gnn

Release 1.0 is the first with a stable public API.

What's Changed in r1.0

• Overall
  • Supported TF/Keras versions moved to >=2.12,<2.16, incompatible Keras v3 raises a clear error.
  • Importing the library no longer leaks private module names.
  • All parts of the GraphSchema protobuf are now exposed undertfgnn.proto.*.
  • Model saving now clearly distinguishes export to inference (pure TF, fully supported) from misc ways of saving for model reuse.
  • Numerous small bug fixes.
• Subgraph sampling: major upgrade
  • New and unified sampler for in-memory and beam-based subgraph sampling.
  • Module tfgnn.experimental.in_memory is removed in favor of the new sampler.
  • New console script tfgnn_sampler replaces the old tfgnn_graph_sampler.
• GraphTensor
  • Most tfgnn.* functions on GraphTensor now work in Keras' Functional API, including the factory methods GraphTensor.from_pieces(...) etc.
  • New static checks for GraphTensor field shapes, opt out with tfgnn.disable_graph_tensor_validation().
  • New runtime checks for GraphTensor field shapes, sizes and index ranges, opt in with tfgnn.enable_graph_tensor_validation_at_runtime().
  • GraphTensor maintains .row_splits_dtype separately from .indices_dtype.
  • The GraphSchema and the I/O functions for tf.Example now support all non-quantized, non-complex floating-point and integer types as well as bool and string.
  • Added convenience wrapper tfgnn.pool_neighbors_to_node().
  • Misc fixes to tfgnn.random_graph_tensor(), now respects component boundaries.
• Runner
  • New tasks for link prediction and node classification/regression based on structured readout.
  • Now comes with API docs.
• Models collection
  • models/contrastive_losses gets multiple extensions, including a triplet loss and API docs.
  • models/multi_head_attention replaces sigmoid with elu+1 in trained scaling.
  • Bug fixes for mixed precision.

Full Changelog: https://github.com/tensorflow/gnn/compare/v0.6.1...v1.0.0

What's Changed in v1.0.2 over v1.0.1

• Bugfixes for use with TF 2.14 and 2.15 in case tf_keras is installed but not used as tf.keras (https://github.com/tensorflow/gnn/commit/ffa453f712fcf07a2a918726fd3cf622e39f9a50) (https://github.com/tensorflow/gnn/commit/e1d921051d22c10d6f322017f496408b78a98452).

Full Changelog: https://github.com/tensorflow/gnn/compare/v1.0.1...v1.0.2

What's Changed in v1.0.1 over v1.0.0

• Bugfix for regression tasks runner.GraphMean*Error: the reduce_type is again passed through correctly (https://github.com/tensorflow/gnn/commit/19c10f2489e3e62efad0fefe27aa9f5002bd2ce5).

Full Changelog: https://github.com/tensorflow/gnn/compare/v1.0.0...v1.0.1

Release 1.0 is the first with a stable public API.

What's Changed in r1.0

• Overall
  • Supported TF/Keras versions moved to >=2.12,<2.16, incompatible Keras v3 raises a clear error.
  • Importing the library no longer leaks private module names.
  • All parts of the GraphSchema protobuf are now exposed undertfgnn.proto.*.
  • Model saving now clearly distinguishes export to inference (pure TF, fully supported) from misc ways of saving for model reuse.
  • Numerous small bug fixes.
• Subgraph sampling: major upgrade
  • New and unified sampler for in-memory and beam-based subgraph sampling.
  • Module tfgnn.experimental.in_memory is removed in favor of the new sampler.
  • New console script tfgnn_sampler replaces the old tfgnn_graph_sampler.
• GraphTensor
  • Most tfgnn.* functions on GraphTensor now work in Keras' Functional API, including the factory methods GraphTensor.from_pieces(...) etc.
  • New static checks for GraphTensor field shapes, opt out with tfgnn.disable_graph_tensor_validation().
  • New runtime checks for GraphTensor field shapes, sizes and index ranges, opt in with tfgnn.enable_graph_tensor_validation_at_runtime().
  • GraphTensor maintains .row_splits_dtype separately from .indices_dtype.
  • The GraphSchema and the I/O functions for tf.Example now support all non-quantized, non-complex floating-point and integer types as well as bool and string.
  • Added convenience wrapper tfgnn.pool_neighbors_to_node().
  • Misc fixes to tfgnn.random_graph_tensor(), now respects component boundaries.
• Runner
  • New tasks for link prediction and node classification/regression based on structured readout.
  • Now comes with API docs.
• Models collection
  • models/contrastive_losses gets multiple extensions, including a triplet loss and API docs.
  • models/multi_head_attention replaces sigmoid with elu+1 in trained scaling.
  • Bug fixes for mixed precision.

