RELEASE.md

Release Notes

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Unreleased

Added

• tff.StructType.items(), this API makes it easier to iterate over tff.StrucType without having to deal with hard to discover and use tff.structure.* APIs.
• The abstract class DPTensorAggregator and the child DPQuantileAggregator (along with the factory class). DPQuantileAggregator is currently a skeleton; future CLs will implement the member functions.
• DPQuantileAggregator::AggregateTensors performs either a push_back or reservoir sampling, depending on size of member buffer_. The reservoir sampling functionality is performed by ::InsertWithReservoirSampling.
• DPQuantileAggregator::MergeWith copies as much data over from the other aggregator's buffer_ until capacity is hit, then performs reservoir sampling.
• tff.program.ComputationArg, which is helpful when creating a federated platform.
• DPQuantileAggregator::ReportWithEpsilonAndDelta implements a DP algorithm to find quantiles by looping over a histogram with growing bucket size.
• DPQuantileAggregator::Serialize and the corresponding DPQuantileAggregatorFactory::Deserialze to save and load aggregator state.
• Ability to disable DP when epsilon is sufficiently large in DPQuantileAggregator::ReportWithEpsilonAndDelta.
• DPTensorAggregatorBundle, a wrapper around one or more instances of DPTensorAggregator, and its factory.
• DPTensorAggregatorBunde::AggregateTensors checks inputs before delegating work to the inner aggregators.
• A new constructor for InputTensorList that takes as input an std::vector of const Tensor*. DPTensorAggregatorBunde::AggregateTensors uses this to split its input across its inner aggregators, which may expect varying sizes.
• DPTensorAggregator::IsCompatible will allow DPTensorAggregatorBundle to check if all inner aggregators are compatible for merge prior to calling their MergeWith functions.
• DPTensorAggregatorBundle::MergeWith checks compatibility before delegating merging to inner aggregators. The compatibility check is done by DPTensorAggregatorBundle::IsCompatible.
• DPTensorAggregatorBundle::Serialize and DPTensorAggregatorBundleFactory::Deserialize enable storage and retrieval of the state of a DPTensorAggregatorBundle.
• DPTensorAggregatorBundle::TakeOutputs calls the inner aggregator's ReportWithEpsilonAndDelta methods and stitches the outputs together.
• Number of round retries to training metrics in tff.learning.programs.train_model.

Fixed

• Buffer overrun in AggVectorIterator when passing in an empty TensorData.

Changed

• DPTensorAggregatorBundleFactory::CreateInternal now checks validity of the epsilon and delta parameters of its given intrinsic.
• When DPGroupByFactory::CreateInternal receives an epsilon at or above kEpsilonThreshold, it no longer bothers splitting it across the inner aggregators.
• Moved the tests of input validity from DPQuantileAggregator to the parent class DPTensorAggregator. This will enable DPTensorAggregatorBundle to check that the input is valid before passing to the aggregators it contains.
• Moved the tests of compatibility from DPQuantileAggregator::MergeWith to DPQuantileAggregator::IsCompatible.
• Updated MeasuredProcessOutput to be a NamedTuple.

Removed

• tff.types.tensorflow_to_type, this function is no longer used.
• tff.program.X, use federated_language.program instead, for each:
  • FederatedDataSource
  • FederatedDataSourceIterator
  • check_in_federated_context
  • ComputationArg
  • contains_only_server_placed_data
  • FederatedContext
  • LoggingReleaseManager
  • MemoryReleaseManager
  • ProgramStateExistsError
  • ProgramStateManager
  • ProgramStateNotFoundError
  • ProgramStateStructure
  • ProgramStateValue
  • DelayedReleaseManager
  • FilteringReleaseManager
  • GroupingReleaseManager
  • NotFilterableError
  • PeriodicReleaseManager
  • ReleasableStructure
  • ReleasableValue
  • ReleaseManager
  • MaterializableStructure
  • MaterializableTypeSignature
  • MaterializableValue
  • MaterializableValueReference
  • materialize_value
  • MaterializedStructure
  • MaterializedValue

Release 0.88.0

Added

• tff.tensorflow.to_type.
• Added pack_args_into_struct and unpack_args_from_struct to the public API under framework.

Changed

• Add round end timestamp to train metrics in tff.learning.programs.train_model.

Deprecated

• tff.types.tensorflow_to_type, use tff.tensorflow.to_type instead.

Changed

• Updated to use an environment-agnostic way to represent a sequence of data.
• Updated JAX computations and contexts to be able to handle sequence types.
• Moved tff.types.structure_from_tensor_type_tree and tff.types.type_to_tf_tensor_specs to the tff.tensorflow package.

Removed

• tff.framework.merge_cardinalities
• tff.framework.CardinalityCarrying
• tff.framework.CardinalityFreeDataDescriptor
• tff.framework.CreateDataDescriptor
• tff.framework.DataDescriptor
• tff.framework.Ingestable

Release 0.87.0

Added

• An implementation of AdamW to tff.learning.optimizers.
• Added Executor class to public API.

Changed

• Support None gradients in tff.learning.optimizers. This mimics the behavior of tf.keras.optimizers - gradients that are None will be skipped, and their corresponding optimizer output (e.g. momentum and weights) will not be updated.
• The behavior of DPGroupingFederatedSum::Clamp: it now sets negatives to 0. Associated test code has been updated. Reason: sensitivity calculation for DP noise was calibrated for non-negative values.
• Change tutorials to use tff.learning.optimizers in conjunction with tff.learning computations.
• tff.simulation.datasets.TestClientData only accepts dictionaries whose leaf nodes are not tf.Tensors.

Fixed

• A bug where tff.learning.optimizers.build_adafactor would update its step counter twice upon every invocation of .next().
• A bug where tensor learning rates for tff.learning.optimizers.build_sgdm would fail with mixed dtype gradients.
• A bug where different optimizers had different behavior on empty weights structures. TFF optimizers now consistently accept and function as no-ops on empty weight structures.
• A bug where tff.simulation.datasets.TestClientData.dataset_computation yielded datasets of indeterminate shape.

Removed

• tff.jax_computation, use tff.jax.computation instead.
• tff.profiler, this API is not used.
• Removed various stale tutorials.
• Removed structure from tff.program.SavedModelFileReleaseManager's get_value method parameters.
• Removed support for tf.keras.optimizers in tff.learning.

