- Added
tff.learning.models.FunctionalModelsupport totff.learning.algorithms.build_mime_lite. - Updated
tensorflow-privacyto version0.8.6. - 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.
- Deleted deprecated
tff.learning.framework.build_encoded_broadcast_process_from_model. - Deprecated
tff.learning.ModelWeightsand aliastff.learning.framework.ModelWeights, has now moved totff.learning.models.ModelWeights. Code should be updated before the next release.
- Fixed a bug with variable creation order of metrics in
tff.learning.models.functional_model_from_keras. - Improved
tff.tf_computationtracing to also tracefunctools.partialobjects.
- Colab compatibility: TFF requires Python 3.9 while Colab runtime uses Python 3.7.
- Added support for Python 3.10.
- Improved support for
numpyvalues in thetff.programAPI. - 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.
- Updated to use
jaxandjaxlibversion0.3.14. - Renamed
tff.program.CoroValueReferencetotff.program.AwaitableValueReferenceto reflect the relaxed contract.
- 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.BaselineTaskhave 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.
This release contains contributions from many people at Google, as well as: Madhava Jay, nbishdev@
- Added support for
tff.learning.models.FunctionalModeltotff.learning.algorithms.build_fed_sgdandtff.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.
- Updated
tensorflowto version2.10.
- Improved documentation for
tff.analytics.heavy_hitters.iblt.build_iblt_computation(). - Fixed incorrect docstring of
tff.federated_select. - Fixed typo in federated program example.
- Added get/set_hparams methods to
tff.learning.templates.ClientWorkProcess. - Added
tff.learning.algorithms.build_mime_lite_with_optimizer_schedule. - Updated
tensorflow-privacyto version0.8.5. - Added
tff.learning.entropy_compression_aggregator. - Added
tff.aggregators.EliasGammaEncodedSumFactory. - Added
tff.program.ClientIdDataSourceandtff.program.ClientIdDataSourceIterator, for working with a data source of ids of federated clients.
- Removed prototype IREE backend.
- Added new dependency on TensorFlow Compression.
- Fixed implementation of the
loading_remote_datatutorial. - Corrected the docstring of
tff.simulation.datasets.stackoverflow.get_synthetic.
- TFF's Python 3.9 typing conflicts with Colab's Python 3.7 runtime.
- Updated to use
Bazelversion5.3.0. - 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.pyto explicitly fail topip installin Python 3.10. This has always been failing at runtime, but now explicitly fails to install usingpip. - Refreshed loading_remote_data notebook content and added content for
FederatedDataSource. - Added a TFF
type_signatureattribute to objects of typeMapReduceForm. - Added a series of slides to the GitHub repo (so not part of the PIP package) which detail a technical deep dive into TFF.
- Bumped tf-privacy version to
0.8.4. - Bumped tf-model-optimization version to
0.8.3. - Removed
initializefromMapReduceForm. SequenceTypenow automatically casts anyStructWithPythonTypethat contains alistto atuplefortf.datacompatibility.- Unified the
model_fnandmodelparameters oftff.learning.algorithms.build_weighted_fed_avg. MapReduceFormnow takes atype_signatureargument in its constructor, and no longer takes aninitializeargument.MapReduceFormno longer contains aninitializeattribute.
- Relaxed overly strict type equivalence check to assignability in TFF-TF code generation.
- Extend
tff.analytics.heavy_hitters.ibltwithcreate_chunkerAPI for encoding non-Unicode strings. - Extend
tff.aggregators.DifferentiallyPrivateFactory.tree_aggregationwith an optionalrecord_aggregation_factoryargument.
- Replaced
ModularClippingSumFactorywithSecureModularSumFactoryintff.analytics.build_hierarchical_histogram_process.
- TFF's python 3.9 typing conflicts with Colab's Python 3.7 support.
- Add a MimeLite implementation that allows from optimizer learning rate
scheduling in
tff.learning.algorithms.build_mime_lite_with_optimizer_schedule.
- None
- None
- TFF's python 3.9 typing conflicts with Colab's Python 3.7 support.
- Added
ReleaseManagers to make authoring program logic more convenient. - Updated TFFs
attrsdependency to version21.4.0. - Update TFFs
tensorflow-privacydependency to version0.8.1.
- Changed
tff.learning.BatchOutputfrom an attrs class to a namedtuple. - Removed unused
tff.learning.framework.parameter_count_from_modelAPI.
