v0.4.0rc1

Removes explicit requirement for google-vizier to stay compatible with google Colab.

Full Changelog: v0.4.0rc0...v0.4.0rc1

v0.4.0rc0

What's Changed

• Expose tfgnn.dataset_from_generator helper. by @copybara-service in #115
• Add triple converter for RDF-style input. by @copybara-service in #119
• Doc fix: The the intro page describes GNNs in some generality. by @copybara-service in #123
• Enables disabled runner tests from RNG changes in Keras by @copybara-service in #122
• Fixes error with named features in homogeneous constructor by @copybara-service in #121
• Let DatasetProviders raise FileNotFoundError if no files match pattern. by @copybara-service in #139
• Helper function for random reordering within segment ids. by @copybara-service in #135
• Updates Bazel dep to 2.8.0. by @copybara-service in #150
• In a GraphTensor guide example, add tf.constant(...) to clarify by @copybara-service in #162
• Shorten function name to_sampling_spec to build, maintaining backward-compatibility. by @copybara-service in #159
• Add streaming API to read Unigraph artifacts (node sets, edge sets) as iterables. by @copybara-service in #166
• Unsupervised (i.e.: model generated target vs. dataset generated label) refactors. by @copybara-service in #165
• Check checkpoint_options.checkpoint_dir: it may be unfilled in a caller's checkpoint_options. by @copybara-service in #172
• Multitask refactors. by @copybara-service in #181
• s/class_id/sparse_class_id/g by @copybara-service in #189
• Make runner/examples/ogbn/mag/train.py parametric in the GNN model type. by @copybara-service in #191
• Fix typo. by @copybara-service in #192
• Use strings to set BQ read method. by @copybara-service in #199
• Fix a typo/ by @copybara-service in #206
• Local tf import to avoid global session. by @copybara-service in #208

Full Changelog: v0.3.0...v0.4.0rc0

v0.3.0

What's Changed

• An initial version of transformer-style multi-head attention. in #55
• Add TPUStrategy support to the runner. in #51
• Add a keys variant for transformer-style multi-head attention. in #56
• Enable dataflow prime and add edge_aggregation_method choice to example. in #68
• Fix typos in gnn_modelling.md by @dgoekmen in #70
• In type spec deserialization, ensure class DType in #83
• Integer mathematics (i.e., TPU-friendly) for adding self-loops on GraphTensors. in #75
• Export symbol tfgnn.is_ragged_tensor(). in #87
• Add per-class metrics for multi-class classification. in #81
• Initial support for graph pieces shapes relaxation (nodes and edges). Helper function to easy dataset creation from scalar graph pieces generator. in #84
• Adds tfgnn.keras.layers.AddSelfLoops in #86
• Remove tfgnn.keras.layers.TotalSize in favor of .total_size and in #96
• Improve the error message when num_classes and class_names are both None. in #99
• Compute SizeConstraints for TPU padding before calling distribute_datasets_from_function. in #77
• Builder pattern for SamplingSpec proto. in #104
• Ensure that L2 regularization for GCN is correctly [de]serialized. in #105
• Pin google-api-core version due to protobuf v4.21.0 tensorflow incompatibility. in #109

New Contributors

• @dgoekmen made their first contribution in #70

Full Changelog: v0.2.0.dev1...v0.3.0

v0.2.1

Removes GCP dependencies from setup.py and updates apache-beam.

Full Changelog: v0.2.0...v0.2.1

v0.2.0

Summary

• Added 3 popular model architectures from literature (GATv2, GraphSAGE, and GCN) and a commonly-used MPNN architecture from within Google.
• Adds an orchestrator toolkit for low-code solutions on common graph learning tasks
• Added a flexible sampler implementation which supports edge-based and node-based sampling.

Full Changelog: v0.2.0.dev1...v0.2.0

v0.2.0rc0

Initial release candidate (rc0) for 0.2.0 release. Includes a new sampler, documentation, and examples.

v0.2.0.dev1

Pre-release for v0.2.0 of TensorFlow GNN with example models (GCN, GATv2, GraphSAGE) and a runner.

This is a v0 release, with no API compatibility guarantees.