Release 0.86.0

Added

• tff.tensorflow.transform_args and tff.tensorflow.transform_result, these functions are intended to be used when instantiating and execution context in a TensorFlow environment.

Changed

• Replaced the tensor on the Value protobuf with an array field and updated the serialization logic to use this new field.
• tff.program.FileProgramStateManager to be able to keep program states at a specified interval (every k states).

Release 0.85.0

Added

• The dp_noise_mechanisms header and source files: contains functions that generate differential_privacy::LaplaceMechanism or differential_privacy::GaussianMechanism, based upon privacy parameters and norm bounds. Each of these functions return a DPHistogramBundle struct, which contains the mechanism, the threshold needed for DP open-domain histograms, and a boolean indicating whether Laplace noise was used.
• Added some TFF executor classes to the public API (CPPExecutorFactory, ResourceManagingExecutorFactory, RemoteExecutor, RemoteExecutorGrpcStub).
• Added support for bfloat16 dtypes from the ml_dtypes package.

Fixed

• A bug where tf.string was mistakenly allowed as a dtype to tff.types.TensorType. This now must be np.str_.

Changed

• tff.Computation and tff.framework.ConcreteComputation to be able to transform the arguments to the computation and result of the computation.
• DPClosedDomainHistogram::Report and DPOpenDomainHistogram::Report: they both use the DPHistogramBundles produced by the CreateDPHistogramBundle function in dp_noise_mechanisms.
• DPGroupByFactory::CreateInternal: when delta is not provided, check if the right norm bounds are provided to compute L1 sensitivity (for the Laplace mech).
• CreateRemoteExecutorStack now allows the composing executor to be specified and assigns client values to leaf executors such that all leaf executors receive the same number of clients, except for potentially the last leaf executor, which may receive fewer clients.
• Allow tff.learning.programs.train_model to accept a should_discard_round function to decide whether a round should be discarded and retried.

Removed

• tff.structure.to_container_recursive, this should not be used externally.

Release 0.84.0

Added

• TFF executor classes to the public API (ComposingExecutor, ExecutorTestBase, MockExecutor, ThreadPool).
• Compiler transformation helper functions to the public API (replace_intrinsics_with_bodies, unique_name_generator, transform_preorder, to_call_dominant).
• A method to get the number of checkpoints aggregated to the CheckpointAggregator API.
• The function DPClosedDomainHistogram::IncrementDomainIndices. It allows calling code to iterate through the domain of composite keys (in a do-while loop).

Changed

• Renamed the boolean use_experimental_simulation_loop parameter to loop_implementation that accepts an tff.learning.LoopImplementation enum for all tff.learning.algorithms methods.
• Modified the model output release frequency to every 10 rounds and the final round in tff.learning.programs.train_model.
• Loosened the kEpsilonThreshold constant and updated the tests of DPOpenDomainHistogram accordingly.
• The behavior of DPClosedDomainHistogram::Report(): it now produces an aggregate for each possible combinations of keys. Those composite keys that GroupByAggregator did not already assign an aggregate to are assigned 0. Future CL will add noise.
• Modified tff.learning.algorithms.build_weighted_fed_avg to generate different training graphs when use_experimental_simulation_loop=True and model_fn is of type tff.learning.models.FunctionalModel.

Fixed

• tff.learning.programs.EvaluationManager raised an error when the version IDs of two state-saving operations were the same.
• tff.jax.computation raised an error when the computation has unused arguments.
• The tff.backends.xla execution stack raised an error when single element structures are returned from tff.jax.computation wrapped methods.

Release 0.83.0

Changed

• The tff.learning.programs.train_model program logic to save a deep copy of the data source iterator within the program state.
• The file-backed native program components to not flatten and unflatten values.

Removed

• Unused functions from tensorflow_utils.
• Serializing raw tf.Tensor values to the Value protobuf.
• Partial support for dataclasses.

Release 0.82.0

Added

• A serialized raw array content field to the Array proto.
• A function to DPCompositeKeyCombiner that allows retrieval of an ordinal. Intended for use by the closed-domain DP histogram aggregation core.
• Constants for invalid ordinals and default l0_bound_.
• New DPClosedDomainHistogram class. Sibling of DPOpenDomainHistogram that is constructed from DP parameters plus domain information. No noising yet.

Changed

• How DPCompositeKeyCombiner handles invalid l0_bound_ values.
• The default l0_bound_ value in DPCompositeKeyCombiner to new constant.
• Organization of DP histogram code. Previously, open-domain histogram class + factory class lived side-by-side in dp_group_by_aggregator.h/cc. Now split into dp_open_domain_histogram.h/cc and dp_group_by_factory.h/cc, which will ease future addition of code for closed-domain histogram.
• Moved tff.federated_secure_modular_sum to the mapreduce backend, use tff.backends.mapreduce.federated_secure_modular_sum instead.
• DPGroupByAggregator changes how it checks the intrinsic based on number of domain tensors in the parameter field.
• DPGroupByFactory is now responsible for checking number and type of the parameters in the DPGroupingFederatedSum intrinsic, since the factory is now accessing those parameters.
• Type of domain_tensors in DPCompositeKeyCombiner::GetOrdinal is now TensorSpan (alias of absl::Span<const Tensor>). This will make it possible to retrieve the slice of intrinsic.parameters that contains the domain information and pass it to DPClosedDomainHistogram.
• Switched type of indices in GetOrdinal from FixedArray<size_t> to FixedArray<int64_t>, to better align with internal standards.

Release 0.81.0

Added

• A helper function to get a vector of strings for the elements of a tensor in order to aid in formatting.
• A field string_val to the tensor proto to allow representing string values explicitly.

Changed

• The format of the release notes (i.e., RELEASE.md) to be based on https://keepachangelog.com/en/1.1.0/.
• Moved constraint on linfinity_bound from DPGroupingFederatedSumFactory to DPGroupByFactory, because closed-domain histogram algorithm will use DPGroupingFederatedSum but not demand a positive linfinity_bound.

Removed

• The dependency on semantic-version.
• The tff.async_utils package, use asyncio instead.

Release 0.80.0

Breaking Changes

• Moved 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

Major Features and Improvements

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

Breaking Changes

• Removed farmhashpy dependency.
• 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.

Release 0.78.0

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.
• Added a new field, content, to the Data building block and updated tests.