- Add tests for
namedtuples in thetff.programpackage. - Add
num_subroundsparameter to the mergeable context, allowing callers to optionally sequentialize subrounds. - Add metrics support to
tff.learning.models.FunctionalModel, including updates to the helper functioncreate_functional_metric_fnsand the downstream callertff.learning.algorithms.build_weighted_fed_avg.
- Fix typo in the types constructed for testing the
tff.programpackage. - 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.Structintegration with JAX pytrees to not flatten the entire structure. - Use Python 3.7 compatible type annotations until Colab updates to Python 3.9.
- Update the
MemoryReleaseManagerto savetype_signaturewhen releasing values. - Add a
type_signatureparameter to theReleaseManager.releasemethod. - Unify retryability logic between TFF-C++ and TFF-Python.
- Update the TFF contributions and collaboration links to point to the Discord server.
- Move Python executor stacks file to
python_executor_stacks.pyinexecutor_stacksdirectory.
- Ensure that dataset deserialization will return ragged and sparse tensors, as needed according to the TFF type of the dataset.
- Make
metric_finalizersuse metric constructors if available.
- Updated tutorials to use
tff.learning.algorithmsAPI. - Asynchronous TFF execution contexts no longer assume a single global cardinality; they concurrently invoke any computation for which concurrency is requested.
- Removed
tff.learning.build_federated_averaging_process; users should migrate totff.learning.algorithms.build_weighted_fed_avg.
- 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.
- Serialization / deserialization of tf.data.Datasets yielding non-dense
tensors for multimachine runtime may encounter issues:
tff.framework.deserialize_valuemay fail to deserialize tf.data.Datasets yielding RaggedTensors or SparseTensors.tff.framework.serialize_valuemay fail to serialize tf.data.Datasets yielding SparseTensors.
- New Colab notebook illustrating how to use
DataBackendto load remote datasets. - Added a CreateDataDescriptor helper function.
- Added a worker binary serving the TFF-C++ executor service.
- Fixed bug with intermediate aggregation and controller failures, causing hangs.
- 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
programpackage. - Adds convenience function for setting a local async CPP execution context.
- Move jax components into a non-experimental namespace.
- Switch compilation flag
_GLIBCXX_USE_CXX11_ABIto1.
- 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
DataBackendwith 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.
- Remove usage of
tff.test.TestCase,tff.test.main(), and deletetest_casemodule. - Update test utility docstrings to use consistent vocabulary.
- Update to TensorFlow 2.9.0
- Rename up
compiler/test_utilstocompiler/building_block_test_utils. - Remove some unnecessary usage of
pytype: skip-file. - Specify the
Nonereturn type ofReleaseManager.release. - Remove usage of deprecated numpy types.
- Replace depreciated
random_integerswithrandint.
- Fix numpy warning.
- Added
asyncio.runcall 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_aggregatorto support structures with mixed floating dtypes. - Create a mirror of
tff.simulation.compose_dataset_computation_with_iterative_processfortff.learning.templates.LearningProcess. - Added logging of expected sequential computations to local TFF-C++ runtime.
- Moved asserts from
tff.test.TestCasetotff.test.*as functions. - Removed
assert_type_assignable_fromfunction. - Moved
assert_nested_struct_eqto thetype_conversions_testmodule. - Removed
client_train_processand fedavg_ds_loop comparisons.
- Fixed comparisons to enums in the benchmarks package.
- Fixed
async_utils.SharedAwaitableexception raiser. - Fixed various lint errors.
- Deprecated
tff.learning.build_federated_averaging_process. - Added an API to convert
tf.keras.metrics.Metricto a set of puretf.functions.
- Renamed
ProgramStateManager.versiontoProgramStateManager.get_versions.
- Fixed the "datasets/" path in the working with TFF's ClientData tutorial.
- Updated .bazelversion to
5.1.1. - Updated the
tff.programAPI to useasyncio. - Exposed new APIs in the
tff.frameworkpackage:tff.framework.CardinalitiesType.tff.framework.PlacementLiteral.tff.framework.merge_cardinalities.
tff.analytics- Added new
analytic_gauss_stddevAPI.
- Added new
- Renamed
ProgramStateManager.versiontoProgramStateManager.get_versions.
- Fixed some Python lint errors related to linting Python 3.9.
- Cleaned up stale TODOs throughout the codebase.
