Added Coursera course Machine Learning with TensorFlow on GCP

Make sure custom losses + metrics get autographed, even though autograph marks code in core Tensorflow with `DoNotConvert`

If this changes and autograph starts converting keras code, this won't be needed. (e.g. maybe after the repo split)

Fixes #37440

PiperOrigin-RevId: 309518788
Change-Id: I42ce3e208e5c2fd10c4de008d9087d089ee4c591

Make sure custom losses + metrics get autographed, even though autograph marks code in core Tensorflow with `DoNotConvert`

If this changes and autograph starts converting keras code, this won't be needed. (e.g. maybe after the repo split)

Fixes #37440

PiperOrigin-RevId: 309516574
Change-Id: Ia82ee60fed225957dec0de4d378d6cac57d7f16e

This change adds an `OpKernelContext::executor_type()` method, which (by analogy with `OpKernelContext::runner()` and `OpKernelContext::run_all_kernels_inline()`) enables function calls within a kernel to inherit that option from the calling context. As a concrete example, this enables a WhileOp or IfOp running with the SINGLE_THREADED_EXECUTOR executor_type to use the same optimizations in the branch/cond/body functions of those ops.

PiperOrigin-RevId: 309515120
Change-Id: I11b0b3ee458dd8ea1cdc9284f8acdb293c5cb770

The "DeleteIterator" op can block on work in the inter-op threadpool, e.g. waiting for function execution to terminate. As a result, we need to ensure that this blocking happens either on a separately created thread for the purpose (or a donated calling thread), which converting it to a `HybridAsyncOpKernel` achieves.

PiperOrigin-RevId: 309514952
Change-Id: I24f056bfbfbf4b1df0332432cc43aa96be123d6a

This fixes the tensorflow_serving OSS BUILD, which otherwise fails with the error:

```
ERROR: /tmpfs/tmp/bazel/external/org_tensorflow/tensorflow/python/keras/api/BUILD:121:1: Couldn't build file external/org_tensorflow/tensorflow/python/keras/api/_v1/__init__.py: Executing genrule @org_tensorflow//tensorflow/python/keras/api:keras_python_api_gen_compat_v1 failed (Exit 1)
> > ImportError: /tmpfs/tmp/bazel/execroot/tf_serving/bazel-out/host/bin/external/org_tensorflow/tensorflow/python/keras/api/create_tensorflow.python_api_keras_python_api_gen_compat_v1.runfiles/org_tensorflow/tensorflow/python/_pywrap_quantize_training.so: undefined symbol: _ZN10tensorflow38DoQuantizeTrainingOnSerializedGraphDefERKNSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEEEiS7_PS5_
```
PiperOrigin-RevId: 309514904
Change-Id: Ib4834d8e03f9193c7dafe096c697d9e5ef45f875

PiperOrigin-RevId: 309513623
Change-Id: If25c55f4700cb13906b001e1fb03d801fb84bf29

PiperOrigin-RevId: 309513590
Change-Id: I63c86395a92176fa7165b4ae9909b50fbe319f39

Create interfaces for distributed iterators and iterables. These interfaces are private for now. Use them to break the circular dependency with AutoGraph.

PiperOrigin-RevId: 309510977
Change-Id: Id2e3190e98be6dfa1adf804411d90812a96c30c4

PiperOrigin-RevId: 309510456
Change-Id: Ifc3683fd575072f6af2250667f6875c9c7399f6c

PiperOrigin-RevId: 309506396
Change-Id: I1652fd13a4d398236058c1806ff900576d39172b

PiperOrigin-RevId: 309505494
Change-Id: I939fb769f8338d99402592858e4d7b8e7a1aa56c

PiperOrigin-RevId: 309503752
Change-Id: Id1e4c5f24a90ec82068e8a5ed9579b5b1c84403d

PiperOrigin-RevId: 309501684
Change-Id: I14a7482a4bf18530d0d60a468cf3b81b7a7e7593

Correctness > speed. And using log1p improves accuracy significantly, especially of gradients.

*** Original change description ***

Partial rollback of change to softplus functor. The large speedup observed was only true for arguments greather than ~5. For arguments less than ~5, a significant slowdown would occur.

PiperOrigin-RevId: 309499644
Change-Id: Ie137dd5e2fa798f8c1e0330375e3e0b6dbcc31a6

Fixes a test whose randomness was affected by this.

PiperOrigin-RevId: 309495955
Change-Id: I4e184c9ca40b0102cf5eb895a040f9f2a09e753a

PiperOrigin-RevId: 309494739
Change-Id: Icae7c227f0515cefce403ab7f99176ad0b8eb4b2

Correctness > speed. And using log1p improves accuracy significantly, especially of gradients.

*** Original change description ***

Partial rollback of change to softplus functor. The large speedup observed was only true for arguments greather than ~5. For arguments less than ~5, a significant slowdown would occur.

PiperOrigin-RevId: 309487961
Change-Id: Ibe949c578a002d25250ba519acd16597981d1a4e

PiperOrigin-RevId: 309485102
Change-Id: I1234400cf973644d3c6aa6e09568aea3da3f3f63

mode.