Bug Fixes

• Fixed #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

• Fixed 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.

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.

Release 0.70.0

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.

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.framework.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.

Release 0.40.0

Major Features and Improvements

• Skip model updates that are non-finite in tff.learning.templates.build_apply_optimizer_finalizer.

Breaking Changes

• Removed deprecated APIs in tff.learning.framework
• Update the Python package scripts to use Python 3.10 by default.
• Remove module wildcard imports from init.py files in TFF.
• Update the Python package scripts to use Python 3.10 by default.

Bug Fixes

• Remove functools.wraps within tff.tf_computation.
• Fix typo in iNaturalist dataset docstring.

Release 0.39.0

Major Features and Improvements

• Added tff.learning.models.FunctionModel support to all methods in tff.learning.algorithms.
• Added support for tf.data.DataSpec to tff.types.infer_unplaced_type.
• Use a tensorflow::ThreadPool for parallelism inside the C++ TensorFlowExecutor.
• Introduced a new tff.experimental_tf_fn_computation tracing decorator that uses FunctionDef instead of GraphDef tracing, providing tf.function auto-control-dependencies.
• Renamed number_of_clients to num_clients in the federated program API.
• Replaced the following API with composers API in tff.learning.templates.
  • tff.learning.framework.build_model_delta_optimizer_process
  • tff.learning.framework.ClientDeltaFn

Bug Fixes

• Fixed a bug in the “Client-efficient large-model federated learning” tutorial to use the correct dense shape.

Release 0.38.0

Major Features and Improvements

• Added tff.learning.models.FunctionalModel support to tff.learning.algorithms.build_mime_lite.
• Updated tensorflow-privacy to version 0.8.6.
• Added an abstract interface describing an asynchronous context
• Removed references to tff.framework.Context.
• Added tff.simulation.datasets.gldv2.get_synthetic.
• Added prefetching data source in tff.program.PrefetchingDataSource.

Breaking Changes

• Deleted deprecated tff.learning.framework.build_encoded_broadcast_process_from_model.
• Deprecated tff.learning.ModelWeights and alias tff.learning.framework.ModelWeights, has now moved to tff.learning.models.ModelWeights. Code should be updated before the next release.

Bug Fixes

• Fixed a bug with variable creation order of metrics in tff.learning.models.functional_model_from_keras.
• Improved tff.tf_computation tracing to also trace functools.partial objects.

Known Bugs

• Colab compatibility: TFF requires Python 3.9 while Colab runtime uses Python 3.7.

Release 0.37.0

Major Features and Improvements

• Added support for Python 3.10.
• Improved support for numpy values in the tff.program API.
• Increased dataset serialization size limit to 100MB.
• Added a new method tff.learning.ModelWeights.convert_variables_to_arrays.
• Added new metrics aggregation factories under tff.learning.metrics.
• Parallelized aggregation in tff.framework.ComposingExecutorFactory.

Breaking Changes

• Updated to use jax and jaxlib version 0.3.14.
• Renamed tff.program.CoroValueReference to tff.program.AwaitableValueReference to reflect the relaxed contract.

Bug Fixes

• Improved documentation for tff.simulation.build_uniform_sampling_fn, tff.learning.robust_aggregator, tff.aggregators.PrivateQuantileEstimationProcess.
• Fixed documentation bug for tutorial “High-performance Simulation with Kubernetes”.
• Fixed bug where momentum hyperparameters were added to SGDM optimizer when momentum was set to 0.
• Removed assertion that preprocessed datasets in a tff.simulation.baselines.BaselineTask have the same element structure.
• Fixed a memory leak when moving numpy arrays across the Python and C++ boundary in the C++ executor.
• Fixed bug in the federated program API when using multiple release managers to release the same value.

Thanks to our Contributors

This release contains contributions from many people at Google, as well as: Madhava Jay, nbishdev@

Release 0.36.0

Major Features and Improvements

• Added support for tff.learning.models.FunctionalModel to tff.learning.algorithms.build_fed_sgd and tff.learning.algorithms.build_fed_prox.
• Increased the gRPC message limit from 1 GB to 2 GB.
• Added hyperparameter getters/setters to various components in tff.learning.

Breaking Changes

• Updated tensorflow to version 2.10.

Bug Fixes

• Improved documentation for tff.analytics.heavy_hitters.iblt.build_iblt_computation().
• Fixed incorrect docstring of tff.federated_select.
• Fixed typo in federated program example.

Release 0.35.0

Major Features and Improvements

• Added get/set_hparams methods to tff.learning.templates.ClientWorkProcess.
• Added tff.learning.algorithms.build_mime_lite_with_optimizer_schedule.
• Updated tensorflow-privacy to version 0.8.5.
• Added tff.learning.entropy_compression_aggregator.
• Added tff.aggregators.EliasGammaEncodedSumFactory.
• Added tff.program.ClientIdDataSource and tff.program.ClientIdDataSourceIterator, for working with a data source of ids of federated clients.

Breaking Changes

• Removed prototype IREE backend.
• Added new dependency on TensorFlow Compression.

Bug Fixes

• Fixed implementation of the loading_remote_data tutorial.
• Corrected the docstring of tff.simulation.datasets.stackoverflow.get_synthetic.

Known Bugs

• TFF's Python 3.9 typing conflicts with Colab's Python 3.7 runtime.

Release 0.34.0

Major Features and Improvements

• Updated to use Bazel version 5.3.0.
• Updated the conventions used to specify the version of a Python dependency, see https://github.com/tensorflow/federated/blob/main/requirements.txt for more information.
• Updated the setup.py to explicitly fail to pip install in Python 3.10. This has always been failing at runtime, but now explicitly fails to install using pip.
• Refreshed loading_remote_data notebook content and added content for FederatedDataSource.
• Added a TFF type_signature attribute to objects of type MapReduceForm.
• Added a series of slides to the GitHub repo (so not part of the PIP package) which detail a technical deep dive into TFF.

Breaking Changes

• Bumped tf-privacy version to 0.8.4.
• Bumped tf-model-optimization version to 0.8.3.
• Removed initialize from MapReduceForm.
• SequenceType now automatically casts any StructWithPythonType that contains a list to a tuple for tf.data compatibility.
• Unified the model_fn and model parameters of tff.learning.algorithms.build_weighted_fed_avg.
• MapReduceForm now takes a type_signature argument in its constructor, and no longer takes an initialize argument.
• MapReduceForm no longer contains an initialize attribute.