- 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.
tff.analytics- Added new
tff.analytics.IbltFactoryaggregation factory. - Added new IBTL tensor encoder/decoder libraries and uses them in
tff.analytics.heavy_hitters.iblt.build_iblt_computation.
- Added new
tff.aggregator- Added
as_weighted_aggregatorto thetff.aggregator.FactoryAPI.
- Added
tff.learning- Improved compilation and execution performance of
tff.learning.metrics.secure_sum_then_finalizeby grouping tensors by DType. - Added
set_model_weightsmethod and default implementation totff.learning.templates.LearningProcess. - Added a new
reset_metricsattribute totff.learning.Model. - Added
schedule_learning_ratetotff.learning.optimizers. - Added new
tff.learning.ddp_secure_aggregatorfor Distributed Differential Privacy.
- Improved compilation and execution performance of
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.
- Removed support for Python 3.7 and 3.8, TFF supports 3.9 and later.
- Removed deprecated attributes
report_local_outputsandfederated_output_computationfromtff.learning.Model - Removed the
ingestmethod fromtff.Context
- Multiple typos in tests, code comments, and pydoc.
- Sequences (datasets) of SparseTensors don't work on the C++ runtime.
- Computations when
CLIENTScardinality is zero doesn't work on the Python runtime. - Assigning variables to a Keras model after construction inside a
model_fnresults in a non-deterministic graph.
- Added
tff.programAPI; 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.analyticsAPI for federated analytics, including private heavy hitters algorithms. - Added
tff.learning.algorithmsAPI, including TFF implementations of FedProx, FedAvg with learning rate scheduling, federated k-Means, and MimeLite. - Added
tff.learning.metricsAPI to support easy configuration of cross-client metrics aggregation via the newmetrics_aggregatorargument. - Added
metrics_aggregatorargument totff.learning.build_federated_averaging_processandtff.learning.build_federated_evaluation. - Added
report_local_unfinalized_metricsandmetric_finalizersmethods totff.learning.Modeland deprecatedreport_local_outputsandfederated_output_computation. - Added
tff.learning.optimizersAPI for building purely functional optimizers and implementations of SGD, Adagrad, Rmsprop, Adam, Yogi, - Added
tff.learning.reconstructionAPI for building partially local federated learning algorithms, including Federated Reconstruction. - Added
tff.learning.templatesAPI to support building learning algorithms in a modular fashion. - Added
tff.simulation.baselinesAPI to support evaluating learning algorithms on a suite of representative tasks. - Added
tff.aggregators.DifferentiallyPrivateFactory.tree_aggregationto support the DP-FTRL algorithm. - Added
tff.aggrgators.SecureModularSumFactory - Added
tff.aggregators.DiscreteFourierTransformFactoryandtff.aggregators.HadamardTransformFactoryto support rotation-based aggregators. - Added
tff.aggregators.concat_factoryfor 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.ClientDatasubclasses. - Added
tff.simulation.datasets.ClientData.serializable_dataset_fnattribute, enabling dataset creation within TF/TFF computations. - Added
debug_measurementsoption to aggregators intff.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_zipto invocation of user-defined TFF federated computations. - Added utilities to
tff.simulation.datasetsfor saving federated datasets to a SQL database compatible withtff.simulation.datasets.SqlClientData. - Added
tff.learning.models.FunctionalModelandtff.learning.models.functional_model_from_keras. - Increased max flow of tensors. Tensors now flow here, there, and everywhere.
- Updated the Python dependencies:
- Updated
absl-pyto version1.0.0. - Updated
attrsto version21.2.0. - Added
farmhashpyversion0.4.0. - Updated
jaxto version0.2.27. - Updated
jaxlibto version0.1.76. - Updated
numpyto version1.21.4. - Removed
retrying. - Updated
tensorflow-model-optimizationto version0.7.1. - Updated
tensorflow-model-optimizationto version0.7.3. - Updated
tensorflowto version2.8.0. - Added support for building many dependencies including
tensorflowusing Bazel. - Updated the Bazel dependencies:
- Updated
rules_pythonto version0.5.0. - Updated
com_google_protobufto versionv3.18.0-rc1. - Added
absl_pyversion1.0.0. - Added
com_google_googletestversionrelease-1.11.0. - Added
io_bazel_rules_goversionv0.29.0. - Added
bazel_skylibversion1.0.3. - Added
pybind11_abseil. - Added
pybind11_bazel. - Added
pybind11_protobuf. - Added
com_google_abslversion20211102.0. - Added
tensorflow_orgversionv2.8.0.
