PiperOrigin-RevId: 382409145

We would switch hash types when saving and loading because of a config
error.

PiperOrigin-RevId: 382784449

All other "int" output from KPLs (Hashing, StringLookup, IntegerLookup and TextVectorization)
is int64, we should be consistent.

PiperOrigin-RevId: 382778202

When output_sequence_length is set, for dense tensor input, we would error
out when the static shape contains None on any dimension. We can fix that
by using the dynamic shape to calculate padding for output_sequence_length.

PiperOrigin-RevId: 382347472

PiperOrigin-RevId: 382327442

PiperOrigin-RevId: 381424814

PiperOrigin-RevId: 381369501

PiperOrigin-RevId: 381368343

PiperOrigin-RevId: 381360615

This simplifies the calling code, and ensures that we print all tags found the summary events if the assertion fails, which helps debug cases where summaries are still being logged but under a tag name that differs from the expected one.

PiperOrigin-RevId: 381352128

PiperOrigin-RevId: 381292911

PiperOrigin-RevId: 381276605

PiperOrigin-RevId: 381274594

PiperOrigin-RevId: 381160816

This change also enables models with multiple input arguments to be saved, for example:

```
class Subclass(keras.Model):

  def call(self, a, b):
    ...
```

PiperOrigin-RevId: 381148805

PiperOrigin-RevId: 381088871

PiperOrigin-RevId: 381065953

PiperOrigin-RevId: 381061383

PiperOrigin-RevId: 381059294

PiperOrigin-RevId: 381058890

PiperOrigin-RevId: 381049091

ShardedVariable: add unit test to verify return value of assign_xxx can be passed as control dependencies.

Add AUC metric test when variables are sharded.

PiperOrigin-RevId: 381046405

set `tf.expand_dims argument` `axis` to an integer instead of a list in the multi_head_attention model.

PiperOrigin-RevId: 380957913

PiperOrigin-RevId: 380952115

This was always broken for numpy vocabulary inputs, and recently broke for lists as well.

PiperOrigin-RevId: 380946490

PiperOrigin-RevId: 380936605

1. was giving an `UnboundLocal` error.
2. If `fpath` and `untar_fpath` are set to the same value the `untar` doesn't run.
So ensure that `fname` always has `.tar.gz`, and `untar_fpath` never does.

The old behavior (before the previous change) was to always add `.tar.gz`, giving `.tar.gz.tar.gz` files. That seems wrong too.

Fixes tensorflow_docs/site/en/tutorials/load_data/text.ipynb

PiperOrigin-RevId: 380934634

PiperOrigin-RevId: 380933127

PiperOrigin-RevId: 380918577

PiperOrigin-RevId: 380914431

PiperOrigin-RevId: 380914130

It is only support for Discretization and Normalization and not even tested
on those classes. Removing it gives us a cleaner API interface for release.

PiperOrigin-RevId: 380906031

We recently switched to creating our tables in an init_scope (which is
good, we never want to creat them at graph execution time), but we also
need to place the initialization ops inside the init_scope.

PiperOrigin-RevId: 380899225

The date for forward compat already take effective.

PiperOrigin-RevId: 380894653

PiperOrigin-RevId: 380894633

PiperOrigin-RevId: 380869424

PiperOrigin-RevId: 380865941

PiperOrigin-RevId: 380851544

- introduction to keras for engineers
- introduction to keras for researchers

PiperOrigin-RevId: 380842991

PiperOrigin-RevId: 380838692