Bug Fixes

• Relaxed overly strict type equivalence check to assignability in TFF-TF code generation.

Release 0.33.0

Major Features and Improvements

• Extend tff.analytics.heavy_hitters.iblt with create_chunker API for encoding non-Unicode strings.
• Extend tff.aggregators.DifferentiallyPrivateFactory.tree_aggregation with an optional record_aggregation_factory argument.

Breaking Changes

• Replaced ModularClippingSumFactory with SecureModularSumFactory in tff.analytics.build_hierarchical_histogram_process.

Known Bugs

• TFF's python 3.9 typing conflicts with Colab's Python 3.7 support.

Release 0.32.0

Major Features and Improvements

• Add a MimeLite implementation that allows from optimizer learning rate scheduling in tff.learning.algorithms.build_mime_lite_with_optimizer_schedule.

Breaking Changes

• None

Bug Fixes

• None

Known Bugs

• TFF's python 3.9 typing conflicts with Colab's Python 3.7 support.

Release 0.31.0

Major Features and Improvements

• Added ReleaseManagers to make authoring program logic more convenient.
• Updated TFFs attrs dependency to version 21.4.0.
• Update TFFs tensorflow-privacy dependency to version 0.8.1.

Breaking Changes

• Changed tff.learning.BatchOutput from an attrs class to a namedtuple.
• Removed unused tff.learning.framework.parameter_count_from_model API.

Release 0.30.0

Major Features and Improvements

• Add tests for namedtuples in the tff.program package.
• Add num_subrounds parameter to the mergeable context, allowing callers to optionally sequentialize subrounds.
• Add metrics support to tff.learning.models.FunctionalModel, including updates to the helper function create_functional_metric_fns and the downstream caller tff.learning.algorithms.build_weighted_fed_avg.

Bug Fixes

• Fix typo in the types constructed for testing the tff.program package.
• Fix some program example typos.
• Fix tests that don't seem to be running under the CI.
• Fix naming bug for Python mergeable execution.
• Ensure exceptions raised from remote executor stub implement gRPC error interface.
• Update tff.structure.Struct integration with JAX pytrees to not flatten the entire structure.
• Use Python 3.7 compatible type annotations until Colab updates to Python 3.9.

Release 0.29.0

Major Features and Improvements

• Update the MemoryReleaseManager to save type_signature when releasing values.
• Add a type_signature parameter to the ReleaseManager.release method.
• Unify retryability logic between TFF-C++ and TFF-Python.
• Update the TFF contributions and collaboration links to point to the Discord server.

Breaking Changes

• Move Python executor stacks file to python_executor_stacks.py in executor_stacks directory.

Bug Fixes

• Ensure that dataset deserialization will return ragged and sparse tensors, as needed according to the TFF type of the dataset.
• Make metric_finalizers use metric constructors if available.

Release 0.28.0

Major Features and Improvements

• Updated tutorials to use tff.learning.algorithms API.
• Asynchronous TFF execution contexts no longer assume a single global cardinality; they concurrently invoke any computation for which concurrency is requested.

Breaking Changes

• Removed tff.learning.build_federated_averaging_process; users should migrate to tff.learning.algorithms.build_weighted_fed_avg.

Bug Fixes

• Clarified semantics for TFF-C++ multimachine Dispose, DisposeExecutor, and executor keying, to avoid raising exceptions and spamming logs in the course of normal operation.
• Fixed unsigned integer overflow for TFF-C++ max_concurrent_computation_calls.
• Normalizes on call-dominant form before attempting to compile to MergeableCompForm, removing spurious failures on dependent-aggregation checking.

Known Bugs

• Serialization / deserialization of tf.data.Datasets yielding non-dense tensors for multimachine runtime may encounter issues:
  • tff.framework.deserialize_value may fail to deserialize tf.data.Datasets yielding RaggedTensors or SparseTensors.
  • tff.framework.serialize_value may fail to serialize tf.data.Datasets yielding SparseTensors.

Release 0.27.0

Major Features and Improvements

• New Colab notebook illustrating how to use DataBackend to load remote datasets.
• Added a CreateDataDescriptor helper function.
• Added a worker binary serving the TFF-C++ executor service.

Bug Fixes

• Fixed bug with intermediate aggregation and controller failures, causing hangs.

Release 0.26.0

Major Features and Improvements

• Updated TensorFlow to 2.9.1.
• Update pybind11 to 2.9.2.
• Re-enable cpp_fast_protos.
• Introduces container class to run coroutines in a dedicated thread, allowing TFF’s synchronous execution interfaces to be used in conjunction with other asyncio code.
• Use latest TFF version in Colab notebook links.
• Rename the helper functions that create test MeasuredProcesses.
• Add a compiler transform checking Tensorflow computations against list of allowed ops.
• Explicitly specify return types in the program package.
• Adds convenience function for setting a local async CPP execution context.
• Move jax components into a non-experimental namespace.

Breaking Changes

• Switch compilation flag _GLIBCXX_USE_CXX11_ABI to 1.

Release 0.25.0

Major Features and Improvements

• Adds error message logging to TFF C++ execution context.
• Adds test coverage for C++ runtime with aggregators.
• Redefines 'workers going down with fixed clients per round' test.
• Add complete examples of using DataBackend with TFF comps.
• Updated the MapReduceForm documentation to include the two additional secure sum intrinsics.
• tff.learning
  • Relax the type check on LearningProcess from strictly SequenceType to also allow structures of SequenceType.

Breaking Changes

• Remove usage of tff.test.TestCase, tff.test.main(), and delete test_case module.
• Update test utility docstrings to use consistent vocabulary.
• Update to TensorFlow 2.9.0
• Rename up compiler/test_utils to compiler/building_block_test_utils.
• Remove some unnecessary usage of pytype: skip-file.
• Specify the None return type of ReleaseManager.release.
• Remove usage of deprecated numpy types.
• Replace depreciated random_integers with randint.

Bug Fixes

• Fix numpy warning.

