PiperOrigin-RevId: 322178848

Update `build_federated_averaging_process` and `build_federated_sgd_process` to return the full metrics suite.

After this change, users of `tff.learning.build_federated_averaging_process`
(and the FedSGD version) will receives a named tuple of three groups of metrics: broadcast, train, and aggregation.

- `train` are the same metrics that were previously returned.
- `broadcast` and `aggregation` are metrics for the broadcast and aggregatin
  process respectively. For example, an aggregation process might report
  metrics about the client updates and the aggregated update (e.g. the global
  norms).

Previously these methods only extracted the training metrics to retain
historical behavior. This change updates the methods and callsites to return
the full metrics set (including those from the broadcast and aggregation
processes).

PiperOrigin-RevId: 317888544

* Regex search replace.
* Removed unused imports.
* Removed BUILD deps.

PiperOrigin-RevId: 314969638

PiperOrigin-RevId: 313410789

Replace `import tensorflow.compat.v2 as tf` with `import tensorflow as tf` in the `learning` module.

PiperOrigin-RevId: 310572253

PiperOrigin-RevId: 310377468

Step 1 of 3.

PiperOrigin-RevId: 297684545

-   Removes `tff.learning.from_compiled_keras_model()` from API.
-   Removes unused `tff.learning._TrainableKerasModel`.
-   Removes usage of `tf.keras.Model.output_names` and
    `tf.keras.Model.loss_functions` in `tff.learning._KerasModel` to be
    compliant with upcoming Keras refactoring.

As a side-effect of this change, TFF will no longer supported (though no known
usages):

1.   Specifying multiple loss functions or loss weights via dictionaries
keyed by model output name. Instead, using ordered sequences of
functions mapping to outputs is still possible.
2.   Sublcassed `tf.keras.Model` instances cannot be used in TFF.

PiperOrigin-RevId: 295824509

* This is required for parameterized test to work because decorators are parsed before the main function is called, if there is tf code in the parameters it may be not obvious why things are failing.

PiperOrigin-RevId: 294983486

PiperOrigin-RevId: 285801872

PiperOrigin-RevId: 285116495

de-duplication strategy in non-eager mode.

TF, under certain circumstances, prepends `_#` numbers to variables with the
same name. This appears to happen in non-eager mode and in some occasions with
Keras. TFF previously used these names for the type signature, which would
cause problems when the same model was created more than once in the same graph
context.

Instead, rely on the ordering of the sequence of variables contained in the
model and use the unnamed structure for types.

PiperOrigin-RevId: 285032395

PiperOrigin-RevId: 280784907

PiperOrigin-RevId: 280767713

- Replace `dict` with `collections.OrderedDict` when used in `tf.data.Dataset`
- Convert tests to eager mode.

PiperOrigin-RevId: 279172544

PiperOrigin-RevId: 266383040

PiperOrigin-RevId: 263172828

PiperOrigin-RevId: 258830309

Change calls from `tensor.numpy()` to `self.evaluate(tensor)`, which is compatible in graph and eager mode.

Additionally tighten up the assertions on nested structures in a few unittests.

PiperOrigin-RevId: 251246960

PiperOrigin-RevId: 250977892

python/learning.

PiperOrigin-RevId: 250908176

PiperOrigin-RevId: 249666998

PiperOrigin-RevId: 244269098

- Annotate many tests to be wrapped in a graph.
- Avoid comparison on the string representation of the Tensor.
- Do not use self.session() provided by the testcase, rather construct a new
  tf.Session().  In eager mode, the returned object is not `tf.Session`, it is
  type `EagerSessionWarner` which fails tff.pytype_check().
- Normalize on tf.test.main() [for pure Tensorflow tests) or test.main() [for TFF + TF tests]
- Use the forward compatible tf.compat.v1.enable_v2_behavior() method.

PiperOrigin-RevId: 244033806

Support state in the function used to do the federated_broadcast. This could allow, e.g., for an adaptive quantization scheme to be used. Also removes some older type-based tests, since we now have execution tests, and the type-based tests are likely to be brittle (needing to be updated due to changes in distant files).

PiperOrigin-RevId: 241843768

Allow stateful aggregation operators in model_delta_optimizer_processes like Federated Averaging. Allowing stateful aggregation functions to be applied to model deltas paves the way for using a differentially private average such as the GaussianAverageQuery of TensorFlow Privacy (where the state includes for example the Privacy Ledger).

PiperOrigin-RevId: 241628028

- Remove variable creation in __init__, not needed
- Add back @tf.contrib.eager.function() annotations and remove tf.control_dependencies
- Reduce number of repeates in unittests so execution time is faster (2 reptitions is sufficient)
- Replace EagerTensor.numpy() calls with tf.TestCase.evaluate() for extracting tensor values in unittests

PiperOrigin-RevId: 241365351

* Rename `func` to `fn` for variables.
* Rename `func` to `function` for module name.

PiperOrigin-RevId: 240143663

PiperOrigin-RevId: 238046002

PiperOrigin-RevId: 237892686

Breakage in colabs caused by shifting the API surface and thus being out of sync between pip package and head. Reverting changes to learning directory and ipynb, keeping them in core.

PiperOrigin-RevId: 237058456

In particular, federated_zip will now simply promote the federation of a 1-tuple of federated values, returning a federated 1-tuple of values.

WARNING: This CL is backwards-incompatible.
Now an IterativeProcess built through `build_federated_optimizer_process` will always return a *tuple* of metrics as the second return value of its next method, as opposed to returning a tuple if there are 2 or more metrics specified and a scalar if there is only 1.
PiperOrigin-RevId: 236763176

PiperOrigin-RevId: 236402517

* Moved test functions into one location.
* Renamed test functions and module to be more concisely named.
* Moved TF 2.0 setup from setup methods to main function.

PiperOrigin-RevId: 230776467

PiperOrigin-RevId: 230591634

* Added missing BUILD dependencies.
* Added six python dependencies.
* Fixed whitespace.

PiperOrigin-RevId: 227725720

Updated FedAvg and FedSgd to ignore non-finite updates and to calculate a client_weight to use in the federated average of deltas.

PiperOrigin-RevId: 227063793

Renaming to clarify the relationship between trainable, non-trainable, and local variables. Also resolves the issue that 'model' and 'initial_model' were different types, which was surprising. In particular:
  * initial_model => initial_weights
  * model_delta => weights_delta
  * ModelVars => ModelWeights
  * enhanced_model.vars => enhanced_model.weights

Also regenerates documentation.

PiperOrigin-RevId: 227040535

Implements the client logic for federated_sgd, and adds a common interface for client optimizers that average model deltas (FedSgd and FedAvg).  Common functionality for building federated optimizers moves to framework/optimizer_utils.py.

PiperOrigin-RevId: 227026404