- Removed support for building source on MacOS.
- Removed support for Python 3.6.
- Removed symbol
tff.framework.type_contains, usetff.types.containsinstead. - Removed many symbols from
tff.simulation, these can be found intff.programinstead. - Removed support for converting non-OrderedDict mapping types to
tff.Values. - Removed
tff.simulation.datasets.ClientData.from_clients_and_fnin favor oftff.simulation.datasets.ClientData.from_clients_and_tf_fn. - Restricted
tff.simulation.datasets.ClientData.preprocessto only support TF-serializable functions. - Removed
tff.backends.reference, and the reference context it contained. - Removed
tff.learning.build_federated_sgd_processin favor oftff.learning.algorithms.build_fed_sgd. - Removed
tff.simulation.run_simulationin favor oftff.simulation.run_training_process. - Removed
tff.learning.framework.EnhancedModel. - Removed
tff.learning.framework.build_stateless_mean.
- 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.
- Introduced new intrinsics:
federated_selectandfederated_secure_select. - New
tff.structure_from_tensor_type_treeto help manipulate structures oftff.TensorTypeinto structures of values. - Many new
tff.aggregatorsfactory implementations. - Introduced
tf.dataconcept for data URIs. - New
tff.typepackage with utilities for working withtff.Typevalues. - Initial experimental support for
tff.jax_computation. - Extend
tff.tf_computationsupport toSpareTensorandRaggedTensor.
- Update gRPC dependency to 1.34.
- Moved
ClientDatainterface and implementations totff.simulation.datasets. - Renamed
tff.utils.update_statetotff.structure.update_struct. - Removed the
tff.utilsnamespace, all symbols have migrated, many totff.aggregators. - Moved infinite EMNIST dataset to federated research repository.
- Removes
rpc_modeargument to remote executors, along with streaming mode. - Removes deprecated
tff.federated_apply. - Removes
tff.federated_reduce, all usages can usetff.federated_aggregate. - Removes
HDF5ClientDataandh5pypip dependency. - Removes
setattrfunctionality ontff.ValueImpl.
- Improved
tf.GraphDefcomparisons. - 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_clientsconsuming unreasonable amounts of memory/compute time.
- Extended the
tff.simulationpackage to add many new tools for running simulations (checkpoints and metrics managers, client sampling functions). - Extended the
tff.aggregatorspackage with a number of new aggregation factories. - Added the
tff.structureAPI to expose theStructclass and related functions. - Added the
tff.profilerAPI to expose useful profiling related functions. - Added the
tff.backends.testpackage to expose backends that focused on testing specifically a way to test a computation with afederated_secure_sumintrinsic. - Added the
tff.experimentalpackage to expose less stable API.
- Replaced the
tff.aggregators.AggregationProcessFactoryabstract base class with thetff.aggregators.UnweightedAggregationFactoryand thetff.aggregators.WeightedAggregationFactoryclasses. - Replaced the
tff.aggregators.ZeroingFactoryclass with atff.aggregators.zeroing_factoryfunction with the same input arguments. - Replaced the
tff.aggregators.ClippingFactoryclass with atff.aggregators.clipping_factoryfunction with the same input arguments. - Updated
tensorflowpackage dependency to2.4.0. - Updated
absl-pypackage dependency to0.10. - Updated
grpciopackage dependency to1.32.0. - Added a
jaxlibpackage dependency at0.1.55. - Updated
numpypackage dependency to1.19.2. - Updated
tensorflow-addonspackage dependency to0.12.0. - Updated
tensorflow-model-optimizationpackage dependency to0.5.0.
- Fixed issue with the
sequence_reduceintrinsic handling federated types.
- New
tff.aggregatorspackage with interfaces for stateful aggregation compositions. - New Google Landmark Dataset
tff.simulations.dataset.gldv2 - New convenience APIs
tff.type_clientsandtff.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_processAPI to move execution of client dataset construction to executor stack leaves. - Extend parameterization of
tff.learning.build_federated_averaging_processwithuse_experimental_simulation_loopargument to better utilize multi-GPU setups.
