PiperOrigin-RevId: 374300438

`asyncio.Task.current_task()` was deprecated in python 3.7 and removed in 3.9.
This was replaced by `asyncio.current_task()`.
https://docs.python.org/3.8/library/asyncio-task.html#asyncio.Task.current_task

PiperOrigin-RevId: 374223589

https://github.com/tensorflow/tensorflow/blob/v2.5.0/tensorflow/tools/pip_package/setup.py#L130

Fix the nightly package to depend on gRPC 1.37 which was updated recently.

PiperOrigin-RevId: 374208257

PiperOrigin-RevId: 374194494

PiperOrigin-RevId: 373895854

Previously, computations could return and correctly
infer the types of sparse and ragged tensors. However,
there was previously no support for interpreting
SparseTensorSpecs or RaggedTensorSpecs. This CL adds
that support as well as a couple tests ensuring
end-to-end functioning of these types with TFF.

PiperOrigin-RevId: 373882381

Add a `latest_round_num` property to the CSVManager, for the purposes of being able to resume logging metrics in the event of a failuire.

PiperOrigin-RevId: 373670178

`Tensorflow` proto message.

This is a first step into turning on static garppler for the native backend compiler pipeline.

PiperOrigin-RevId: 373651610

PiperOrigin-RevId: 373650738

* Removed `tff.utils`, this package has been removed.
* Added `tff.analytics`, this package has been added.
* Removed detailed `tff.backends` items, these are intended to grow and therefore don't scale/maintain well and they all should have the same set of dependencies.
* Updated the `tff.simulation` to make it clear that the simulation package can be used in all types of TFF simulations.

PiperOrigin-RevId: 373600277

This has presented as an issue now that the interning dict lives for the entire life of the program.

PiperOrigin-RevId: 373462906

This change removes the WeakValueDictionary which was causing weird behavior
where lookups would occur after parts of the Python standard library had
already destroyed themselves. `atexit` hooks to clear the dictionary appeared
to fix this at first, but it regressed again (for unknown reasons, this time
on a smaller subset of programs).

TFF types themselves are not particularly large in memory, and are frequently
long-lived anyways (many are in global-scope objects), so the longer lifetimes
are likely not a significant downside.

PiperOrigin-RevId: 373052202

PiperOrigin-RevId: 372951650

These tests make sure that compiler does not compile the aggregator in a way that is not expected, for instance by accidentally duplicating some values to be aggregated.

PiperOrigin-RevId: 372497721

constant or tensor.

PiperOrigin-RevId: 372404564

PiperOrigin-RevId: 372216575

PiperOrigin-RevId: 372023705

PiperOrigin-RevId: 372021667

PiperOrigin-RevId: 371942693

This function is only a small wrapper around the more flexible `tff.simulation.build_uniform_sampling_fn`, which should be used instead.

PiperOrigin-RevId: 371815302

Change the result of `tff.simulation.build_uniform_sampling_fn` to additionally be paratermized by `size`.

This allows `functools.partial(fn, size=X)` to replicate previous behavior,
while also allowing new usages with variable sizes.

PiperOrigin-RevId: 371797271

TransformingClientData now works only with transformation functions that can be serialized as tf.computation.

Also more general and less cumbersome method to split clients into pseudoclients: the user can specify a function to generate a list of pseudoclient ids for each client id. This allows, for example, differing numbers of pseudoclients for different clients.

Also updates infinite_emnist to work with the new transforming_client_data.

PiperOrigin-RevId: 371748160

PiperOrigin-RevId: 371420165

PiperOrigin-RevId: 371258358

This change also fixes a bug where TestClientData could modify its input arguments in-place.

PiperOrigin-RevId: 371161350

PiperOrigin-RevId: 371132404

This change fixes a previous bug that fixed these to be the same.

PiperOrigin-RevId: 370984866

PiperOrigin-RevId: 370934109

PiperOrigin-RevId: 370797360

PiperOrigin-RevId: 370785992

Implement a transformation that will set `disable_meta_optimizer = True` on all PartitionedCall ops in a TensorFlow graph.

TFF graphs are typically short-lived (emphemeral client participation each round),
which means the costs of graph optimized are unable to amortize over the short
usage.

PiperOrigin-RevId: 370783748

Previously, these caches would sometimes get deleted after Python had already
cleared away modules from the standard library, resulting in all manner of
nonsense. Deleting these with an `atexit` hook ensures that they are cleaned up
before the standard library starts deleting itself.

PiperOrigin-RevId: 370743951

Updates all TFF internal callsites to use default_num_clients instead of num_clients, in preparation for parameter name change.

PiperOrigin-RevId: 370690092

PiperOrigin-RevId: 370596723

PiperOrigin-RevId: 370576548

PiperOrigin-RevId: 370572743

PiperOrigin-RevId: 370568025

Add unittest predicates to prevent (and ultimately try to reduce) call `model_fn` multiple times in tff.learning`.

Some `model_fn` can be expensive and we should avoid rebuiling the model many times.

PiperOrigin-RevId: 370566132

PiperOrigin-RevId: 370565395

Adds Skip(0) and some small test disabling to avoid segfault on GPU runtimes when directly invoking tf.functions.

PiperOrigin-RevId: 370525961