Full Changelog: https://github.com/tensorflow/gnn/compare/v0.6.1...v1.0.0

What's Changed in v1.0.2 over v1.0.1

• Bugfixes for use with TF 2.14 and 2.15 in case tf_keras is installed but not used as tf.keras (https://github.com/tensorflow/gnn/commit/ffa453f712fcf07a2a918726fd3cf622e39f9a50) (https://github.com/tensorflow/gnn/commit/e1d921051d22c10d6f322017f496408b78a98452).

Full Changelog: https://github.com/tensorflow/gnn/compare/v1.0.1...v1.0.2rc1

What's Changed in v1.0.1 over v1.0.0

• Bugfix for regression tasks runner.GraphMean*Error: the reduce_type is again passed through correctly (https://github.com/tensorflow/gnn/commit/19c10f2489e3e62efad0fefe27aa9f5002bd2ce5).

Full Changelog: https://github.com/tensorflow/gnn/compare/v1.0.0...v1.0.1

Release 1.0 is the first with a stable public API.

What's Changed in r1.0

• Overall
  • Supported TF/Keras versions moved to >=2.12,<2.16, incompatible Keras v3 raises a clear error.
  • Importing the library no longer leaks private module names.
  • All parts of the GraphSchema protobuf are now exposed undertfgnn.proto.*.
  • Model saving now clearly distinguishes export to inference (pure TF, fully supported) from misc ways of saving for model reuse.
  • Numerous small bug fixes.
• Subgraph sampling: major upgrade
  • New and unified sampler for in-memory and beam-based subgraph sampling.
  • Module tfgnn.experimental.in_memory is removed in favor of the new sampler.
  • New console script tfgnn_sampler replaces the old tfgnn_graph_sampler.
• GraphTensor
  • Most tfgnn.* functions on GraphTensor now work in Keras' Functional API, including the factory methods GraphTensor.from_pieces(...) etc.
  • New static checks for GraphTensor field shapes, opt out with tfgnn.disable_graph_tensor_validation().
  • New runtime checks for GraphTensor field shapes, sizes and index ranges, opt in with tfgnn.enable_graph_tensor_validation_at_runtime().
  • GraphTensor maintains .row_splits_dtype separately from .indices_dtype.
  • The GraphSchema and the I/O functions for tf.Example now support all non-quantized, non-complex floating-point and integer types as well as bool and string.
  • Added convenience wrapper tfgnn.pool_neighbors_to_node().
  • Misc fixes to tfgnn.random_graph_tensor(), now respects component boundaries.
• Runner
  • New tasks for link prediction and node classification/regression based on structured readout.
  • Now comes with API docs.
• Models collection
  • models/contrastive_losses gets multiple extensions, including a triplet loss and API docs.
  • models/multi_head_attention replaces sigmoid with elu+1 in trained scaling.
  • Bug fixes for mixed precision.

Full Changelog: https://github.com/tensorflow/gnn/compare/v0.6.1...v1.0.0

What's Changed in v1.0.2 over v1.0.1

• Bugfix for use with TF 2.14 and 2.15 in case tf_keras is installed but not used as tf.keras (https://github.com/tensorflow/gnn/commit/ffa453f712fcf07a2a918726fd3cf622e39f9a50).