Release 0.24.0

Major Features and Improvements

• Added asyncio.run call to metrics manager release calls to ensure compatibility with https://github.com/tensorflow/federated/commit/a98b5ed6894c536549da06b4cc7ed116105dfe65.
• Added an example and documentation for the Federated Program API.
• Improved model_update_aggregator to support structures with mixed floating dtypes.
• Create a mirror of tff.simulation.compose_dataset_computation_with_iterative_process for tff.learning.templates.LearningProcess.
• Added logging of expected sequential computations to local TFF-C++ runtime.

Breaking Changes

• Moved asserts from tff.test.TestCase to tff.test.* as functions.
• Removed assert_type_assignable_from function.
• Moved assert_nested_struct_eq to the type_conversions_test module.
• Removed client_train_process and fedavg_ds_loop comparisons.

Bug Fixes

• Fixed comparisons to enums in the benchmarks package.
• Fixed async_utils.SharedAwaitable exception raiser.
• Fixed various lint errors.

Release 0.23.0

Major Features and Improvements

• Deprecated tff.learning.build_federated_averaging_process.
• Added an API to convert tf.keras.metrics.Metric to a set of pure tf.functions.

Breaking Changes

• Renamed ProgramStateManager.version to ProgramStateManager.get_versions.

Bug Fixes

• Fixed the "datasets/" path in the working with TFF's ClientData tutorial.

Release 0.22.0

Major Features and Improvements

• Updated .bazelversion to 5.1.1.
• Updated the tff.program API to use asyncio.
• Exposed new APIs in the tff.framework package:
  • tff.framework.CardinalitiesType.
  • tff.framework.PlacementLiteral.
  • tff.framework.merge_cardinalities.
• tff.analytics
  • Added new analytic_gauss_stddev API.

Breaking Changes

• Renamed ProgramStateManager.version to ProgramStateManager.get_versions.

Bug Fixes

• Fixed some Python lint errors related to linting Python 3.9.
• Cleaned up stale TODOs throughout the codebase.

Known Bugs

• Version 0.21.0 currently fails to import in colab if the version of Python is less than Python 3.9. Please use a runtime with a version of Python greater than Python 3.9 or use TFF version 0.20.0.

Release 0.21.0

Major Features and Improvements

• tff.analytics
  • Added new tff.analytics.IbltFactory aggregation factory.
  • Added new IBTL tensor encoder/decoder libraries and uses them in tff.analytics.heavy_hitters.iblt.build_iblt_computation.
• tff.aggregator
  • Added as_weighted_aggregator to the tff.aggregator.Factory API.
• tff.learning
  • Improved compilation and execution performance of tff.learning.metrics.secure_sum_then_finalize by grouping tensors by DType.
  • Added set_model_weights method and default implementation to tff.learning.templates.LearningProcess.
  • Added a new reset_metrics attribute to tff.learning.Model.
  • Added schedule_learning_rate to tff.learning.optimizers.
  • Added new tff.learning.ddp_secure_aggregator for Distributed Differential Privacy.
• tff.simulation
  • Added an option to distort train images in the CIFAR-100 baseline task.
  • Changed the default sequence length for the Shakespeare baseline task to a more reasonable value.
• Core
  • Switched runtime to create new RemoteExecutors with different cardinalities, rather than resetting the cardinality in the remote service.

Breaking Changes

• Removed support for Python 3.7 and 3.8, TFF supports 3.9 and later.
• Removed deprecated attributes report_local_outputs and federated_output_computation from tff.learning.Model
• Removed the ingest method from tff.Context

Bug Fixes

• Multiple typos in tests, code comments, and pydoc.

Known Bugs

• Sequences (datasets) of SparseTensors don't work on the C++ runtime.
• Computations when CLIENTS cardinality is zero doesn't work on the Python runtime.
• Assigning variables to a Keras model after construction inside a model_fn results in a non-deterministic graph.

Release 0.20.0

Major Features and Improvements

• Added tff.program API; this API is still in active development but can be used to compose shared and platform specific: program logic, components, and privacy concepts to create federated programs.
• Added support for Python 3.9.
• Added CelebA and iNaturalist datasets to tff.simulation.datasets.
• Added tff.analytics API for federated analytics, including private heavy hitters algorithms.
• Added tff.learning.algorithms API, including TFF implementations of FedProx, FedAvg with learning rate scheduling, federated k-Means, and MimeLite.
• Added tff.learning.metrics API to support easy configuration of cross-client metrics aggregation via the new metrics_aggregator argument.
• Added metrics_aggregator argument to tff.learning.build_federated_averaging_process and tff.learning.build_federated_evaluation.
• Added report_local_unfinalized_metrics and metric_finalizers methods to tff.learning.Model and deprecated report_local_outputs and federated_output_computation.
• Added tff.learning.optimizers API for building purely functional optimizers and implementations of SGD, Adagrad, Rmsprop, Adam, Yogi,
• Added tff.learning.reconstruction API for building partially local federated learning algorithms, including Federated Reconstruction.
• Added tff.learning.templates API to support building learning algorithms in a modular fashion.
• Added tff.simulation.baselines API to support evaluating learning algorithms on a suite of representative tasks.
• Added tff.aggregators.DifferentiallyPrivateFactory.tree_aggregation to support the DP-FTRL algorithm.
• Added tff.aggregators.SecureModularSumFactory
• Added tff.aggregators.DiscreteFourierTransformFactory and tff.aggregators.HadamardTransformFactory to support rotation-based aggregators.
• Added tff.aggregators.concat_factory for aggregating structures as a single tensor.
• Added tff.backends.native.create_mergeable_comp_execution_context, tff.backends.native.set_mergeable_comp_execution_context; these can be used with a distributed runtime to scale to tens of thousands of clients.
• Improved performance of many tff.simulation.datasets.ClientData subclasses.
• Added tff.simulation.datasets.ClientData.serializable_dataset_fn attribute, enabling dataset creation within TF/TFF computations.
• Added debug_measurements option to aggregators in tff.learning.
• Added support for unambiguous zero-client aggregations.
• Added support for Python dataclasses as function parameters and return values for TFF computations.
• Added automatic insertion of tff.federated_zip to invocation of user-defined TFF federated computations.
• Added utilities to tff.simulation.datasets for saving federated datasets to a SQL database compatible with tff.simulation.datasets.SqlClientData.
• Added tff.learning.models.FunctionalModel and tff.learning.models.functional_model_from_keras.
• Increased max flow of tensors. Tensors now flow here, there, and everywhere.
• Updated the Python dependencies:
• Updated absl-py to version 1.0.0.
• Updated attrs to version 21.2.0.
• Added farmhashpy version 0.4.0.
• Updated jax to version 0.2.27.
• Updated jaxlib to version 0.1.76.
• Updated numpy to version 1.21.4.
• Removed retrying.
• Updated tensorflow-model-optimization to version 0.7.1.
• Updated tensorflow-model-optimization to version 0.7.3.
• Updated tensorflow to version 2.8.0.
• Added support for building many dependencies including tensorflow using Bazel.
• Updated the Bazel dependencies:
• Updated rules_python to version 0.5.0.
• Updated com_google_protobuf to version v3.18.0-rc1.
• Added absl_py version 1.0.0.
• Added com_google_googletest version release-1.11.0.
• Added io_bazel_rules_go version v0.29.0.
• Added bazel_skylib version 1.0.3.
• Added pybind11_abseil.
• Added pybind11_bazel.
• Added pybind11_protobuf.
• Added com_google_absl version 20211102.0.
• Added tensorflow_org version v2.8.0.

