Release 0.80.0

Breaking Changes

• Move the tools package to the root of the repository.
• Updated bazel to version 6.5.0.
• Updated rules_python to version 0.31.0.
• Deleted deprecated tff.learning.build_federated_evaluation, which was superseded by tff.learning.algorithms.build_fed_eval.

Release 0.79.0

• Enable support for models with non-trainable variables in
  tff.learning.models.functional_model_from_keras.

Breaking Changes

• Updated com_github_grpc_grpc to version 1.50.0.
• Moved the TFF repository from https://github.com/tensorflow/federated to
  https://github.com/google-parfait/tensorflow-federated.

Major Features and Improvements

• Moved aggregation from https://github.com/google-parfait/federated-compute
  to TFF to consolidate the federated language and remove circular
  dependencies.

Breaking Changes

• Updated rules_license to version 0.0.8.
• Removed elias_gamma_encode module.
• Removed tensorflow_compression dependency.

Release 0.77.0

Major Features and Improvements

• Added an implementation of __eq__() on building blocks.

Bug Fixes

• Fix #4588: Target Haswell CPU architectures (-march=haswell) instead of
  whatever is native to the build infrastructure to ensure that binaries in
  the pip package and executable on Colab CPU runtimes.

Release 0.76.0

Major Features and Improvements

• Added a Literal to the TFF language, part 2. This change updates the
  tracing and execution portions of TFF to begin using the Literal.
• Added an implementation of the Adafactor optimizer to
  tff.learning.optimizers.build_adafactor
• Added a new field, content, to the Data proto.

Breaking Changes

• Removed the check_foo() methods on building blocks.
• Removed tff.data, this symbol is not used.

Bug Fixes

• Fix a bug where the pip package default executor stack cannot execute
  computations that have Lambdas under sequence_* intrinsics.

Release 0.75.0

Major Features and Improvements

• Updated the type annotation for MaterializedValue to include the Python
  scalar types in addition to the numpy scalar types.
• Added a Literal to the TFF language, part 1.
• Added Literal to the framework package.
• Extended
  tff.learning.algorithms.build_weighted_fed_avg_with_optimizer_schedule to
  support tff.learning.models.FunctionalModel.

Breaking Changes

• Deleted the tff.learning.framework namespace⚰️.

Bug Fixes

• Fixed logic for determining if a value can be cast to a specific dtype.
• Fixed a bug where repeated calls to
  FilePerUserClientData.create_tf_dataset_for_client could blow up memory
  usage

Release 0.74.0

Major Features and Improvements

• Make some of the C++ executor APIs public visibility for downstream repos.
• Moved the DataType protobuf object into its own module. Moving the
  DataType object into its own module allows DataType to be used outside
  of a Computation more easily and prevents a circular dependency between
  Computation and Array which both require a DataType.
• Updated build_apply_optimizer_finalizer to allow custom reject update
  function.
• Relaxed the type requirement of the attributes of ModelWeights to allow
  assigning list or tuples of matching values to other sequence types on
  tf.keras.Model instances.
• Improved the errors raised by JAX computations for various types.
• Updated tutorials to use recommended tff.learning APIs.

Breaking Changes

• Removed the runtime-agnostic support for tf.RaggedTensor and
  tf.SparseTensor.

Full Changelog: https://github.com/tensorflow/federated/compare/v0.73.0...v0.74.0

Release 0.73.0

Major Features and Improvements

• Make some of the C++ executor APIs public visibility for downstream repos.
• tff.learning.algorithms.build_fed_kmeans supports floating point weights,
  enabling compatibility with tff.aggregators using differential privacy.
• Added two new metrics aggregators:
  tff.learning.metrics.finalize_then_sample and
  tff.learning.metrics.FinalizeThenSampleFactory.

Breaking Changes

• Remove the ability to return SequenceType from tff.federated_computation
  decorated callables.

Bug Fixes

• tff.learning algorithms now correctly do not include metrics for clients
  that had zero weight due to model updates containing non-finite values.
  Previously the update was rejected, but the metrics still aggregated.