Full Changelog: https://github.com/tensorflow/gnn/compare/v1.0.1...v1.0.2rc0

What's Changed in v1.0.1 over v1.0.0

• Bugfix for regression tasks runner.GraphMean*Error: the reduce_type is again passed through correctly (https://github.com/tensorflow/gnn/commit/19c10f2489e3e62efad0fefe27aa9f5002bd2ce5).

Full Changelog: https://github.com/tensorflow/gnn/compare/v1.0.0...v1.0.1

Release 1.0 is the first with a stable public API.

What's Changed in r1.0

• Overall
  • Supported TF/Keras versions moved to >=2.12,<2.16, incompatible Keras v3 raises a clear error.
  • Importing the library no longer leaks private module names.
  • All parts of the GraphSchema protobuf are now exposed undertfgnn.proto.*.
  • Model saving now clearly distinguishes export to inference (pure TF, fully supported) from misc ways of saving for model reuse.
  • Numerous small bug fixes.
• Subgraph sampling: major upgrade
  • New and unified sampler for in-memory and beam-based subgraph sampling.
  • Module tfgnn.experimental.in_memory is removed in favor of the new sampler.
  • New console script tfgnn_sampler replaces the old tfgnn_graph_sampler.
• GraphTensor
  • Most tfgnn.* functions on GraphTensor now work in Keras' Functional API, including the factory methods GraphTensor.from_pieces(...) etc.
  • New static checks for GraphTensor field shapes, opt out with tfgnn.disable_graph_tensor_validation().
  • New runtime checks for GraphTensor field shapes, sizes and index ranges, opt in with tfgnn.enable_graph_tensor_validation_at_runtime().
  • GraphTensor maintains .row_splits_dtype separately from .indices_dtype.
  • The GraphSchema and the I/O functions for tf.Example now support all non-quantized, non-complex floating-point and integer types as well as bool and string.
  • Added convenience wrapper tfgnn.pool_neighbors_to_node().
  • Misc fixes to tfgnn.random_graph_tensor(), now respects component boundaries.
• Runner
  • New tasks for link prediction and node classification/regression based on structured readout.
  • Now comes with API docs.
• Models collection
  • models/contrastive_losses gets multiple extensions, including a triplet loss and API docs.
  • models/multi_head_attention replaces sigmoid with elu+1 in trained scaling.
  • Bug fixes for mixed precision.

Full Changelog: https://github.com/tensorflow/gnn/compare/v0.6.1...v1.0.0

What's Changed in v1.0.1 over v1.0.0

• Bugfix for regression tasks runner.GraphMean*Error: the reduce_type is again passed through correctly
  (https://github.com/tensorflow/gnn/commit/19c10f2489e3e62efad0fefe27aa9f5002bd2ce5).

Full Changelog: https://github.com/tensorflow/gnn/compare/v1.0.0...v1.0.1

First release with a stable public API.

What's Changed

• Overall
  • Supported TF/Keras versions moved to >=2.12,<2.16, incompatible Keras v3 raises a clear error.
  • Importing the library no longer leaks private module names.
  • All parts of the GraphSchema protobuf are now exposed undertfgnn.proto.*.
  • Model saving now clearly distinguishes export to inference (pure TF, fully supported) from misc ways of saving for model reuse.
  • Numerous small bug fixes.
• Subgraph sampling: major upgrade
  • New and unified sampler for in-memory and beam-based subgraph sampling.
  • Module tfgnn.experimental.in_memory is removed in favor of the new sampler.
  • New console script tfgnn_sampler replaces the old tfgnn_graph_sampler.
• GraphTensor
  • Most tfgnn.* functions on GraphTensor now work in Keras' Functional API, including the factory methods GraphTensor.from_pieces(...) etc.
  • New static checks for GraphTensor field shapes, opt out with tfgnn.disable_graph_tensor_validation().
  • New runtime checks for GraphTensor field shapes, sizes and index ranges, opt in with tfgnn.enable_graph_tensor_validation_at_runtime().
  • GraphTensor maintains .row_splits_dtype separately from .indices_dtype.
  • The GraphSchema and the I/O functions for tf.Example now support all non-quantized, non-complex floating-point and integer types as well as bool and string.
  • Added convenience wrapper tfgnn.pool_neighbors_to_node().
  • Misc fixes to tfgnn.random_graph_tensor(), now respects component boundaries.
• Runner
  • New tasks for link prediction and node classification/regression based on structured readout.
  • Now comes with API docs.
• Models collection
  • models/contrastive_losses gets multiple extensions, including a triplet loss and API docs.
  • models/multi_head_attention replaces sigmoid with elu+1 in trained scaling.
  • Bug fixes for mixed precision.