Breaking Changes

• Removed support for building source on MacOS.
• Removed support for Python 3.6.
• Removed symbol tff.framework.type_contains, use tff.types.contains instead.
• Removed many symbols from tff.simulation, these can be found in tff.program instead.
• Removed support for converting non-OrderedDict mapping types to tff.Values.
• Removed tff.simulation.datasets.ClientData.from_clients_and_fn in favor of tff.simulation.datasets.ClientData.from_clients_and_tf_fn.
• Restricted tff.simulation.datasets.ClientData.preprocess to only support TF-serializable functions.
• Removed tff.backends.reference, and the reference context it contained.
• Removed tff.learning.build_federated_sgd_process in favor of tff.learning.algorithms.build_fed_sgd.
• Removed tff.simulation.run_simulation in favor of tff.simulation.run_training_process.
• Removed tff.learning.framework.EnhancedModel.
• Removed tff.learning.framework.build_stateless_mean.

Bug Fixes

• Fixed broken links in documentation.
• Fixed many pytype errors.
• Fixed some inconsistencies in Bazel visibility.
• Fixed bug where tff.simulation.datasets.gldv2.load_data() would result in an error.

Release 0.19.0

Major Features and Improvements

• Introduced new intrinsics: federated_select and federated_secure_select.
• New tff.structure_from_tensor_type_tree to help manipulate structures of tff.TensorType into structures of values.
• Many new tff.aggregators factory implementations.
• Introduced tf.data concept for data URIs.
• New tff.type package with utilities for working with tff.Type values.
• Initial experimental support for tff.jax_computation.
• Extend tff.tf_computation support to SpareTensor and RaggedTensor.

Breaking Changes

• Update gRPC dependency to 1.34.
• Moved ClientData interface and implementations to tff.simulation.datasets.
• Renamed tff.utils.update_state to tff.structure.update_struct.
• Removed the tff.utils namespace, all symbols have migrated, many to tff.aggregators.
• Moved infinite EMNIST dataset to federated research repository.
• Removes rpc_mode argument to remote executors, along with streaming mode.
• Removes deprecated tff.federated_apply.
• Removes tff.federated_reduce, all usages can use tff.federated_aggregate.
• Removes HDF5ClientData and h5py pip dependency.
• Removes setattr functionality on tff.ValueImpl.

Bug Fixes

• Improved tf.GraphDef comparisons.
• Force close generators used for sending functions to computation wrappers, avoiding race conditions in Colab.
• Fix tracing libraries asyncio usage to be Python3.9 compatible.
• Fix issue with destruction of type intern pool destructing and abc.
• Fix type interning for tensors with unknown dimensions.
• Fix ClientData.create_dataset_from_all_clients consuming unreasonable amounts of memory/compute time.

Release 0.18.0

Major Features and Improvements

• Extended the tff.simulation package to add many new tools for running simulations (checkpoints and metrics managers, client sampling functions).
• Extended the tff.aggregators package with a number of new aggregation factories.
• Added the tff.structure API to expose the Struct class and related functions.
• Added the tff.profiler API to expose useful profiling related functions.
• Added the tff.backends.test package to expose backends that focused on testing specifically a way to test a computation with a federated_secure_sum intrinsic.
• Added the tff.experimental package to expose less stable API.

Breaking Changes

• Replaced the tff.aggregators.AggregationProcessFactory abstract base class with the tff.aggregators.UnweightedAggregationFactory and the tff.aggregators.WeightedAggregationFactory classes.
• Replaced the tff.aggregators.ZeroingFactory class with a tff.aggregators.zeroing_factory function with the same input arguments.
• Replaced the tff.aggregators.ClippingFactory class with a tff.aggregators.clipping_factory function with the same input arguments.
• Updated tensorflow package dependency to 2.4.0.
• Updated absl-py package dependency to 0.10.
• Updated grpcio package dependency to 1.32.0.
• Added a jaxlib package dependency at 0.1.55.
• Updated numpy package dependency to 1.19.2.
• Updated tensorflow-addons package dependency to 0.12.0.
• Updated tensorflow-model-optimization package dependency to 0.5.0.

Bug Fixes

• Fixed issue with the sequence_reduce intrinsic handling federated types.

Release 0.17.0

Major Features and Improvements

• New tff.aggregators package with interfaces for stateful aggregation compositions.
• New Google Landmark Dataset tff.simulations.dataset.gldv2
• New convenience APIs tff.type_clients and tff.type_at_server
• Invert control of computation tracing methods to produce clearer Python stack traces on error.
• Move executor creation to a factory pattern in executor service, allowing distributed runtimes to be agnostic to number of clients.
• Significant improvements of type serialization/deserialization
• New tff.simulations.compose_dataset_computation_with_iterative_process API to move execution of client dataset construction to executor stack leaves.
• Extend parameterization of tff.learning.build_federated_averaging_process with use_experimental_simulation_loop argument to better utilize multi-GPU setups.

Breaking Changes

• Removed tff.utils.StatefulFn, replaced by tff.templates.MeasuredProcess.
• Removed tff.learning.assign_weights_to_keras_model
• Stop removing OptimizeDataset ops from tff.tf_computations.
• The research/ directory has been moved to http://github.com/google-research/federated.
• Updates to input_spec argument for tff.learning.from_keras_model.
• Updated TensorFlow dependency to 2.3.0.
• Updated TensorFlow Model Optimization dependency to 0.4.0.

