Release 0.60.0

Major Features and Improvements

• DTensor TF executor is now integrated with the default TFF C++ worker.
• Added federated program documentation and guidelines.
• Removed the pytype dependency from TFF.
• tff.learning.algorithms.build_fed_recon_eval now supports TFF optimizers.

Breaking Changes

• Updated tff.types.deserialize_type to not accept/return None.
• Removed the tff.framework.ComputationBuildingBlock.is_foo methods.
• Renamed tff.learning.algorithms.build_personalization_eval to
  tff.learning.algorithms.build_personalization_eval_computation
• tff.learning.models.ReconstructionModel.from_keras_model will now check
  that global and local variables are disjoint, raise ValueError if they are
  not.

Bug Fixes

• Fixed tff.learning.models.ReconstructionModel.has_only_global_variables
  (it was returning incorrect value).

Release 0.59.0

Major Features and Improvements

• Removed compression for worker_binary.
• Allowed tensor and numpy float-like objects in optimizer hyperparameters.
• Improved API/filtering logic in FilteringReleaseManager.

Breaking Changes

• Renamed build_personalization_eval to
  build_personalization_eval_computation.
• Updated tff.to_type implementation and type annotation to not
  accept/return None.

Bug Fixes

• Fixed and documented pytype errors in the program package.
• Fixed bug in how tff.program.NativeFederatedContext handles arguments of
  various types.

Release 0.58.0

Major Features and Improvements

• Updated algorithms built from tff.learning.models.FunctionalModel to allow
  nested outputs.
• Added the PrefetchingDataSource back to the tff.program API now that the
  flakiness has been fixed.

Bug Fixes

• Changed return type of
  tff.simulation.compose_dataset_computation_with_learning_process to be a
  tff.learning.templates.LearningProcess.
• Fixed flaky tests in prefetching_data_source_test.
• Fixed type annotations and lint errors.
• Cleaned up error messages and typing information in
  tff.learning.optimizers.

Release 0.57.0

Major Features and Improvements

• Allow functional models to return a structure.

Breaking Changes

• Removed support for handling attrs as containers in the tff.program API.
• Migrated the personalization_eval module to the algorithms package.
• Deleted the tff.learning.build_local_evaluation API.
• Migrated tff.learning.reconstruction to the tff.learning.algorithms package.
• Updated to dm-tree version 0.1.8.
• Updated to dp-accounting version 0.4.1.
• Updated to tensorflow-privacy version 0.8.9.

Release 0.56.0

Major Features and Improvements

• Added Vizier backed tuning program logic to tff.learning.
• Updated the tff.learning.programs.EvaluationManager to clean up states
  after recording the evaluation is completed.

Breaking Changes

• Removed deprecated tff.learning.framework.ServerState symbol.

Release 0.55.0

Major Features and Improvements

• Removed nest_asyncio dependency from tutorials.
• Added a new
  aggregatorrtff.aggregators.DifferentiallyPrivateFactory.tree_adaptive for
  combining DP-FTRL (https://arxiv.org/abs/2103.00039) and adaptive clipping
  (https://arxiv.org/abs/1905.03871).
• Updated tff.learning.programs.EvaluationManager to set the evaluation
  deadline from the start time.

Breaking Changes

• Python runtime deleted; C++ runtime covers all known use-cases.

Bug Fixes

• Fixed a bug attempting to push tf.data.Dataset iterator ops onto GPUs.

Release 0.54.0

Major Features and Improvements

• Added attributes to tff.learning.programs.EvaluationManager, this enables
  constructing new EvaluationManagers from existing ones.
• Added Subsample Process abstract class and the implementation of Threshold
  Sampling Process Remove introducing dependency on relayout op for control
  edges.
• Replaced usage of attrs in tff.aggregators with typing.NamedTuple.
• Removed introducing dependency on relayout op for control edges.

Breaking Changes

• Removed run_server and server_context from the tff.simulation API.
• Removed the following symbols from the tff.framework API:
  • tff.framework.local_executor_factory
  • tff.framework.DataBackend
  • tff.framework.DataExecutor
  • tff.framework.EagerTFExecutor

Bug Fixes

• Removed use of deprecated tff.learning symbols, and clear cell image
  outputs.

Release 0.53.0

Major Features and Improvements

• Updated TF version to 2.12.0.
• Relaxed runtime type checks on tff.learning.templates.LearningProcess to allow non-sequence CLIENTS arguments.
• tff.simulation.compose_dataset_computation_with_learning_process now returns a tff.learning.templates.LearningProcess.
• Updated the tff.program.FederatedDataSourceIterators so that they can be serialized.