Release 0.72.0

Major Features and Improvements

• Added an async XLA runtime under tff.backends.xla.

Breaking Changes

• Updated tensorflow-privacy version to 0.9.0.
• Removed the deprecated type_signature parameter from the tff.program.ReleaseManager.release method.

Release 0.71.0

Major Features and Improvements

• Added new environment-specific packages to TFF.

Breaking Changes

• Temporarily disable tff.program.PrefetchingDataSource due to flakiness
  from a lack of determinism.
• Removed support for invoking infer_type with TensorFlow values.
• Removed deprecated tff.aggregators.federated_(min|max)symbols, please use
  tff.federated_(min|max) instead.
• Removed support for creating a tff.TensorType using a tf.dtypes.DType.
• Removed tff.check_return_type.

Bug Fixes

• Declared OwnedValueId::INVALID_ID as a static constexpr.

Release 0.69.0

Major Features and Improvements

• The local_unfinalized_metrics_type argument to
  tff.learning.metrics.(secure_)sum_then_finalize is now optional (and is not
  actually used). It will be removed in a future release.

Breaking Changes

• tff.learning.metrics.(secure_)sum_then_finalize now return polymorphic
  computations. They can still be passed into algorithm builders (e.g.
  tff.learning.algorithms.build_weighted_fed_avg) but to be called directly
  they must first be traced with explicit types.
• Removed support for handling tf.TensorSpec using to_type, use
  tensorflow_to_type instead.
• Removed support for calling tff.TensorType using a tf.dtypes.DType.

Release 0.68.0

Major Features and Improvements

• Added tff.types.tensorflow_to_type function to convert structures
  containing tensorflow type specs into a tff.Type.
• Deprecated tff.types.infer_unplaced_type.
• Updated tff.types.ArrayShape to be defined as a Sequence not an
  Iterable, this is because the len of an tff.types.ArrayShape is used
  for comparison.
• Deprecated the type_signature parameter for the
  tff.program.ReleaseManager.release method.

Breaking Changes

• Removed the implementation of tff.Value.__add__.
• Removed the deprecated tff.Type.check_*() functions, use isinstance
  instead.
• Removed tff.types.at_clients and tff.types.at_server functions, use the
  tff.FederatedType constructor instead.
• Removed support for handling tf.data.DatasetSpec, tf.RaggedTensorSpec,
  and tf.SparseTensorSpec using tff.to_type, use
  tff.types.tensorflow_to_type instead.
• Removed support for handling tf.RaggedTensor and tf.SparseTensor using
  infer_type.

Release 0.67.0

Major Features and Improvements

• Updated the representation of a tff.TensorType.dtype to be a np.dtype
  instead of tf.dtypes.Dtype.
• Added tff.program.DelayedReleaseManager.

Breaking Changes

• Removed check_allowed_ops from the framework package.
• Removed check_disallowed_ops from the framework package.
• Removed replace_intrinsics_with_bodies from the framework package.
• Removed get_session_token from the framework package.
• Added a workspace dependency on pybind11_bazel.
• Removed type_from_tensors from the framework package.
• Updated the version of rules_python to 0.23.0.

Bug Fixes

• Temporary pin googleapis-common-protos to version 1.61.0 to work around
  an issue with a transitive dependency.

Release 0.66.0

Breaking Changes

• Removed the capability to pass a tf.TensorShape as the shape of a
  tff.TensorType.

Bug Fixes

• Correctly materialize SERVER placed values out of the C++ execution stack
  when using StreamingRemoteExecutor instead of returning an error about
  placement not found.

Release 0.65.0

Major Features and Improvements

• Update the representation of a tff.TensorType.shape to be a
  tff.types.ArrayShape instead of tf.TensorShape.

• Updated type_to_py_container``to be able to handletff.SequenceTypes` with
  an unknown Python type.