Bug Fixes

• Fixed streaming mode hang in remote executor.
• Wrap collections.namedtuple._asdict calls in collections.OrderedDict to support Python 3.8.
• Correctly serialize/deserialize tff.TensorType with unknown shapes.
• Cleanup TF lookup HashTable resources in TFF execution.
• Fix bug in Shakespeare dataset where OOV and last vocab character were the same.
• Fix TFF ingestion of Keras models with shared embeddings.
• Closed hole in compilation to CanonicalForm.

Known Bugs

• "Federated Learning for Image Classification" tutorial fails to load projector plugin for tensorboard. (https://github.com/tensorflow/federated/issues/914)
• Certain Keras models with activity regularization fail in execution with unliftable error (https://github.com/tensorflow/federated/issues/913).

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

amitport, ronaldseoh

Release 0.16.1

Bug Fixes

• Fixed issue preventing Python lists from being all_equal values.

Release 0.16.0

Major Features and Improvements

• Mirrored user-provided types and minimize usage of AnonymousTuple.

Breaking Changes

• Renamed AnonymousTuple to Struct.

Release 0.15.0

Major Features and Improvements

• Updated tensorflow-addons package dependency to 0.9.0.
• Added API to expose the native backend more conveniently. See tff.backends.native.* for more information.
• Added a compiler argument to the tff.framework.ExecutionContext API and provided a compiler for the native execution environment, which improves TFF’s default concurrency pattern.
• Introduced a new tff.templates.MeasuredProcess concept, a specialization of tff.templates.IterativeProcess.
• Extends tff.learning interfaces to accept tff.templates.MeasuredProcess objects for aggregation and broadcast computations.
• Introduce new convenience method tff.learning.weights_type_from_model.
• Introduced the concept of a tff.framework.FederatingStrategy, which parameterizes the tff.framework.FederatingExecutor so that the implementation of a specific intrinsic is easier to provide.
• Reduced duplication in TFF’s generated ASTs.
• Enabled usage of GPUs on remote workers.
• Documentation improvements.

Breaking Changes

• The IterativeProcess return from tff.learning.build_federated_averaging_process and tff.learning.build_federated_sgd_process now zip the second tuple output (the metrics) to change the result from a structure of federated values to to a federated structure of values.
• Removed tff.framework.set_default_executor function, instead you should use the more convenient tff.backends.native.set_local_execution_context function or manually construct a context an set it using tff.framework.set_default_context.
• The tff.Computation base class now contains an abstract __hash__ method, to ensure compilation results can be cached. Any custom implementations of this interface should be updated accordingly.

Bug Fixes

• Fixed issue for missing variable initialization for variables explicitly not added to any collections.
• Fixed issue where table initializers were not run if the tff.tf_computation decorated function used no variables.

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

jvmcns@

Release 0.14.0

Major Features and Improvements

• Multiple TFF execution speedups.
• New tff.templates.MeasuredProcess specialization of IterativeProcess.
• Increased optimization of the tff.templates.IterativeProcess -> tff.backends.mapreduce.CanonicalForm compiler.

Breaking Changes

• Moved tff.utils.IterativeProcess to tff.templates.IterativeProcess.
• Removed tff.learning.TrainableModel, client optimizers are now arguments on the tff.learning.build_federated_averaging_process.
• Bump required version of pip packages for tensorflow (2.2), numpy (1.18), pandas (0.24), grpcio (1.29).

Bug Fixes

• Issue with GPUs in multimachine simulations not being utilized, and bug on deserializing datasets with GPU-backed runtime.
• TensorFlow lookup table initialization failures.

Known Bugs

• In some situations, TF will attempt to push Datasets inside of tf.functions over GPU device boundaries, which fails by default. This can be hit by certain usages of TFF, e.g. as tracked here.

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

jvmcns@

Release 0.13.1

Bug Fixes

• Fixed issues in tutorial notebooks.

Release 0.13.0

Major Features and Improvements

• Updated absl-py package dependency to 0.9.0.
• Updated h5py package dependency to 2.8.0.
• Updated numpy package dependency to 1.17.5.
• Updated tensorflow-privacy package dependency to 0.2.2.

Breaking Changes

• Deprecated dummy_batch parameter of the tff.learning.from_keras_model function.

Bug Fixes

• Fixed issues with executor service using old executor API.
• Fixed issues with remote executor test using old executor API.
• Fixed issues in tutorial notebooks.

Release 0.12.0

Major Features and Improvements

• Upgraded tensorflow dependency from 2.0.0 to 2.1.0.
• Upgraded tensorflow-addons dependency from 0.6.0 to 0.7.0.
• Upgraded attr dependency from 18.2 to 19.3.
• Upgraded tfmot dependency from 0.1.3 to 0.2.1.
• Added a federated partition of the CIFAR-100 dataset to tff.simulation.datasets.cifar100.
• Made the high performance, parallel executor the default (replacing the reference executor).
• Added a new tff.learning.build_personalization_eval for evaluating model personalization strategies.
• Added new federated intrinsic tff.federated_secure_sum.
• tff.learning.build_federated_averaing_process() now takes a client_optimizer_fn and a tff.learning.Model. tff.learning.TrainableModel is now deprecated.
• Improved performance in the high performance executor stack.
• Implemented and exposed tff.framework.ExecutorFactory; all tff.framework...executor_factory calls now return an instance of this class.
• Added remote_executor_example binary which demonstrates using the RemoteExecutor across multi-machine deployments.
• Added close() method to the Executor, allowing subclasses to proactively release resources.
• Updated documentation and scripts for creating Docker images of the TFF runtime.
• Automatically call tff.federated_zip on inputs to other federated intrinsics.

Breaking Changes

• Dropped support for Python2.
• Renamed tff.framework.create_local_executor (and similar methods) to tff.framework.local_executor_factory.
• Deprecated federated_apply(), instead use federated_map() for all placements.

Bug Fixes

• Fixed problem with different instances of the same model having different named types. tff.learning.ModelWeights no longer names the tuple fields returned for model weights, instead relying on an ordered list.
• tff.sequence_* on unplaced types now correctly returns a tff.Value.

Known Bugs

• tff.sequence_*.. operations are not implemented yet on the new high-performance executor stack.
• A subset of previously-allowed lambda captures are no longer supported on the new execution stack.

