name.
• Model Registry: Snowpark dataframe for sample input or input for predict method that contains a column with Snowflake NUMBER(precision, scale) data type where scale = 0 will not lead to error, and will now correctly recognized as INT64 data type in model signature.
• Model Registry: Fix an issue that prevent model logged in the system whose default encoding is not UTF-8 compatible from deploying.
• Model Registry: Added earlier and better error message when any file name in the model or the file name of model itself contains characters that are unable to be encoded using ASCII. It is currently not supported to deploy such a model.

1.0.2 (2023-06-22)

Behavior Changes

• Model Registry: Prohibit non-snowflake-native models from being logged.
• Model Registry: _use_local_snowml parameter in options of deploy() has been removed.
• Model Registry: A default False embed_local_ml_library parameter has been added to the options of log_model(). With this set to False (default), the version of the local snowflake-ml-python library will be recorded and used when deploying the model. With this set to True, local snowflake-ml-python library will be embedded into the logged model, and will be used when you load or deploy the model.

New Features

• Model Registry: A new optional argument named code_paths has been added to the arguments of log_model() for users to specify additional code paths to be imported when loading and deploying the model.
• Model Registry: A new optional argument named options has been added to the arguments of log_model() to specify any additional options when saving the model.
• Model Development: Added metrics:
  • d2_absolute_error_score
  • d2_pinball_score
  • explained_variance_score
  • mean_absolute_error
  • mean_absolute_percentage_error
  • mean_squared_error

Bug Fixes

• Model Development: accuracy_score() now works when given label column names are lists of a single value.

1.0.1 (2023-06-16)

Behavior Changes

• Model Development: Changed Metrics APIs to imitate sklearn metrics modules:
  • accuracy_score(), confusion_matrix(), precision_recall_fscore_support(), precision_score() methods move from respective modules to metrics.classification.
• Model Registry: The default table/stage created by the Registry now uses "SYSTEM" as a prefix.
• Model Registry: get_model_history() method as been enhanced to include the history of model deployment.

New Features

• Model Registry: A default False flag named replace_udf has been added to the options of deploy(). Setting this to True will allow overwrite existing UDF with the same name when deploying.
• Model Development: Added metrics:
  • f1_score
  • fbeta_score
  • recall_score
  • roc_auc_score
  • roc_curve
  • log_loss
  • precision_recall_curve
• Model Registry: A new argument named permanent has been added to the argument of deploy(). Setting this to True allows the creation of a permanent deployment without needing to specify the UDF location.
• Model Registry: A new method list_deployments() has been added to enumerate all permanent deployments originating from a specific model.
• Model Registry: A new method get_deployment() has been added to fetch a deployment by its deployment name.
• Model Registry: A new method delete_deployment() has been added to remove an existing permanent deployment.

1.0.0 (2023-06-09)

Behavior Changes

• Model Registry: predict() method moves from Registry to ModelReference.
• Model Registry: _snowml_wheel_path parameter in options of deploy(), is replaced with _use_local_snowml with default value of False. Setting this to True will have the same effect of uploading local SnowML code when executing model in the warehouse.
• Model Registry: Removed id field from ModelReference constructor.
• Model Development: Preprocessing and Metrics move to the modeling package: snowflake.ml.modeling.preprocessing and snowflake.ml.modeling.metrics.
• Model Development: get_sklearn_object() method is renamed to to_sklearn(), to_xgboost(), and to_lightgbm() for respective native models.

New Features

• Added PolynomialFeatures transformer to the snowflake.ml.modeling.preprocessing module.
• Added metrics:
  • accuracy_score
  • confusion_matrix
  • precision_recall_fscore_support
  • precision_score

Bug Fixes

• Model Registry: Model version can now be any string (not required to be a valid identifier)
• Model Deployment: deploy() & predict() methods now correctly escapes identifiers

0.3.2 (2023-05-23)

Behavior Changes

• Use cloudpickle to serialize and deserialize models throughout the codebase and removed dependency on joblib.

New Features

• Model Deployment: Added support for snowflake.ml models.

0.3.1 (2023-05-18)

Behavior Changes

• Standardized registry API with following
  • Create & open registry taking same set of arguments
  • Create & Open can choose schema to use
  • Set_tag, set_metric, etc now explicitly calls out arg name as metric_name, tag_name, metric_name, etc.

New Features

• Changes to support python 3.9, 3.10
• Added kBinsDiscretizer
• Support for deployment of XGBoost models & int8 types of data

0.3.0 (2023-05-11)

Behavior Changes

• Big Model Registry Refresh
  • Fixed API discrepancies between register_model & log_model.
  • Model can be referred by Name + Version (no opaque internal id is required)

New Features

• Model Registry: Added support save/load/deploy SKL & XGB Models

0.2.3 (2023-04-27)

Bug Fixes

• Allow using OneHotEncoder along with sklearn style estimators in a pipeline.

New Features

• Model Registry: Added support for delete_model. Use delete_artifact = False to not delete the underlying model data but just unregister.

0.2.2 (2023-04-11)

New Features

• Initial version of snowflake-ml modeling package.
  • Provide support for training most of scikit-learn and xgboost estimators and transformers.

Bug Fixes

• Minor fixes in preprocessing package.

0.2.1 (2023-03-23)

New Features

• New in Preprocessing:
  • SimpleImputer
  • Covariance Matrix
• Optimization of Ordinal Encoder client computations.

Bug Fixes

• Minor fixes in OneHotEncoder.

0.2.0 (2023-02-27)

New Features

• Model Registry
• PyTorch & Tensorflow connector file generic FileSet API
• New to Preprocessing:
  • Binarizer
  • Normalizer
  • Pearson correlation Matrix
• Optimization in Ordinal Encoder to cache vocabulary in temp tables.

0.1.3 (2023-02-02)

New Features

• Initial version of transformers including:
  • Label Encoder
  • Max Abs Scaler
  • Min Max Scaler
  • One Hot Encoder
  • Ordinal Encoder
  • Robust Scaler
  • Standard Scaler