- Removed
tff.utils.StatefulFn, replaced bytff.templates.MeasuredProcess. - Removed
tff.learning.assign_weights_to_keras_model - Stop removing
OptimizeDatasetops fromtff.tf_computations. - The
research/directory has been moved to http://github.com/google-research/federated. - Updates to
input_specargument fortff.learning.from_keras_model. - Updated TensorFlow dependency to
2.3.0. - Updated TensorFlow Model Optimization dependency to
0.4.0.
- Fixed streaming mode hang in remote executor.
- Wrap
collections.namedtuple._asdictcalls incollections.OrderedDictto support Python 3.8. - Correctly serialize/deserialize
tff.TensorTypewith 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.
- "Federated Learning for Image Classification" tutorial fails to load
projectorplugin 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).
This release contains contributions from many people at Google, as well as:
amitport, ronaldseoh
- Fixed issue preventing Python
lists from beingall_equalvalues.
- Mirrored user-provided types and minimize usage of
AnonymousTuple.
- Renamed
AnonymousTupletoStruct.
- Updated
tensorflow-addonspackage dependency to0.9.0. - Added API to expose the native backend more conveniently. See
tff.backends.native.*for more information. - Added a compiler argument to the
tff.framework.ExecutionContextAPI and provided a compiler for the native execution environment, which improves TFF’s default concurrency pattern. - Introduced a new
tff.templates.MeasuredProcessconcept, a specialization oftff.templates.IterativeProcess. - Extends
tff.learninginterfaces to accepttff.templates.MeasuredProcessobjects 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 thetff.framework.FederatingExecutorso 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.
- The
IterativeProcessreturn fromtff.learning.build_federated_averaging_processandtff.learning.build_federated_sgd_processnow 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_executorfunction, instead you should use the more convenienttff.backends.native.set_local_execution_contextfunction or manually construct a context an set it usingtff.framework.set_default_context. - The
tff.Computationbase class now contains an abstract__hash__method, to ensure compilation results can be cached. Any custom implementations of this interface should be updated accordingly.
- 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_computationdecorated function used no variables.
This release contains contributions from many people at Google, as well as:
jvmcns@
- Multiple TFF execution speedups.
- New
tff.templates.MeasuredProcessspecialization ofIterativeProcess. - Increased optimization of the
tff.templates.IterativeProcess->tff.backends.mapreduce.CanonicalFormcompiler.
- Moved
tff.utils.IterativeProcesstotff.templates.IterativeProcess. - Removed
tff.learning.TrainableModel, client optimizers are now arguments on thetff.learning.build_federated_averaging_process. - Bump required version of pip packages for tensorflow (2.2), numpy (1.18), pandas (0.24), grpcio (1.29).
- Issue with GPUs in multimachine simulations not being utilized, and bug on deserializing datasets with GPU-backed runtime.
- TensorFlow lookup table initialization failures.
- 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.
This release contains contributions from many people at Google, as well as:
jvmcns@
- Fixed issues in tutorial notebooks.
- Updated
absl-pypackage dependency to0.9.0. - Updated
h5pypackage dependency to2.8.0. - Updated
numpypackage dependency to1.17.5. - Updated
tensorflow-privacypackage dependency to0.2.2.
- Deprecated
dummy_batchparameter of thetff.learning.from_keras_modelfunction.
- Fixed issues with executor service using old executor API.
- Fixed issues with remote executor test using old executor API.
- Fixed issues in tutorial notebooks.
- Upgraded tensorflow dependency from
2.0.0to2.1.0. - Upgraded tensorflow-addons dependency from
0.6.0to0.7.0. - Upgraded attr dependency from
18.2to19.3. - Upgraded tfmot dependency from
0.1.3to0.2.1. - 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_evalfor evaluating model personalization strategies. - Added new federated intrinsic
tff.federated_secure_sum. tff.learning.build_federated_averaing_process()now takes aclient_optimizer_fnand atff.learning.Model.tff.learning.TrainableModelis now deprecated.- Improved performance in the high performance executor stack.
- Implemented and exposed
tff.framework.ExecutorFactory; alltff.framework...executor_factorycalls now return an instance of this class. - Added
remote_executor_examplebinary 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_zipon inputs to other federated intrinsics.
- Dropped support for Python2.
- Renamed
tff.framework.create_local_executor(and similar methods) totff.framework.local_executor_factory. - Deprecated
federated_apply(), instead usefederated_map()for all placements.