Breaking Changes

• Moved tff.structure_from_tensor_type_tree to
  tff.types.structure_from_tensor_type_tree.
• Remove the capability to pass an int as the shape of a tff.TensorType.

Release 0.64.0

Major Features and Improvements

• Updated the TFF project and the Python package to be compatible with Python
  3.11.
• Updated train_process to train_process_factory in vizier program logic
  to support multiple trials in parallel.
• Added support for using non-OrderedDict mapping types.

Breaking Changes

• Updated the version of grpc to v1.59.1.
• Updated the version of bazel to 6.1.0.
• Updated the version of tensorflow to 2.14.0.
• Updated the version of numpy to ~=1.25.
• Updated the version of com_google_googletest to 1.12.1.

Bug Fixes

• Fixed import path for Vizier in federated program example.
• Fixed serialization of TenshorShape in error message to be human readable.
• Fixed bug in tff.program.FileProgramStateManager removing old program
  state.

Release 0.63.0

Major Features and Improvements

• Added tff.federated_min and tff.federated_max intrinsics.
• Added a get_value() method to tff.program.SavedModelFileReleaseManager, for retrieving values that were previously released.
• Added tff.program.PeriodicReleaseManager to release values at regular intervals.
• Learning program logic now saves next evaluation time so that it can be loaded upon computation restarts.
• DistributeAggregateForm now skips normalizing the all_equal bit.
• Added parallelism to Vizier program logic.
• Enabled building protos with certain Bazel versions.

Breaking Changes

• Updated the version of attrs to 23.1.
• Updated the version of cachetools to ~=5.3.
• Updated the version of dp-accounting to 0.4.3.
• Updated the version of google-vizier to 0.1.11.
• Updated the version of jax to 0.4.14.
• Updated the version of portpicker to ~=1.6.
• Updated the version of tensorflow to 2.13.0.
• Updated the version of tensorflow-model-optimization to 0.7.5.
• Updated the version of tensorflow-privacy to 0.8.11.
• Updated the version of typing-extensions to ~=4.5.0.
• Increased TF_CUDA_VERSION to 12.
• Removed the tff.program.Capabilities enum from the iterator.
• Deleted Python executors.
• Removed the deprecated is_{foo} functions from tff.Types. Users should use isinstance checks instead.
• Deleted go-related BUILD targets for TFF proto files.

Bug Fixes

• Removed non-existent GCP doc from TFF guides.
• Cleaned up exceptions in the tff.program package for consistency and clarity.
• Fixed various pytype errors.
• Fixed various undefined-variable lint errors.
• Fixed a UnicodeDecodeError in the FedRecon tutorial.
• Fixed Python 3.11 related errors.

Release 0.62.0

Breaking Changes

• Removed context argument from
  tff.learning.algorithms.build_personalization_eval_computation. Now a
  personalization function only takes a model, a train dataset, and a test
  dataset as arguments.

Bug Fixes

• Fix a rare infinite loop issue caused by very small float values when using
  tff.learning.ddp_secure_aggregator.

Release 0.61.0

Major Features and Improvements

• Updated the type annotation for the dtype parameter to tff.TensorType.
• Added adaptive tuning function to ThresholdSampling class.
• Added
  tff.learning.models.ReconstructionModel.from_keras_model_and_variables,
  which provides a way to get a ReconstructionModel from a Keras model and
  lists of global and local trainable/non_trainable variables.
• Switched tff.learning.algorithms.build_fed_recon_eval to use a stateful
  metrics aggregator.

Breaking Changes

• Removed tff.learning.models.ReconstructionModel.from_keras_model, which
  has been replaced by
  tff.learning.models.ReconstructionModel.from_keras_model_and_layers.
• Removed the following functions from the py_typecheck module: check_len,
  check_callable, is_dataclass,is_attrs, check_subclass and
  check_not_none. They are unused or can be replaced by Python type
  annotations.

Bug Fixes

• Fixed a small bug in the TFF SGDM optimizer, to only track momentum as a
  hparam if it is both specified and nonzero.