Deep Learning Pipelines migration guide

This page includes tips for migrating from the open source Deep Learning Pipelines package that was included in Databricks Runtime 6.6 ML and below. Parts of the Deep Learning Pipelines library sparkdl were removed in Databricks Runtime 7.0 ML (EoS), specifically, the Transformers and Estimators used in Apache Spark ML pipelines.

This page is not a resource for general information about deep learning pipelines on Azure Databricks.

Read images

The Deep Learning Pipelines package includes an image reader sparkdl.image.imageIO, which was removed in Databricks Runtime 7.0 ML (EoS).

Instead, use the image data source or binary file data source from Apache Spark. Many of the example notebooks in Load data for machine learning and deep learning show use cases of these two data sources.

Transfer learning

The Deep Learning Pipelines package includes a Spark ML Transformer sparkdl.DeepImageFeaturizer for facilitating transfer learning with deep learning models. DeepImageFeaturizer was removed in Databricks Runtime 7.0 ML (EoS).

Instead, use pandas UDFs to perform featurization with deep learning models. pandas UDFs, and their newer variant Scalar Iterator pandas UDFs, offer more flexible APIs, support more deep learning libraries, and provide better performance.

See Featurization for transfer learning for examples of transfer learning with pandas UDFs.

Distributed hyperparameter tuning

The Deep Learning Pipelines package includes a Spark ML Estimator sparkdl.KerasImageFileEstimator for tuning hyperparameters using Spark ML tuning utilities. KerasImageFileEstimator was removed in Databricks Runtime 7.0 ML (EoS).

Instead, use Hyperopt to distribute hyperparameter tuning for deep learning models.

Distributed inference

The Deep Learning Pipelines package includes several Spark ML Transformers for distributing inference, all of which were removed in Databricks Runtime 7.0 ML (EoS):

• DeepImagePredictor
• TFImageTransformer
• KerasImageFileTransformer
• TFTransformer
• KerasTransformer

Instead, use pandas UDFs to run inference on Spark DataFrames, following the examples in Deploy models for batch inference and prediction.

Deploy models as SQL UDFs

The Deep Learning Pipelines package includes a utility sparkdl.udf.keras_image_model.registerKerasImageUDF for deploying a deep learning model as a UDF callable from Spark SQL. registerKerasImageUDF was removed in Databricks Runtime 7.0 ML (EoS).

Instead, use MLflow to export the model as a UDF, following the example in scikit-learn model deployment on Azure ML.