Snowflake-ml-python

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.

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.

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- Model Registry: Fix out-of-date doc about `platform` argument in the `deploy` function.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.

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.

Bug Fixes

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

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.