Breaking Changes

• Deleted the forward_pass attribute from the FunctionalModel interface.
• Removed from_keras_model, MetricsFinalizersType, BatchOutput, Model, and ModelWeights symbols from the tff.learning package. Users should instead use the tff.learning.models package for these symbols.
• Removed deprecated tff.learning.federated_aggregate_keras_metric function.
• Removed implicit attribute forwarding on tff.simulation.compose_dataset_computation_with_learning_process.
• Removed deprecated tff.framework.remote_executor_factory_from_stubs.
• Removed deprecated tff.backends.xla APIs.
• Renamed the tff.backends.test APIs to: tff.backends.test.(create|set)_(sync|async)_test_cpp_execution_context.

Release 0.52.0

Major Features and Improvements

• Exposed tff.backends.mapreduce.consolidate_and_extract_local_processing as
  public API.
• Updated the federated program API to be able to handle tff.Serializable
  objects.

Breaking Changes

• Deprecated handling attrs classes as containers in the tff.program API.
• Updated tff.learning.algorithms implementations to use
  tff.learning.models.FunctionalModel.loss instead of
  FunctionalModel.forward_pass.

Bug Fixes

• Avoid using sys.stdout and sys.stderr in subprocess.Popen when
  executing inside an IPython context.
• Added a SequenceExecutor to the C++ execution stack to handle sequence_*
  intrinsics.

Release 0.51.0

Major Features and Improvements

• Enabled, improved, and fixed Python type annotations in some modules.
• Added the interface of loss_fn to tff.learning.models.FunctionalModel,
  along with serialization and deserialization methods.
• Updated the composing executor to forward the type field of Intrinsic
  protos that are sent to child executors.
• Added an executor binding for DTensor based executor.

Breaking Changes

• Deprecated tff.framework.DataBackend. Python execution is deprecated
  instead use CPP Execution.

Bug Fixes

• Fixed the formulation of the JIT constructed mapped selection computation
  that is sent to the remote machine in streaming mode.
• Fixed the usage of np.bytes_ types that incorrectly truncate byte string
  with null terminator.

Release 0.50.0

Major Features and Improvements

• Added client learning rate measurements to
  tff.learning.algorithms.build_weighted_fed_avg_with_optimizer_schedule
• Added support for streaming federated structure values to the C++
  RemoteExecutor.
• Added a C++ executor for executing TF graphs using TF2 DTensor APIs when
  layout information is specified for input parameters or variables in the
  graph.

Breaking Changes

• Deprecated the following API, Python execution is deprecated instead use CPP
  execution:
  • tff.framework.local_executor_factory
  • tff.framework.remote_executor_factory_from_stubs
  • tff.framework.DataExecutor
  • tff.framework.EagerTFExecutor
• Removed the following API, Python execution is deprecated instead use CPP
  execution:
  • tff.backends.native.create_local_python_execution_context
  • `tff.backends.native.create_remote_python_execution_context
  • tff.framework.remote_executor_factory
• Remove the executors_errors module from the tff.framework API, use
  tff.framework.RetryableError instead.

Bug Fixes

• Fixed potential lifetime issue in C++ RemoteExecutor
• Enabled and fixed python type annotations in many packages.
• Fixed one-off error in evaluation criteria in training program logic.

Release 0.49.0

Major Features and Improvements

• Created the Baselines API of the GLDv2 (landmark) dataset for simulation,
  with a GLDv2 preprocessing function, a GLDv2 tasks function, and a Google
  mirror of the GLDv2 baselines tasks.

Breaking Changes

• Temporarily removed tff.program.PrefetchingDataSource, the
  PrefetchingDataSourceIterator tests are flaky and it's not immediately clear
  if this is due to the implementation of the PrefetchingDataSourceIterator or
  due to a bug in the test.
• Deprecated the following API, Python execution is deprecated instead use CPP
  execution:
  • tff.backends.native.create_local_python_execution_context
  • tff.backends.native.create_remote_python_execution_context
  • tff.backends.native.create_remote_async_python_execution_context
  • tff.backends.native.set_remote_async_python_execution_context
• Removed the following API, Python execution is deprecated instead use CPP
  execution:
  • tff.backends.native.set_local_python_execution_context
  • tff.backends.native.set_remote_python_execution_context
  • tff.frameowrk.FederatingExecutor
  • tff.framework.ComposingExecutorFactory
  • tff.framework.ExecutorValue
  • tff.framework.Executor
  • tff.framework.FederatedComposingStrategy
  • tff.framework.FederatedResolvingStrategy
  • tff.framework.FederatingStrategy
  • tff.framework.ReconstructOnChangeExecutorFactory
  • tff.framework.ReferenceResolvingExecutor
  • tff.framework.RemoteExecutor
  • tff.framework.ResourceManagingExecutorFactory
  • tff.framework.ThreadDelegatingExecutor
  • tff.framework.TransformingExecutor
  • tff.framework.UnplacedExecutorFactory
• Removed duplicate API from tff.framework, instead use:
  • tff.types.type_from_tensors
  • tff.types.type_to_tf_tensor_specs
  • tff.types.deserialize_type
  • tff.types.serialize_type
• Renamed tff.learning.Model to tff.learning.models.VariableModel.
• Renamed the
  cpp_execution_context.(create|set)_local_async_cpp_execution_context
  function to match the name of
  execution_context.(create|set)_(sync|async)_local_cpp_execution_context.