When forming segments for implicit batch mode, the inputs of the operations in a
segment need to have a rank of not less than 2 and the implicit batch size of
the operations need to be consistent. Previously, we do not perform such
checking during segmentation. We may reject to build a TRTEngineOp for a
segment that doesn't meet such a requirement. We may even crash when executing
a TRTEngineOp built for a segment that doesn't meet such a requirement as shown
in the bug here.

Add test cases.

PiperOrigin-RevId: 309484466
Change-Id: I715817e9359d41075300911ca44aacb410e63ea0

PiperOrigin-RevId: 309482235
Change-Id: I54f4d159eaa88bed00cbf410037f025b1f2eb5ca

Previously, `RunSync()` was missing the optimization from e5c6881c that enables a single-component multi-device function from being executed as a local function. When 624f1a0f switched all tf.data functions to be considered as multi-device functions, the optimization became load-bearing, because occasionally it will dispatch a function to a ProcessFunctionLibraryRuntime that does not actually support multi-device functions, because it has no rendezvous factory.

The fix is simple: update the call to `GetHandleOnDevice()` in `PrepareRunSync()` to pass the optional `include_multi_device = true` argument.

PiperOrigin-RevId: 309481098
Change-Id: Id12d01cffc50399bab5711c7de4756c542873ec2

1) Reorder 2 successive adds with constants to allow constant folding.
2) Remove trivial Add/Sub/Mul/Div.

PiperOrigin-RevId: 309480295
Change-Id: Ie84f5fd4da79dc80a3a5ee806051bd63b5b49888

PiperOrigin-RevId: 309478605
Change-Id: Ic2b716cf6d4b418f0c05e9224e3f974fea9cc8d7

PiperOrigin-RevId: 309476584
Change-Id: I250ab2972a94cd9ce7527e2fb3de215d93b15c85

PiperOrigin-RevId: 309476085
Change-Id: Ice1d773353cdd58b0cfeb004f01c394677d6d3a8

Correctness > speed. And using log1p improves accuracy significantly, especially of gradients.

*** Original change description ***

Partial rollback of change to softplus functor. The large speedup observed was only true for arguments greather than ~5. For arguments less than ~5, a significant slowdown would occur.

PiperOrigin-RevId: 309472556
Change-Id: I015752231ebe7eb961c624fdc3d3a8e2f57101c6

PiperOrigin-RevId: 309472392
Change-Id: Ic8a8702f699e2138e5d7a4145eba6c15decf645f

[StreamExecutor] Allow HostExecutor users to control the stack sizes of threads used for HostStream via.

Also include non_portable_tags in the keys used when creating an Executor. There seems to be no good reason that it is omitted.

Will fix https://github.com/google/jax/issues/432 when included in a jaxlib release.

PiperOrigin-RevId: 309472318
Change-Id: Ia2535616047390d6bf6f2da82a666a321dcc9f5d

PiperOrigin-RevId: 309471787
Change-Id: If49367d3f6a56bd5fa361d8c23627e631c9feb8f

PiperOrigin-RevId: 309471131
Change-Id: I1085115abff5630b1ce94ef951a75d4f82bdc862

PiperOrigin-RevId: 309469850
Change-Id: I7054a3238dfbd9d94efb506e11d7ee91d811edd4

PiperOrigin-RevId: 309462718
Change-Id: I11e53b12b0b7ee7d6725a547c8016123d7f6682b

PiperOrigin-RevId: 309458102
Change-Id: I6b406a6b17726e409ab4798f40dd5f0791c696db

PiperOrigin-RevId: 309457859
Change-Id: I44bbc5804bb51b8248a52e9b2a0e863641905d17

This is related to https://github.com/tensorflow/tensorboard/pull/3564

1. Add DebuggedGraph.get_op_creation_digest()
2. Remove DebuggedGraph.get_op_type(), which is superseded by
   DebuggedGraph.get_op_creation_digest() and is not used anywhere.
3. Add DebuggedGraph.add_op_consumers() and DebuggedGraph.get_op_consumers()
   to enable efficient tracking of the downstream consuming ops of a graph
   op.
4. Add host_name and stack_frame_ids to data class GraphOpCreationDigest.

PiperOrigin-RevId: 309455936
Change-Id: I104084c1ef8b887f69733702a2f4c3190fa5402f

This CL introduces TfLiteHexagonDelegateOptionsDefault() method that returns
an instance of TfLiteHexagonDelegateOptions with default values.
Without an API for defaults, every new field added will break consumers of
hexagon delegate, as they will be filled with garbage values.

This is also similar to TfLite GPU delegate API.

PiperOrigin-RevId: 309453808
Change-Id: I12275991b34f175e2d1abbe40068287238f44684

PiperOrigin-RevId: 309452719
Change-Id: I1ac91638464b2a3a1914cee5cadb0353b307e6df

PiperOrigin-RevId: 309451157
Change-Id: Ie70aa64a31a832bd65e75d8d7a46895ee99cf00c

Previously the context was deleting some objects before custom devices had a chance to release their references. This change just moves custom device destruction to be a bit earlier.

PiperOrigin-RevId: 309450790
Change-Id: Id95d7918c5cbe87f6948ef302d6373f55bbb13ab