ML Workbench Release Notes

TigerGraph ML Workbench 1.4

Released May 16th, 2023.

Updates made at the pyTigerGraph layer are denoted with (pyTigerGraph), and ML Workbench functionality is denoted with (ML Workbench).

If you are using the Graph Data Science dataloaders, you must upgrade both the ML Workbench and pyTigerGraph to 1.4 at the same time. There is an incompatibility between v1.3 and v1.4 of pyTigerGraph and the corresponding ML Workbench versions.

New features

  • Built in Graph ML models and Trainer (pyTigerGraph)

  • Transforms (pyTigerGraph)

    • PyGTemporalTransform to create a sequence of subgraphs for a given batch of data, in a temporal manner.

    • NodePieceMLPTransform to transform a batch produced by a NodePiece dataloader into a batch that can be fed into a PyTorch multilayer perceptron.

  • SSL Support in Kafka Dataloaders (ML Workbench & pyTigerGraph)

    • Two-way SSL encryption via Kerberos.

  • Collaborative dataloaders (ML Workbench & pyTigerGraph)

    • Use dataloaders on multiple machines to pull batches from the same Kafka queue. Helpful for data distributed model training.

  • Additional Dataloader Support (See pyTigerGraph Release Notes)

    • DATETIME support in dataloaders.

    • Optional distributed_query parameter in dataloaders if running on distributed database clusters.

    • stop() function in dataloaders.

Updated features

  • Dataloader factory produces multiple dataloaders if filter_by is a list of different filters. (pyTigerGraph)

  • Improved the scalability of the NodePiece dataloader. (pyTigerGraph)

TigerGraph ML Workbench 1.3

Released April 2023

New features

  • Two new data loader classes: HGTLoader and NodePiece.

  • Callback functions to all data loader factories: users can write functions to process the batch in a background thread before it is passed into the training loop.

  • A delimiter parameter to all data loader factories: users can choose what delimiter they want to separate attributes they are loading from the graph.

Updated features

  • Template query support in the featurizer (requires TigerGraph Database version 3.9+).

  • Data splitters to automatically perform a schema change if needed to add attributes to the database.

  • Support for multi-edges in the upsertEdges() function.

  • Better error messaging.

Fixes

  • Fixed SSL certificate handling when custom certificate is used.

  • Removed ANSI escape characters from output of .gsql() calls.

  • Improved the logic for shuffling vertices in dataloaders when filter_by attribute was used.

  • Improved the batching algorithm in dataloaders so that the output batches have a consistent batch size.

  • Changed getVertexCount() endpoint to more scalable solution.

TigerGraph ML Workbench 1.1 (September 2022)

New features

  • TensorFlow support for homogeneous GNNs via the Spektral library

  • Heterogeneous Graph Dataloading support for DGL

  • Support for lists of strings in dataloaders

  • Activator for both Community and Enterprise editions of the ML Workbench (see https://act.tigergraphlabs.com for details)

Updated features

  • Fixed KeyError when creating a data loader on a graph where PrimaryIdAsAttribute is False

  • Error catch if Kafka dataloader doesn’t run in asynchronous mode

  • Schema now refreshes during dataloader instantiation and featurizer attribute addition

  • Connection instantiation time reduced

  • Reinstall query if it is disabled

  • Confirm Kafka topic is created before subscription

  • Streamlined Kafka resource usage

  • Allow multiple consumers on the same data

  • Improved deprecation warnings

TigerGraph ML Workbench 1.0 (August 2022)

New features

Soft launch of TigerGraph ML Workbench on Cloud, an end-to-end Kubeflow-managed cloud platform for training and serving machine learning models.

  • Complete KubeFlow integration

  • Fully-managed infrastructure orchestrated by Kubernetes

  • Connection to TigerGraph Cloud Solutions

  • Cloud-hosted Jupyter Notebooks

  • TensorBoard integration

  • Experiments with AutoML (beta)

Known Issues

  • When creating a new Notebook, the user is prompted for Configurations and Affinity/Tolerations. These have no effect on the notebook.