Snowpark ML release notes¶

Bug fixes¶

Model Registry bug fixes:

• Fix “401 Unauthorized” issue when deploying a model to Snowpark Container Services.

Feature Store bug fixes:

• Some exceptions in property setters have been downgraded to warnings, allowing you to change desc, refresh_freq, and warehouse in “draft” feature views.

Modeling bug fixes:

• Fix issues with calling OneHotEncoder and OrdinalEncoder with a dictionary as the categories parameter and the data in a pandas DataFrame.

New features¶

New Model Registry features:

• Allow overriding device_map and device when loading Hugging Face pipeline models.

• Add set_alias and unset_alias methods to ModelVersion instances to manage the model version’s aliases.

• Add partitioned_inference_api decorator to create partitioned inference methods in models.

New Feature Store features:

• New refresh_freq, refresh_mode, and scheduling_state columns have been added to the output of the list_feature_views method.

• The update_feature_view method now supports updating a feature view’s description.

• New methods refresh_feature_view and get_refresh_history manage updates of feature views.

• New method generate_training_set generates table-backed feature snapshots. generate_dataset(..., output_type="table") has been deprecated and generates a DeprecationWarning.

New Modeling features:

• OneHotEncoder and OrdinalEncoder now accept a list of array-like values for the categories argument.

Bug fixes¶

Model Registry bug fixes:

• Fix an issue causing incorrect results when using a pandas Dataframe with over 100,000 rows as the input of ModelVersion.run method in Stored Procedures.

Modeling bug fixes:

• Fix an issue with passing categories to OneHotEncoder and OrdinalEncoder as a dictionary or as a pandas DataFrame.

New features¶

New Model Registry features:

• Model Registry now supports timestamp (TIMESTAMP_NTZ) columns in input and output data.

New modeling features:

• OneHotEncoder and OrdinalEncoder now support a list of array-like values for the categories argument.

New Dataset features:

• DatasetVersion instances now have label_cols and exclude_cols properties.

Bug fixes¶

Model Registry bug fixes:

• Fixed an issue that prevented calls to log_model in a stored procedure.

Modeling bug fixes:

• Quick fix for import snowflake.ml.modeling.parameters.enable_anonymous_sproc not working due to package dependency error.

New features¶

New Model Registry features:

• log_model, get_model, and delete_model methods now support fully-qualified names.

New modeling features:

• You can now use an anonymous stored procedure during fitting, so that modeling does not require privileges to operate on the registry schema. Call import snowflake.ml.modeling.parameters.enable_anonymous_sproc to enable this feature.

Bug fixes¶

Model registry bug fixes:

• Fix issue with loading older models.

Behavior changes¶

Model Registry behavior changes:

• The fit_transform method can now return either a Snowpark DataFrame or a pandas DataFrame, matching the kind of DataFrame passed to the method.

New features¶

New Model Registry features:

• Added support for exporting models from the registry (ModelVersion.export).

• Added support for loading the underlying model object (ModelVersion.load).

• Added support for renaming models (Model.rename).

Bug fixes¶

Model Registry bug fixes:

• Fixed the “invalid parameter SHOW_MODEL_DETAILS_IN_SHOW_VERSIONS_IN_MODEL” error.

New features¶

New Model Registry features:

• Added support for catboost models (catboost.CatBoostClassifier, catboost.CatBoostRegressor).

• Added support for lightgbm models (lightgbm.Booster, lightgbm.LightGBMClassifier, lightgbm.LightGBMRegressor).

Bug fixes¶

Model Registry bug fixes:

• Fixed bug that caused relax_version option to not work.

Behavior changes¶

Model Registry behavior changes:

• The apply method is no longer included as a target method by default when logging an XGBoost model. If you need this method available in logged models, included it manually in the target-methods option:

  log_model(..., options={"target_methods": ["apply", ...]})
  

  Copy

New features¶

New model registry features:

• The registry now supports logging sentence transformer models (sentence_transformers.SentenceTransformer).