Full Changelog: https://github.com/tensorflow/gnn/compare/v0.6.1...v1.0.0

First release with a stable public API.

What's Changed

• Overall
  • Supported TF/Keras versions moved to >=2.12,<2.16, incompatible Keras v3 raises a clear error.
  • Importing the library no longer leaks private module names.
  • All parts of the GraphSchema protobuf are now exposed undertfgnn.proto.*.
  • Model saving now clearly distinguishes export to inference (pure TF, fully supported) from misc ways of saving for model reuse.
  • Numerous small bug fixes.
• Subgraph sampling: major upgrade
  • New and unified sampler for in-memory and beam-based subgraph sampling.
  • Module tfgnn.experimental.in_memory is removed in favor of the new sampler.
  • New console script tfgnn_sampler replaces the old tfgnn_graph_sampler.
• GraphTensor
  • Most tfgnn.* functions on GraphTensor now work in Keras' Functional API, including the factory methods GraphTensor.from_pieces(...) etc.
  • New static checks for GraphTensor field shapes, opt out with tfgnn.disable_graph_tensor_validation().
  • New runtime checks for GraphTensor field shapes, sizes and index ranges, opt in with tfgnn.enable_graph_tensor_validation_at_runtime().
  • GraphTensor maintains .row_splits_dtype separately from .indices_dtype.
  • The GraphSchema and the I/O functions for tf.Example now support all non-quantized, non-complex floating-point and integer types as well as bool and string.
  • Added convenience wrapper tfgnn.pool_neighbors_to_node().
  • Misc fixes to tfgnn.random_graph_tensor(), now respects component boundaries.
• Runner
  • New tasks for link prediction and node classification/regression based on structured readout.
  • Now comes with API docs.
• Models collection
  • models/contrastive_losses gets multiple extensions, including a triplet loss and API docs.
  • models/multi_head_attention replaces sigmoid with elu+1 in trained scaling.
  • Bug fixes for mixed precision.

Full Changelog: https://github.com/tensorflow/gnn/compare/v0.6.1...v1.0.0

Initial release candidate for v1.0.0.

Early developmental release of tensorflow-gnn 1.0.0 code; docs still unfinished.

• import tensorflow_gnn now checks if the version of tf.keras is compatible.

Early developmental release of tensorflow-gnn 1.0.0 code; docs still outdated.