- Fixed problem with different instances of the same model having different
named types.
tff.learning.ModelWeightsno longer names the tuple fields returned for model weights, instead relying on an ordered list. tff.sequence_*on unplaced types now correctly returns atff.Value.
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.
- Python 2 support is now deprecated and will be removed in a future release.
federated_mapnow works with bothtff.SERVERandtff.CLIENTplacements.federated_zipreceived significant performance improvements and now works recursively.- Added retry logic to gRPC calls in the execution stack.
collections.OrderedDictis now required in many places rather than standard Python dictionaries.
- Fixed computation of the number of examples when Keras is using multiple inputs.
- Fixed an assumption that
tff.framework.Tupleis returned fromIterativeProcess.next. - Fixed argument packing in polymorphic invocations on the new executor API.
- Fixed support for
dir()inValueImpl. - Fixed a number of issues in the Colab / Jupyter notebook tutorials.
- Updated to use
grpcio1.24.3.
- Add a
federated_sampleaggregation that is used to collect a sample of client values on the server using reservoir sampling. - Updated to use
tensorflow2.0.0andtensorflow-addons0.6.0instead of the coorisponding nightly package in thesetup.pyfor releasing TFF Python packages. - Updated to use
tensorflow-privacy0.2.0. - Added support for
attr.sclasses type annotations. - Updated streaming
Executemethod ontff.framework.ExecutorServiceto be asynchronous. - PY2 and PY3 compatability.
- 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.
examplespackage split tosimulationandresearch.
- Various error message string improvements.
- Dataset serialization fixed for V1/V2 datasets.
tff.federated_aggregatesupportsaccumulate,mergeandreportmethods with signatures containing tensors with undefined dimensions.
- 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.
- Removed support for the deprecated
tf.data.Datasetstring iterator handle. - Bumps the required versions of grpcio and tf-nightly.
- Fixes in notebooks, typos, etc.
- Assorted fixes to align with TF 2.0.
- Fixes thread cleanup on process exit in the high-performance executor.
This release contains contributions from many people at Google, as well as:
Gui-U@, Krishna Pillutla, Sergii Khomenko.
- High-performance simulation components and tutorials.
- Refactoring/consolidation in utility functions in tff.framework.
- Switches some of the tutorials to new PY3-only executor stack components.
- Includes the
examplesdirectory in the pip package. - Pip installs for TensorFlow and TFF in turorials.
- Patches for asyncio in tutorials for use in Jupyter notebooks.
- Python 3 compatibility issues.
- Support for
federated_map_all_equalin 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.
This release contains contributions from many people at Google.
- Support for multiple outputs and loss functions in
kerasmodels. - Support for stateful broadcast and aggregation functions in federated averaging and federated SGD APIs.
tff.utils.update_stateextended to handle more generalstatearguments.- Addition of
tff.utils.federated_minandtff.utils.federated_max. - Shuffle
client_idsincreate_tf_dataset_from_all_clientsby default to aid optimization.
- Dependencies added to
requirements.txt; in particular,grpcioandportpicker.
- Removes dependency on
tf.data.experimental.NestedStructure.
This release contains contributions from many people at Google, as well as:
Dheeraj R Reddy, @Squadrick.
- Removed source level TF dependencies and switched from
tensorflowtotf-nightlydependency. - Add support for
attrmodule in TFF type system. - Introduced new
tff.frameworkinterface layer. - New AST transformations and optimizations.
- Preserve Python container usage in
tff.tf_computation.
- Updated TFF model to reflect Keras
tf.keras.model.weightsorder. - Keras model with multiple inputs. #416
- New
tff.simulation.TransformingClientDataAPI and associated inifinite EMNIST dataset (see http://tensorflow.org/federated/api_docs/python/tff for details)
- Normalized
functofnacross the repository (rename some parameters and functions)
- 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
- Rename
tff.learning.federated_averagetotff.learning.federated_mean. - Rename 'func' arguments to 'fn' throughout the API.
- Assorted fixes to typos in documentation and setup scripts.
- Updated to use TensorFlow version 1.13.1.
- Implemented Federated SGD in
tff.learning.build_federated_sgd_process().
next()function oftff.utils.IteratedProcesss returned bybuild_federated_*_process()no longer unwraps single value tuples (always returns a tuple).
- 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 bybuild_federated_*_process(). - Clear cached Keras sessions when wrapping Keras models to avoid referencing stale graphs.
Initial public release.