Tensorflow-federated

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.

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_computation`s.
* 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

Bug Fixes

* Fixed issue preventing Python `list`s from being `all_equal` values.

Major Features and Improvements

* Mirrored user-provided types and minimize usage of `AnonymousTuple`.

Breaking Changes

* Renamed `AnonymousTuple` to `Struct`.

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

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](https://github.com/tensorflow/federated/issues/832).

Thanks to our Contributors

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

jvmcns