What's Changed

• Typo fix is_writen -> is_written by @copybara-service in https://github.com/tensorflow/gnn/pull/304
• For examples/notebooks, install the latest release (not release candidate). by @copybara-service in https://github.com/tensorflow/gnn/pull/306
• Add OGBN-MAG colab links by @copybara-service in https://github.com/tensorflow/gnn/pull/305
• Partially re-enable some tests for TF2.9. by @copybara-service in https://github.com/tensorflow/gnn/pull/316
• Expose integrated gradient function to enable building custom exporter with it. by @copybara-service in https://github.com/tensorflow/gnn/pull/319
• Add causal masking when training with labels. by @copybara-service in https://github.com/tensorflow/gnn/pull/321
• Fixes double-install of tf-nightly and tensorflow in nightly test script. by @copybara-service in https://github.com/tensorflow/gnn/pull/323
• GCN model: Different normalizations of the adjacency matrix is added. The value of 'degree_normalization' controls how the edges are normalized. The default value of 'degree_normalization' is set to 'in_out'. Note that the default value gives a different output for asymmetric graphs compared to the previous default value. To reinstate the old behaviour of the code for asymmetric graphs, set 'degree_normalization' to 'in_in'. The 'normalize' flag is removed. To make sure the edges are not normalized, set the 'degree_normalization' to 'none'. by @copybara-service in https://github.com/tensorflow/gnn/pull/326
• For eventual multi-task support: by @copybara-service in https://github.com/tensorflow/gnn/pull/335
• This change is a no-op by @copybara-service in https://github.com/tensorflow/gnn/pull/262
• Raise exception if a key has been provided, but no dictionary is available for ResidualNextState. by @copybara-service in https://github.com/tensorflow/gnn/pull/353
• Hooks for HGT hyperparameter tuning by @copybara-service in https://github.com/tensorflow/gnn/pull/356
• Update TF-GNN models to best practices: by @copybara-service in https://github.com/tensorflow/gnn/pull/358
• Remove the unnecessary epsilon parameter. by @copybara-service in https://github.com/tensorflow/gnn/pull/363
• Call validate_graph_tensor_spec_for_readout() from validate_schema(). by @copybara-service in https://github.com/tensorflow/gnn/pull/364
• Remove the function random_graph_tensor_from_schema(). by @copybara-service in https://github.com/tensorflow/gnn/pull/367
• Fix typo: s/contined/continued/ by @copybara-service in https://github.com/tensorflow/gnn/pull/379
• Merge DGI layers with others. by @copybara-service in https://github.com/tensorflow/gnn/pull/385
• Add code for upcoming generalized pool() and broadcast() operations. by @copybara-service in https://github.com/tensorflow/gnn/pull/384
• Add more granular shuffling. by @copybara-service in https://github.com/tensorflow/gnn/pull/387
• Unit-test broadcast_node_to_edges() for Adjacency and HyperAdjacency. by @copybara-service in https://github.com/tensorflow/gnn/pull/394
• Expose RunResult. by @copybara-service in https://github.com/tensorflow/gnn/pull/395
• Merge DGI layers with others. by @copybara-service in https://github.com/tensorflow/gnn/pull/402
• Updates TF version for newest stable tests to 2.12 by @copybara-service in https://github.com/tensorflow/gnn/pull/410
• Allow passing collection of receiver tags to ConvGNNBuilder by @copybara-service in https://github.com/tensorflow/gnn/pull/412
• Fix the behavior of shuffle_tensor for 0 shapes. This happens with make_empty_feature and shuffling rate not in {0, 1}. by @copybara-service in https://github.com/tensorflow/gnn/pull/416
• Correctly handle auxiliary graph pieces in convert_to_line_graph by @copybara-service in https://github.com/tensorflow/gnn/pull/422
• TFLite tests for vanilla_mpnn and mt_albis by @copybara-service in https://github.com/tensorflow/gnn/pull/425
• Allows specialization how composite tensors are converted for the bulk executor. by @copybara-service in https://github.com/tensorflow/gnn/pull/432
• Silence some pytype errors. by @copybara-service in https://github.com/tensorflow/gnn/pull/439
• Update attribution.py testing: the preprocessing model created by orchestration returns a Sequence[GraphTensor]. by @copybara-service in https://github.com/tensorflow/gnn/pull/442
• Add Keras wrappers ReadoutNamedIntoFeature and AddReadoutFromFirstNode by @copybara-service in https://github.com/tensorflow/gnn/pull/452
• Adjust maxval += 1 in the case of an integer random counterfactual. by @copybara-service in https://github.com/tensorflow/gnn/pull/457
• Switch tfgnn.pool() to implementation v2 and remove v1. by @copybara-service in https://github.com/tensorflow/gnn/pull/464
• Enable test case: subgraph sampling model, saving and loading, as tf.keras.Model. by @copybara-service in https://github.com/tensorflow/gnn/pull/468
• Utils to read unigraph-formatted data into sampler v2 input format and a driver program used like graph_sampler.py by @copybara-service in https://github.com/tensorflow/gnn/pull/469
• Align context indices dtype resolution between GraphTensor and GraphTensorSpec by @copybara-service in https://github.com/tensorflow/gnn/pull/477
• Redefines indices_dtype as two concepts: by @copybara-service in https://github.com/tensorflow/gnn/pull/486
• Add convenience function. by @copybara-service in https://github.com/tensorflow/gnn/pull/495
• Fix type annotation: SimpleConv(..., receiver_feature=None) is supported by @copybara-service in https://github.com/tensorflow/gnn/pull/506
• Fixes type annotation syntax in runner link prediction by @copybara-service in https://github.com/tensorflow/gnn/pull/511
• Disable test that fails for unrelated reasons. by @copybara-service in https://github.com/tensorflow/gnn/pull/518
• Sampler pieces are removed from tfgnn core package. by @copybara-service in https://github.com/tensorflow/gnn/pull/519
• Re-enables edge sampling tests. by @copybara-service in https://github.com/tensorflow/gnn/pull/529
• Implement the Coherence metric. by @copybara-service in https://github.com/tensorflow/gnn/pull/533
• Fix for inconsistent ragged row partitions of edges and indices to select unique edges. by @copybara-service in https://github.com/tensorflow/gnn/pull/535

