PiperOrigin-RevId: 411172320

This is a method to create a `tff.learning.models.FunctionalModel` from a `tf.keras.Model`.

PiperOrigin-RevId: 411169398

The CachingExecutor formerly played a critical role in the TFF execution stack, as it prevented multiple execution of the myriad duplicate computations in TFF's ASTs. Nowadays, TFF no longer generates these types of duplicate computations, and it is no longer to cache at this level to ensure performance.  Additionally, this type of caching plays poorly with nondeterministic computations.

Prior to this CL, the CachingExecutor was already unused outside of tests. This should be a no-op for TFF customers.

PiperOrigin-RevId: 411112532

PiperOrigin-RevId: 411090036

This change should enable the MacOS presubmit to succeed.

PiperOrigin-RevId: 410926931

In the tff.program library, there is usage of various `flatten` functions:

* `tf.nest.flatten`
* `tree.flatten`
* `structure_utils.flatten`, which is a wrapper around `tree.flatten_with_path`

This change does two things:

1. Consolidates usage of various `flatten` functions to the `tree` package, `tree` is used over `tf.nest` because it has more functionality.
2. Renames `structure_utils.flatten` to `structure_utils.flatten_with_name` to disambiguate this functionality.

PiperOrigin-RevId: 410910920

PiperOrigin-RevId: 410894245

PiperOrigin-RevId: 410631896

PiperOrigin-RevId: 410595976

PiperOrigin-RevId: 410587320

PiperOrigin-RevId: 410538549

PiperOrigin-RevId: 410538498

Now uses dtype.name to match known-rank case (used dtype previously), and adjusts from [None] to (shape=None).

This will change the string representation e.g. of TensorType(tf.float32, shape=None) from:

tf.float32[None]

to

float32(shape=None).

For reference, a TensorType(tf.float32, shape=[None]) (IE, rank-1 tensor of unknown size) will remain represnted as:

float32[?].

PiperOrigin-RevId: 410428097

This copied from TF's .bazelrc.

PiperOrigin-RevId: 410381396

PiperOrigin-RevId: 410361295

PiperOrigin-RevId: 410350199

PiperOrigin-RevId: 410327209

PiperOrigin-RevId: 410312882

PiperOrigin-RevId: 410271764

PiperOrigin-RevId: 410254008

PiperOrigin-RevId: 409593450

This will be used as the default metric aggregator for most iterative process/eval computation builders provided by TFF learning.

PiperOrigin-RevId: 409444231

- Moves a generic check to py_typecheck.py or as private module level helpers specific to concrete optimizers.
- Corrects structural checks on weights and gradients, and adds them to public API.

PiperOrigin-RevId: 409418948

PiperOrigin-RevId: 409407858

PiperOrigin-RevId: 409155243

* Added unit tests.
* Removed sorting the key of the returned value when it is flattened.
* Updated type annotations.

PiperOrigin-RevId: 408970683

PiperOrigin-RevId: 408917062

- Remove unused pybind_protobuf wrappings for methods without protocol buffers.
- Fix header only dep to include implementation.
- Update bazelrc to avoid ubp compilation error.

PiperOrigin-RevId: 408671302

PiperOrigin-RevId: 408518267

PiperOrigin-RevId: 408462877

* Added documentation about how value references are materialized.
* Added documentation about how values and program state is flattened or unflattend.
* Added documentation about the format of the released values and saved program state.

PiperOrigin-RevId: 408459647

* Added `CSVFileReleaseManager`.
* Added tests.
* Updated datatypes for all tests in the `program` package.

PiperOrigin-RevId: 408384664

First of a stream of CLs with boilerplate work for sequence executor.

PiperOrigin-RevId: 407896764

Note that this does not parallelize materializing or merging of child results: those
steps are still serialized.

PiperOrigin-RevId: 407878128

As an aside, this CL adds a new `main` function for tests running in OSS.
This was necessary in order to use absl flags in OSS C++ tests because OSS
gtest_main intercepts argc and argv but does not parse absl flags. For
details, see https://github.com/google/googletest/issues/3646.

PiperOrigin-RevId: 407875061

PiperOrigin-RevId: 407853543

Previously, failure to make `ExecutorValue` instances copy-constructible would result in confusing template instantiation errors. This `static_assert` makes the
requirement explicit and provides the user with a clear, actionable suggestion.

Example error:

```
ERROR: third_party/tensorflow_federated/cc/core/impl/executors/executor.h:230:3 static_assert failed due to requirement 'std::is_copy_constructible<std::unique_ptr<int, std::default_delete<int>>>::value' "`ExecutorValue`s (the type parameter passed to `ExecutorBase`) must be copy-constructible. Consider wrapping with `std::shared_ptr`, `std::shared_future`, or other similar container like: `class MyExecutor: public ExecutorBase<std::shared_ptr<MyExecutorValue>>`"
  static_assert(std::is_copy_constructible<ExecutorValue>::value,
  ^             ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
third_party/tensorflow_federated/cc/core/impl/executors/executor.h:356:26 in instantiation of template class 'tensorflow_federated::ExecutorBase<std::unique_ptr<int>>' requested here
class MyExecutor: public ExecutorBase<std::unique_ptr<int>> {
                         ^
```
PiperOrigin-RevId: 407853500

PiperOrigin-RevId: 407830629

Without this, some compilers complain with:
```
In file included from tensorflow_federated/cc/core/impl/executors/composing_executor_test.cc:31:
./tensorflow_federated/cc/core/impl/executors/value_test_utils.h:83:22: error: call to constructor of 'tensorflow::Tensor' is ambiguous
  tensorflow::Tensor tensor(tensor_constructor_args...);
                     ^      ~~~~~~~~~~~~~~~~~~~~~~~
tensorflow_federated/cc/core/impl/executors/composing_executor_test.cc:510:26: note: in instantiation of function template specialization 'tensorflow_federated::testing::TensorV<unsigned long>' requested here
      v0::Value result = TensorV(client_results.size());
                         ^
```

PiperOrigin-RevId: 407829908

PiperOrigin-RevId: 407826333