2.6.7

ADS

• Fixed a bug in PyTorchModel. The score.py failed when torch.Tensor was used as input data.
• Added support for flexible shapes for Data Flow Jobs.
• Loading a model from Model Catalog (GenericModel.from_model_catalog()) and Model Deployment (GenericModel.from_model_deployment()) no longer requires a model file name.
• Switched from using cx_Oracle interface to the oracledb driver to connect to Oracle Databases.
• Added support for image attribute for the PyTorchModel.predict() and TensorFlowModel.predict() methods. Images can now be directly passed to the model Deployment predict.

The following APIs are deprecated:

• OracleAutoMLProvider

2.6.6

ADS

• Added SparkPipelineModel model serialization class for fast and easy model deployment.
• Added support for flexible shapes for Jobs and Model Deployments.
• Added support for freeform_tags and defined_tags for Model Deployments.
• Added the populate_schema() method to the GenericModel class. Populate input and output schemas for model artifacts.
• The ADSString was added to the Feature types system. Use the enhanced string class functionalities such as regular expression (RegEx) matching and natural language parsing within Pandas dataframes and series.
• Saving model does not require iPython dependencies

Following APIs are deprecated:

• DatasetFactory.open
• ADSModel.prepare
• ads.common.model_export_util.prepare_generic_model

2.6.5

ADS

• OCI SDK updated from version 2.59.0 to version 2.82.0.

2.6.4

ADS

• Added support for large models with artifact size between 2 and 6 GB. The large models can be saved to the Model Catalog, downloaded from the Model Catalog, and deployed as a Model Deployment resource.
• Added delete() method to the GenericModel class. Deletes models and associated model deployments.
• The Model Input Schema is improved to return features sorted by the order attribute.
• Added user-friendly default names for created Jobs, Model Deployments, and Models.

2.6.3

ADS

• Deprecated the ads.dataflow.DataFlow class. It has been superseded by the ads.jobs.DataFlow class.
• Added prepare_save_deploy() method to the GenericModel class. Prepare model artifacts and deploy the model with one command.
• Added support for binary payloads in model deployment.
• Updated AutoMLModel, GenericModel, LightgbmModel, PyTorchModel, SklearnModel, TensorflowModel, and XgboostModel classes to support binary payloads in model deployment.
• The maximum runtime for a Job can be limited with the with_maximum_runtime_in_minutes() method in the CondaRuntime, DataFlowNotebookRuntime, DataFlowRuntime, GitPythonRuntime, NotebookRuntime, and ScriptRuntime classes.
• The ads.jobs.DataFlow class supports Published conda environments.

2.6.2

ADS

• Added from_model_deployment() method to the GenericModel class. Now you can load a model directly from an existing model deployment.

• Moved dependencies from being default into optional installation groups:

  • all-optional
  • bds
  • boosted
  • data
  • geo
  • notebook
  • onnx
  • opctl
  • optuna
  • tensorflow
  • text
  • torch
  • viz

  Use python3 -m pip install oracle-ads[XXX] where XXX are the group names.

2.6.1

ADS

• Added support for running a container as jobs using ads.jobs.ContainerRuntime.
• The ModelArtifact class is deprecated. Use the model serialization classes (GenericModel, PyTorchModel, SklearnModel, etc.).

2.5.10

ADS

• Added BDSSecretKeeper to store and save configuration parameters to connect to Big Data service to the vault.
• Added the krbcontext and refresh_ticket functions to configure Kerberos authentication for the Big Data service.
• Added authentication options to logging APIs to allow you to pass in the OCI API key configuration or signer.
• Added the configuration file path option to the set_auth method to allow to change the path of the OCI configuration.
• Fixed a bug in AutoML for Ttext datasets.
• Fixed bug in import ads.jobs to notify users installing ADS optional dependencies.
• Fixed a bug in the generated score.py file, where Pandas dataframe's dtypes changed when deserializing. Now you can recover it from the input schema.
• Updated requirements to oci>=2.59.0.

2.5.9

ADS

• Added framework specific model serialization to add more inputs to the generated score.py file.

• Added the following framework-specific model classes:

  • AutoMLModel
  • SKlearnModel
  • XGBoostModel
  • LightGBMModel
  • PyTorchModel
  • TensorFlowModel

• For any framework not included in the preceding list, added another class:

  • GenericModel

• These model classes include methods specific to the frameworks that improve deployment speed. Some example methods are:

  • Prepare (the artifacts)
  • Save (metadata and model to model catalog)
  • Deploy (the models quickly with this method)
  • Predict (perform inference operations)

• Added support to create jobs with managed egress.

• Shortened the time for streaming large number of logs for job run logging.

2.5.8

ADS

• Fixed bug in automatic extraction of taxonomy metadata for Sklearn models.
• Fixed bug in jobs NotebookRuntime when using non-ASCII encoding.
• Added compatibility with Python 3.8 and 3.9.
• Added an enhanced string class, called ADSString. It adds functionality such as regular expression (RegEx) matching, and natural language processing (NLP) parsing. The class can be expanded by registering custom plugins to perform custom string processing actions.