Full Changelog: https://github.com/tensorflow/gnn/compare/v0.5.1...v0.6.0

Full Changelog: https://github.com/tensorflow/gnn/compare/v0.6.0rc0...v0.6.0rc1

What's Changed

• Typo fix is_writen -> is_written by @copybara-service in https://github.com/tensorflow/gnn/pull/304
• For examples/notebooks, install the latest release (not release candidate). by @copybara-service in https://github.com/tensorflow/gnn/pull/306
• Add OGBN-MAG colab links by @copybara-service in https://github.com/tensorflow/gnn/pull/305
• Partially re-enable some tests for TF2.9. by @copybara-service in https://github.com/tensorflow/gnn/pull/316
• Expose integrated gradient function to enable building custom exporter with it. by @copybara-service in https://github.com/tensorflow/gnn/pull/319
• Add causal masking when training with labels. by @copybara-service in https://github.com/tensorflow/gnn/pull/321
• Fixes double-install of tf-nightly and tensorflow in nightly test script. by @copybara-service in https://github.com/tensorflow/gnn/pull/323
• GCN model: Different normalizations of the adjacency matrix is added. The value of 'degree_normalization' controls how the edges are normalized. The default value of 'degree_normalization' is set to 'in_out'. Note that the default value gives a different output for asymmetric graphs compared to the previous default value. To reinstate the old behaviour of the code for asymmetric graphs, set 'degree_normalization' to 'in_in'. The 'normalize' flag is removed. To make sure the edges are not normalized, set the 'degree_normalization' to 'none'. by @copybara-service in https://github.com/tensorflow/gnn/pull/326
• For eventual multi-task support: by @copybara-service in https://github.com/tensorflow/gnn/pull/335
• This change is a no-op by @copybara-service in https://github.com/tensorflow/gnn/pull/262
• Raise exception if a key has been provided, but no dictionary is available for ResidualNextState. by @copybara-service in https://github.com/tensorflow/gnn/pull/353
• Hooks for HGT hyperparameter tuning by @copybara-service in https://github.com/tensorflow/gnn/pull/356
• Update TF-GNN models to best practices: by @copybara-service in https://github.com/tensorflow/gnn/pull/358
• Remove the unnecessary epsilon parameter. by @copybara-service in https://github.com/tensorflow/gnn/pull/363
• Call validate_graph_tensor_spec_for_readout() from validate_schema(). by @copybara-service in https://github.com/tensorflow/gnn/pull/364
• Remove the function random_graph_tensor_from_schema(). by @copybara-service in https://github.com/tensorflow/gnn/pull/367
• Fix typo: s/contined/continued/ by @copybara-service in https://github.com/tensorflow/gnn/pull/379
• Merge DGI layers with others. by @copybara-service in https://github.com/tensorflow/gnn/pull/385
• Add code for upcoming generalized pool() and broadcast() operations. by @copybara-service in https://github.com/tensorflow/gnn/pull/384
• Add more granular shuffling. by @copybara-service in https://github.com/tensorflow/gnn/pull/387
• Unit-test broadcast_node_to_edges() for Adjacency and HyperAdjacency. by @copybara-service in https://github.com/tensorflow/gnn/pull/394
