PiperOrigin-RevId: 275079302

PiperOrigin-RevId: 275073766

PiperOrigin-RevId: 275055825

In the notation:

  <weights=float32[10,5],bias[5]>@server

'bias' appears to be the name component for a tensor element of a named tuple. [5] describes the shape, but the dtype is missing.

PiperOrigin-RevId: 275054396

PiperOrigin-RevId: 274912142

PiperOrigin-RevId: 274909205

This is problematic when an experiment is resumed from a previous checkpoint,
since the state of the PRNG is not properly saved/restored. This is tricky to
get right across all libraries (numpy, tensorflow, python); so avoid setting
seeds for now.

PiperOrigin-RevId: 274900765

PiperOrigin-RevId: 274896939

PiperOrigin-RevId: 274883244

This was optional, is causing analysis errors, and could cause a exception to be raised.

PiperOrigin-RevId: 274881555

PiperOrigin-RevId: 274871205

PiperOrigin-RevId: 274846444

PiperOrigin-RevId: 274828751

Before:
  The supplied function with argspec SimpleArgSpec(args=[], varargs=[], kwargs=[], defaults=[]) cannot accept a value of type string* single argument or multiple positional and/or keyword arguments.

After:
  The supplied function "my_tf_fn" with argspec () cannot accept a value of type string* as either a single argument or multiple positional and/or keyword arguments.
PiperOrigin-RevId: 274814269

PiperOrigin-RevId: 274714639

PiperOrigin-RevId: 274674077

PiperOrigin-RevId: 274673083

PiperOrigin-RevId: 274669504

* Added CheckpointManager API.
* Added FileCehckpointManager implementation.
* Updated an experiment to:
  * Use the new CheckpointManager objects.
  * Separate evaluation from the metrics consumer/hook.
  * Not store metrics in the checkpoint.

TODO: Update the remaining experiments to use the checkpoint manager.
TODO: Remove checkpoint utils.
PiperOrigin-RevId: 274620556

PiperOrigin-RevId: 274618323

PiperOrigin-RevId: 274615965

Previously, to_type only supported passing attr.s class instances with type information embedded into the instance fields (i.e. tff.to_type(MyAttrsClass(my_field=tf.int32))). This change extends support to also accept attr.s class type objects when the field types have been defined, i.e:

  @attr.s
  class MyAttrsClass(object):
    my_field = attr.ib(type=tf.int32)

  tff.to_type(MyAttrsClass)

PiperOrigin-RevId: 274598038

PiperOrigin-RevId: 274262570

Removes duplication tf data dataset representations, going instead through public constant in tensorflow_utils.

PiperOrigin-RevId: 274221123

* Updated the type annotations for `utils_impl`.
* Updated the build target to be of kind `pytype_strict_library`.
* Updated copybara configuration.

PiperOrigin-RevId: 274214750

This is a prerequisite for making TensorShape and Dimension more strict
about the types of their arguments.

PiperOrigin-RevId: 274176324

should preserve the percentage of both in the final sample.

PiperOrigin-RevId: 274054312

Add a federated_sample aggregation that is used to collect a sample of client values on the server using reservoir sampling.

PiperOrigin-RevId: 274022855

Adds utility function to core/utils/differential_privacy for creating DPQuery with adaptive clipping and/or per-vector clipping. Also cleans up research/baselines/emnist/run_experiment.

PiperOrigin-RevId: 273777778

PiperOrigin-RevId: 273753109

Adds a some test coverage and handles corenr cases better.

PiperOrigin-RevId: 273647845

Normally, **kwargs only accepts a dict for kwargs, but currently autograph wraps the keyword arguments in a dict call -- `dict(kwargs)` -- this is to be fixed soon. Since dict accepts both dicts and lists of key-value pairs, this effectively masked the bug.

PiperOrigin-RevId: 273636344

This follows a similar convetion to tf.kersas.Model.summary(): https://www.tensorflow.org/api_docs/python/tf/keras/Model#summary

PiperOrigin-RevId: 273609432

PiperOrigin-RevId: 273559700

PiperOrigin-RevId: 273405515

PiperOrigin-RevId: 273403022

PiperOrigin-RevId: 273383857

PiperOrigin-RevId: 273360892

PiperOrigin-RevId: 273346886

PiperOrigin-RevId: 273310053