Bug Fixes

• Fixed bug in FLAIR download URLs.
• Enabled and fixed python type annotations in many packages.

Release 0.48.0

Major Features and Improvements

• Implemented divisive split logic needed by DistributeAggregateForm, which is
  currently under development and will replace MapReduceForm and BroadcastForm
  in the future.

Breaking Changes

• Renamed the cpp_execution_context.(create|set)_local_cpp_execution_context
  function to match the name of
  execution_context.(create|set)_(sync|async)_local_cpp_execution_context.
• Deleted the sizing Python execution context and executor.
• Deleted the thread debugging Python execution context and executor.
• Removed ExecutorService from the public API.
• Deleted the local async python execution context.

Bug Fixes

• Enabled and fixed python type annotations in some modules in the
  executors, types, and core package.

Release 0.47.0

Major Features and Improvements

• Added a LayoutMap message in the computation proto for TensorFlow
  DTensor based execution.

Breaking Changes

• Removed the compiler_fn parameter from the high level
  *_mergeable_execution_context functions.

Bug Fixes

• Aligned the context types allowed by the
  tff.program.NativeFederatedContext and the
  tff.program.PrefetchingDataSource.
• Updated build_functional_model_delta_update to use ReduceDataset ops to
  rely on MLIR Bridge for XLA compilation and TPU usage.

Release 0.46.0

Major Features and Improvements

• Added parameter and implementation for C++ remote executor to stream the
  values in a structure across the gRPC interface.
• Added tff.backends.native.desugar_and_transform_to_native to the public
  API.
• Replaced GroupNorm implementation with implementation from Keras.
• Added tff.simulations.datasets.flair APIs for the FLAIR dataset.

Breaking Changes

• Removed file extension for model_output_manager used in
  tff.learning.programs

Bug Fixes

• Enabled and fixed python type annotations in some modules.
• Changed tff.learning.algorithms.build_weighted_fed_prox parameter
  validation to allow proximal_strength = 0.0, matching the pydoc.

Release 0.45.0

Major Features and Improvements

• Integrated the CppToPythonExecutorBridge into the CPPExecutorFactory.
• Changed Python Remote Executor to decompose and stream structures in Compute
  and CreateValue when _stream_structs is true. Added a bool parameter
  stream_structs to
  tff.backends.native.set_localhost_cpp_execution_context() API.

Breaking Changes

• Renamed tff.backends.native.set_localhost_cpp_execution_context() to
  backends.native.set_sync_local_cpp_execution_context().
• Renamed tff.framework.ExecutionContext to
  tff.framework.SyncExecutionContext to be consistent with
  tff.framework.AsyncExecutionContext.
• Removed the SyncSerializeAndExecuteCPPContext and
  AsyncSerializeAndExecuteCPPContext classes.

Bug Fixes

• Fixed usage of typing.Generic in the learning package.
• Enabled pytype analysis for some modules.
• Fixed lint and formatting issues for some modules.

Release 0.44.0

Major Features and Improvements

• Improved the Python type annotations for tff.program API.
• Extended the metrics interface on FunctionalModel to accept the entire BatchOutput structure from the model forward_pass (not just the predictions).
• Introduced a DTensor Executor.

Bug Fixes

• Fixed async RuntimeWarning in the tff.program.NativeFederatedContext.

Release 0.43.0

Major Features and Improvements

• Improve serialization method to allow structures larger than 2 GiB (~500
  million model parameters):
  • tff.learning.models.FunctionalModel
  • tff.programs.FileProgramStateManager

Bug Fixes

• Fix a bug using copy.deepcopy for structures of awaitables (non-pickable)
  in tff.learning.programs.
• Fix a bug when resuming an evaluation in
  tff.learning.programs.EvaluationManager where the restarted evaluation
  would overwrite released metrics.

Release 0.42.0

Major Features and Improvements

• Reduced memory usage for entropy compression.
• Updated com_google_protobuf version to v3.19.0.
• Removed dependency on six.

Breaking Changes

• Removed default value for the key parameter from the abstract base class
  tff.program.ReleaseManager.

Bug Fixes

• Fixed a whitespace syntax issue with shutting down a process when using the
  localhost C++ execution context.
• Modified tff.simulation.build_uniform_sampling_fn so that the output
  raises on non-integer inputs.
• Only wait a subprocess instance if it is not None.

Release 0.41.0

Major Features and Improvements

• TFF-C++ runtime now installed by default. Note that this release will have a
  significantly larger PIP package size.
• Introduce tff.learning.programs for federated program-logic using the
  tff.program APIs.
• Updated tensorflow to version 2.11.0.
• Updated tensorflow_compression to version 2.11.0.
• Updated bazel_skylib to version 1.3.0.