• Expose RunResult. by @copybara-service in https://github.com/tensorflow/gnn/pull/395
• Merge DGI layers with others. by @copybara-service in https://github.com/tensorflow/gnn/pull/402
• Updates TF version for newest stable tests to 2.12 by @copybara-service in https://github.com/tensorflow/gnn/pull/410
• Allow passing collection of receiver tags to ConvGNNBuilder by @copybara-service in https://github.com/tensorflow/gnn/pull/412
• Fix the behavior of shuffle_tensor for 0 shapes. This happens with make_empty_feature and shuffling rate not in {0, 1}. by @copybara-service in https://github.com/tensorflow/gnn/pull/416
• Correctly handle auxiliary graph pieces in convert_to_line_graph by @copybara-service in https://github.com/tensorflow/gnn/pull/422
• TFLite tests for vanilla_mpnn and mt_albis by @copybara-service in https://github.com/tensorflow/gnn/pull/425
• Allows specialization how composite tensors are converted for the bulk executor. by @copybara-service in https://github.com/tensorflow/gnn/pull/432
• Silence some pytype errors. by @copybara-service in https://github.com/tensorflow/gnn/pull/439
• Update attribution.py testing: the preprocessing model created by orchestration returns a Sequence[GraphTensor]. by @copybara-service in https://github.com/tensorflow/gnn/pull/442
• Add Keras wrappers ReadoutNamedIntoFeature and AddReadoutFromFirstNode by @copybara-service in https://github.com/tensorflow/gnn/pull/452
• Adjust maxval += 1 in the case of an integer random counterfactual. by @copybara-service in https://github.com/tensorflow/gnn/pull/457
• Switch tfgnn.pool() to implementation v2 and remove v1. by @copybara-service in https://github.com/tensorflow/gnn/pull/464
• Enable test case: subgraph sampling model, saving and loading, as tf.keras.Model. by @copybara-service in https://github.com/tensorflow/gnn/pull/468
• Utils to read unigraph-formatted data into sampler v2 input format and a driver program used like graph_sampler.py by @copybara-service in https://github.com/tensorflow/gnn/pull/469
• Align context indices dtype resolution between GraphTensor and GraphTensorSpec by @copybara-service in https://github.com/tensorflow/gnn/pull/477
• Redefines indices_dtype as two concepts: by @copybara-service in https://github.com/tensorflow/gnn/pull/486
• Add convenience function. by @copybara-service in https://github.com/tensorflow/gnn/pull/495
• Fix type annotation: SimpleConv(..., receiver_feature=None) is supported by @copybara-service in https://github.com/tensorflow/gnn/pull/506
• Fixes type annotation syntax in runner link prediction by @copybara-service in https://github.com/tensorflow/gnn/pull/511
• Disable test that fails for unrelated reasons. by @copybara-service in https://github.com/tensorflow/gnn/pull/518
• Sampler pieces are removed from tfgnn core package. by @copybara-service in https://github.com/tensorflow/gnn/pull/519
• Re-enables edge sampling tests. by @copybara-service in https://github.com/tensorflow/gnn/pull/529
• Implement the Coherence metric. by @copybara-service in https://github.com/tensorflow/gnn/pull/533
• Fix for inconsistent ragged row partitions of edges and indices to select unique edges. by @copybara-service in https://github.com/tensorflow/gnn/pull/535

