PiperOrigin-RevId: 388752613

PiperOrigin-RevId: 388490972

PiperOrigin-RevId: 388483415

Clarifies the relationship between the batch_argument argument to predict_on_batch and forward_pass, and the property input_spec.

PiperOrigin-RevId: 388318403

PiperOrigin-RevId: 388268647

This implements new `tff.learning.models.save` and `tff.learning.models.load`
APIs that use TensorFlow's SavedModel format for serializing models for
portability.

This has a side-effect of also making the models easily convertible for TFLite
flatbuffers for inference deployments.

PiperOrigin-RevId: 388250327

The `select_output_from_lambda` function previously had
surprisingly different semantics on outer-tuples from
nested tuples: outermost tuples or lists indicated a
structure of results, while inner tuples or lists
indicated paths to elements. No distinction was made
between tuples and lists.

This CL changes to use tuples for multi-index
selection paths and lists for selecting multiple
different values. This simplifies some call sites.

Includes test cleanup to remove outdated references
to "Federated" and to test the new selection abilities.

PiperOrigin-RevId: 388160914

PiperOrigin-RevId: 387933152

PiperOrigin-RevId: 387923093

PiperOrigin-RevId: 387831186

PiperOrigin-RevId: 387815431

It is still a good idea to signal that these targets intentionally do not support Python 2.

PiperOrigin-RevId: 387673807

PiperOrigin-RevId: 387633452

PiperOrigin-RevId: 387466843

These value indices were fragile and would all require replacement
upon the addition of new intrinsics to MapReduceForm. This change
provides more descriptive names and stablilizes the selection of
values to be resilient to new intrinsic additions.

This change is designed to make it eaiser to add MapReduceForm
support for federated_secure_sum and federated_secure_modular_sum.

PiperOrigin-RevId: 387458471

PiperOrigin-RevId: 387380625

PiperOrigin-RevId: 387240061

PiperOrigin-RevId: 387229169

PiperOrigin-RevId: 387222252

PiperOrigin-RevId: 387209652

Minor fix of mismatch in code and description of `building_your_own_federated_leanring_algorithm` tutorial.

PiperOrigin-RevId: 387203200

Previously, the tensor specs provided to the optimizer were for the entire model weights, while the expectation is that it shuold manipulate only trainable model weights.

This also restructures the code to make sure optimizer does not work with lists but tuples instead, as if the optimizer state contains list of tensors, the "for _ in data:" pattern will try to stack all of the tensors, raising error if the tensors to not have the same rank - e.g. when they include bias and kernel of a layer being optimzied.

PiperOrigin-RevId: 387019484

PiperOrigin-RevId: 386911399

PiperOrigin-RevId: 386870506

PiperOrigin-RevId: 386647579

Add helper function to compare TFF and Keras optimizer in optimizer tests. Modify SGD to accept `momentum=0` to be easily used in parameterized tests.

PiperOrigin-RevId: 386586370

PiperOrigin-RevId: 386573163

PiperOrigin-RevId: 386557393

Regex used: "u?int[0123456789]+[^_] "

PiperOrigin-RevId: 386550153

PiperOrigin-RevId: 386548020

PiperOrigin-RevId: 386546510

After some discussion internal to the team, we decided that the value of this function and the checks it adds is not worth the extra maintenance burden as we modify MapReduceForm to support new intrinsics (including `federated_secure_sum` with `min_input`, `federated_secure_modular_sum`, and other primitive federated aggregations).

PiperOrigin-RevId: 386544630

PiperOrigin-RevId: 386533240

PiperOrigin-RevId: 386503038

PiperOrigin-RevId: 386407218

PiperOrigin-RevId: 386339610

This now uniquifies the first graph during composition (previously the code
didn't). A new unittest has been added which catches this, showing that nested
compisitions could result in duplicate shared names if composed in a particular
order.

PiperOrigin-RevId: 386237260

PiperOrigin-RevId: 386061700

Update tff.learning reconstruction training_process to accept TFF optimizer abstraction. Also make TFF SGD accept 0 learning rate to be used in construction test.

PiperOrigin-RevId: 385926783

This change is the first towards leveraging the ability to split some TFF computations into subrounds to run logical rounds at much larger scale than is currently feasible.

PiperOrigin-RevId: 385894645