Release 0.11.0

Major Features and Improvements

• Python 2 support is now deprecated and will be removed in a future release.
• federated_map now works with both tff.SERVER and tff.CLIENT placements.
• federated_zip received significant performance improvements and now works recursively.
• Added retry logic to gRPC calls in the execution stack.

Breaking Changes

• collections.OrderedDict is now required in many places rather than standard Python dictionaries.

Bug Fixes

• Fixed computation of the number of examples when Keras is using multiple inputs.
• Fixed an assumption that tff.framework.Tuple is returned from IterativeProcess.next.
• Fixed argument packing in polymorphic invocations on the new executor API.
• Fixed support for dir() in ValueImpl.
• Fixed a number of issues in the Colab / Jupyter notebook tutorials.

Release 0.10.1

Bug Fixes

• Updated to use grpcio 1.24.3.

Release 0.10.0

Major Features and Improvements

• Add a federated_sample aggregation that is used to collect a sample of client values on the server using reservoir sampling.
• Updated to use tensorflow 2.0.0 and tensorflow-addons 0.6.0 instead of the coorisponding nightly package in the setup.py for releasing TFF Python packages.
• Updated to use tensorflow-privacy 0.2.0.
• Added support for attr.s classes type annotations.
• Updated streaming Execute method on tff.framework.ExecutorService to be asynchronous.
• PY2 and PY3 compatibility.

Release 0.9.0

Major Features and Improvements

• TFF is now fully compatible and dependent on TensorFlow 2.0
• Add stateful aggregation with differential privacy using TensorFlow Privacy (https://pypi.org/project/tensorflow-privacy/).
• Additional stateful aggregation lwith compression using TensorFlow Model Optimization (https://pypi.org/project/tensorflow-model-optimization/).
• Improved executor stack for simulations, documentation and scripts for starting simulations on GCP.
• New libraries for creating synthetic IID and non-IID datsets in simulation.

Breaking Changes

• examples package split to simulation and research.

Bug Fixes

• Various error message string improvements.
• Dataset serialization fixed for V1/V2 datasets.
• tff.federated_aggregate supports accumulate, merge and report methods with signatures containing tensors with undefined dimensions.

Release 0.8.0

Major Features and Improvements

• Improvements in the executor stack: caching, deduplication, bi-directional streaming mode, ability to specify physical devices.
• Components for integration with custom mapreduce backends (tff.backends.mapreduce).
• Improvements in simulation dataset APIs: ConcreteClientData, random seeds, stack overflow dataset, updated documentation.
• Utilities for encoding and various flavors of aggregation.

Breaking Changes

• Removed support for the deprecated tf.data.Dataset string iterator handle.
• Bumps the required versions of grpcio and tf-nightly.

Bug Fixes

• Fixes in notebooks, typos, etc.
• Assorted fixes to align with TF 2.0.
• Fixes thread cleanup on process exit in the high-performance executor.

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Gui-U@, Krishna Pillutla, Sergii Khomenko.

Release 0.7.0

Major Features and Improvements

• High-performance simulation components and tutorials.

Breaking Changes

• Refactoring/consolidation in utility functions in tff.framework.
• Switches some of the tutorials to new PY3-only executor stack components.

Bug Fixes

• Includes the examples directory in the pip package.
• Pip installs for TensorFlow and TFF in tutorials.
• Patches for asyncio in tutorials for use in Jupyter notebooks.
• Python 3 compatibility issues.
• Support for federated_map_all_equal in the reference executor.
• Adds missing implementations of generic constants and operator intrinsics.
• Fixes missed link in compatibility section of readme.
• Adds some of the missing intrinsic reductions.

Thanks to our Contributors

This release contains contributions from many people at Google.

Release 0.6.0

Major Features and Improvements

• Support for multiple outputs and loss functions in keras models.
• Support for stateful broadcast and aggregation functions in federated averaging and federated SGD APIs.
• tff.utils.update_state extended to handle more general state arguments.
• Addition of tff.utils.federated_min and tff.utils.federated_max.
• Shuffle client_ids in create_tf_dataset_from_all_clients by default to aid optimization.

Breaking Changes

• Dependencies added to requirements.txt; in particular, grpcio and portpicker.

Bug Fixes

• Removes dependency on tf.data.experimental.NestedStructure.

Thanks to our Contributors

This release contains contributions from many people at Google, as well as:

Dheeraj R Reddy, @Squadrick.

Release 0.5.0

Major Features and Improvements

• Removed source level TF dependencies and switched from tensorflow to tf-nightly dependency.
• Add support for attr module in TFF type system.
• Introduced new tff.framework interface layer.
• New AST transformations and optimizations.
• Preserve Python container usage in tff.tf_computation.

Bug Fixes

• Updated TFF model to reflect Keras tf.keras.model.weights order.
• Keras model with multiple inputs. #416

Release 0.4.0

Major Features and Improvements

• New tff.simulation.TransformingClientData API and associated infinite EMNIST dataset (see http://tensorflow.org/federated/api_docs/python/tff for details)

Breaking Change

• Normalized func to fn across the repository (rename some parameters and functions)

Bug Fixes

• Wrapped Keras models can now be used with tff.learning.build_federated_evaluation
• Keras models with non-trainable variables in intermediate layers (e.g. BatchNormalization) can be assigned back to Keras models with tff.learning.ModelWeights.assign_weights_to

Release 0.3.0

Breaking Changes

• Rename tff.learning.federated_average to tff.learning.federated_mean.
• Rename 'func' arguments to 'fn' throughout the API.

Bug Fixes

• Assorted fixes to typos in documentation and setup scripts.

Release 0.2.0

Major Features and Improvements

• Updated to use TensorFlow version 1.13.1.
• Implemented Federated SGD in tff.learning.build_federated_sgd_process().

Breaking Changes

• next() function of tff.utils.IteratedProcesss returned by build_federated_*_process() no longer unwraps single value tuples (always returns a tuple).

Bug Fixes

• Modify setup.py to require TensorFlow 1.x and not upgrade to 2.0 alpha.
• Stop unpacking single value tuples in next() function of objects returned by build_federated_*_process().
• Clear cached Keras sessions when wrapping Keras models to avoid referencing stale graphs.

Release 0.1.0

• Initial public release.