• The version_name argument is no longer required when logging a model. A random human-readable ID is generated if none is provided.

Bug fixes¶

Model registry bug fixes:

• Fix issue where, when multiple models are called in the same query, models after the first returned incorrect results. This fix is applied when models are logged and does not benefit existing models; you must log your models again to correct this behavior.

Modeling bug fixes:

• Fix bug in registering a model where only methods mentioned in save_model were added to the model signature for Snowpark ML models.

• Fix bug in batch inference methods such as such as predict and predict_log_probe where, when n_jobs was not 1, the methods would not be executed.

• Fix bug in batch inference methods where they could not infer datatypes when the first row of data contained NULL.

• The output column names from distributed hyperparameter optimization are now correctly matched with the Snowflake identifier.

• Relaxed the versions of dependencies of distributed hyperparameter optimization methods; these were too strict and caused these methods to fail.

• scikit-learn is now listed as a dependency of the LightGBM package.

New features¶

FileSet/FileSystem updates:

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

Behavior changes¶

Model registry behavior changes:

• As previously announced, the default for the relax_version option (in the options argument of log_model) is now True, allowing more reliable deployment in most cases by permitting dependency versions available in Snowflake.

• When running model methods, value range based input validation (which prevents input from overflowing) is now optional. This should improve performance and should not lead to issues for most types of models. To enable validation, pass the named argument strict_input_validation=True when calling the model’s run method.

Model development behavior changes:

• The fit_predict method now returns either a pandas or a Snowpark DataFrame, depending on the type of the input data, and is available on all classes where it is available in the underlying scikit-learn, xgboost, or lightgbm class.

New features and updates¶

FileSet/FileSystem updates:

• Instances of snowflake.ml.fileset.sfcfs.SFFileSystem can now be serialized with pickle.

Bug fixes¶

Model registry bug fixes:

• Fix a problem with importing log_model in some circumstances.

• Fix an incorrect error message when validating input Snowpark DataFrame with an array feature.

Model development bug fixes:

• Relax package versions for all inference methods when the installed version of a dependency is not available in the Snowflake conda channel.

New features and updates¶

Model development updates:

• All modeling classes now include a score_samples method to calculate the log-likelihood of the given samples.

Model registry updates:

• Decimal type features are automatically cast (with a warning) to a DOUBLE or FLOAT instead of producing an error.

• Improve error message for currently-unsupported pip-requirements option.

• You can now delete a version of a model.

Bug fixes¶

Model development fixes:

• precision_recall_fscore_support returned incorrect results with average="samples".

Model registry fixes:

• Descriptions, models, and tags were not retrieved correctly in newly-created registries under the private preview model registry API due to a recent Snowflake behavior change.

New features and updates¶

Model registry updates:

• You can now specify external access integrations when deploying a model to Snowpark Container Services using the private preview registry API, allowing models to access the internet to retrieve dependencies during deployment. The following endpoints are required for all deployments:

  • docker.com:80

  • docker.com:443

  • anaconda.com:80

  • anaconda.com:443

  • anaconda.org:80

  • anaconda.org:443

  • pypi.org:80

  • pypi.org:443

  For models derived from HuggingFacePipeLineModel, the following endpoints are required.

  • huggingface.com:80

  • huggingface.com:443

  • huggingface.co:80

  • huggingface.co:443

New features and updates¶

Model development updates:

• Infer column data type for transformers when possible.

Model registry updates:

• relax_version option (in options argument of log_model) relaxes dependencies of stated versions to allow newer minor versions when set to True.

New features and updates¶

Public preview release of model registry. See Snowflake Model Registry. The previous private preview release of the model registry has been deprecated, but will continue to be supported while it includes features not yet available in the public preview version.

Model development updates:

• Added support for fit_predict method in AgglomerativeClustering, DBSCAN, and OPTICS classes.

• Added support for fit_transform method in MDS, SpectralEmbedding and TSNE class.