Full Changelog: https://github.com/tensorflow/gnn/compare/v0.5.1...v0.6.0rc0

Full Changelog: https://github.com/tensorflow/gnn/compare/v0.5.0...v0.5.1

Full Changelog: https://github.com/tensorflow/gnn/compare/v0.5.0rc0...v0.5.0

What's Changed

• Code Tutorial #1: Running TF-GNN on single machine. by @copybara-service in https://github.com/tensorflow/gnn/pull/215
• Minor tutorial site changes. by @copybara-service in https://github.com/tensorflow/gnn/pull/218
• Add link to slides by @copybara-service in https://github.com/tensorflow/gnn/pull/220
• Add GPT-GNN folder for the new autoregressive edge-prediction Task. by @copybara-service in https://github.com/tensorflow/gnn/pull/232
• Add compatibility tests for oldest supported Python and TensorFlow version (in this case Python 3.7 and TensorFlow 2.8) by @copybara-service in https://github.com/tensorflow/gnn/pull/234
• Add MultiHeadAttention model option to xm_launcher by @copybara-service in https://github.com/tensorflow/gnn/pull/235
• Satisfy pytype by having kwarg Convolve(extra_receiver_ops=...) where unused. by @copybara-service in https://github.com/tensorflow/gnn/pull/237
• Fix embedded paths to point to LoG location instead. by @copybara-service in https://github.com/tensorflow/gnn/pull/240
• Small colab fixes. by @copybara-service in https://github.com/tensorflow/gnn/pull/242
• Fix link and copy as_dataset implementation that is not yet in pip package. by @copybara-service in https://github.com/tensorflow/gnn/pull/246
• Use Keras layers (vs. tensor functions) for pooling. by @copybara-service in https://github.com/tensorflow/gnn/pull/247
• Move interfaces from orchestration.py to interfaces.py. by @copybara-service in https://github.com/tensorflow/gnn/pull/248
• Exhaustively test Tasks cross distribution strategy. by @copybara-service in https://github.com/tensorflow/gnn/pull/250
• Add save_options to ModelExporters by @copybara-service in https://github.com/tensorflow/gnn/pull/254
• fixes tag filters in test script by @copybara-service in https://github.com/tensorflow/gnn/pull/256
• Fix overriding return type. by @copybara-service in https://github.com/tensorflow/gnn/pull/265
• Apply DGI Task random seed for feature shuffling operations. Also add the missing tests for distributed strategy to ensure model correctness is deterministically validated. by @copybara-service in https://github.com/tensorflow/gnn/pull/182
• Expose check_compatible_with_schema_pb to the public interface. by @copybara-service in https://github.com/tensorflow/gnn/pull/271
• Enabling GCNConv layer to work with weighted adjacency. by @copybara-service in https://github.com/tensorflow/gnn/pull/274
• Add a tensorflow_gnn/models/constrastive_losses/... with a Deep Graph Infomax implementation. by @copybara-service in https://github.com/tensorflow/gnn/pull/285
• Negative sampling operation that returns a tensor of unconnected nodes for each node-id in the adjacency list. by @copybara-service in https://github.com/tensorflow/gnn/pull/158

Full Changelog: https://github.com/tensorflow/gnn/compare/v0.4.0...v0.5.0rc0

Full Changelog: https://github.com/tensorflow/gnn/compare/v0.4.0...v0.4.1

Full Changelog: https://github.com/tensorflow/gnn/compare/v0.4.0rc2...v0.4.0

Bugfix for python 3.7 compatibility

Full Changelog: https://github.com/tensorflow/gnn/compare/v0.4.0rc1...v0.4.0rc2

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

Full Changelog: https://github.com/tensorflow/gnn/compare/v0.4.0rc0...v0.4.0rc1