platform/upstream/tensorflow.git
7 years ago[TF:XLA:CPU] Alias tensorflow::gtl to gtl for brevity
Sanjoy Das [Tue, 23 Jan 2018 23:20:43 +0000 (15:20 -0800)]
[TF:XLA:CPU] Alias tensorflow::gtl to gtl for brevity

PiperOrigin-RevId: 183001493

7 years agoAdds regularization_losses in remaining heads.
A. Unique TensorFlower [Tue, 23 Jan 2018 23:11:22 +0000 (15:11 -0800)]
Adds regularization_losses in remaining heads.

PiperOrigin-RevId: 183000075

7 years agoSwitch models/ to use new test-specification format
A. Unique TensorFlower [Tue, 23 Jan 2018 23:08:26 +0000 (15:08 -0800)]
Switch models/ to use new test-specification format

PiperOrigin-RevId: 182999650

7 years agoLabeledTensor: don't test Tensor.shape object, test value instead.
Skye Wanderman-Milne [Tue, 23 Jan 2018 23:02:46 +0000 (15:02 -0800)]
LabeledTensor: don't test Tensor.shape object, test value instead.

With the C API enabled, a new TensorShape object may be generated
every Tensor.get_shape call.

PiperOrigin-RevId: 182998609

7 years agoTweaks to PR 16101 which adds stream selection support for `tf.contrib.ffmpeg.decode_...
Yong Tang [Tue, 23 Jan 2018 23:01:14 +0000 (15:01 -0800)]
Tweaks to PR 16101 which adds stream selection support for `tf.contrib.ffmpeg.decode_audio`.

PiperOrigin-RevId: 182998275

7 years agoMake bijectors/reshape_test.py work with C API enabled.
Skye Wanderman-Milne [Tue, 23 Jan 2018 22:54:22 +0000 (14:54 -0800)]
Make bijectors/reshape_test.py work with C API enabled.

Some of the error messages change with the C API enabled due to slight
differences in shape inference (in this case, the C API catches shape
errors sooner). This change refactors some of the tests to expect
different error messages in different cases.

PiperOrigin-RevId: 182997183

7 years agoTurn private Operation and Tensor fields into properties.
Skye Wanderman-Milne [Tue, 23 Jan 2018 22:31:15 +0000 (14:31 -0800)]
Turn private Operation and Tensor fields into properties.

Unfortunately there are many uses of these private fields. Turning
them into properties that call public API methods will allow enabling
the C API before fixing all such uses.

PiperOrigin-RevId: 182993539

7 years agoReplace the use of ConsumeValueOrDie with TF_ASSERT_OK_AND_ASSIGN.
A. Unique TensorFlower [Tue, 23 Jan 2018 22:25:36 +0000 (14:25 -0800)]
Replace the use of ConsumeValueOrDie with TF_ASSERT_OK_AND_ASSIGN.

PiperOrigin-RevId: 182992618

7 years agoUpdating the update_version script to act appropriately for rc0 when it comes
Amit Patankar [Tue, 23 Jan 2018 22:24:53 +0000 (14:24 -0800)]
Updating the update_version script to act appropriately for rc0 when it comes
to the install sources tested configurations table along with a warning. Using perl rather than sed as it supports regex look behind.

PiperOrigin-RevId: 182992467

7 years agoC API: don't return source/sink control edges.
Skye Wanderman-Milne [Tue, 23 Jan 2018 22:06:13 +0000 (14:06 -0800)]
C API: don't return source/sink control edges.

PiperOrigin-RevId: 182989157

7 years agoFix StridedSlice.
Yao Zhang [Tue, 23 Jan 2018 21:51:44 +0000 (13:51 -0800)]
Fix StridedSlice.
(1) Use input index instead of input name to make sure the naming is unique,
because in rare cases, two inputs may connect to the same node.
(2) Consecutive if statements each should have a return and should not fall
through; this created un-wanted behavior. Use switch and break now.

PiperOrigin-RevId: 182986609

7 years agoAdd support for conv and matmul operators with variable filter.
Nupur Garg [Tue, 23 Jan 2018 21:35:08 +0000 (13:35 -0800)]
Add support for conv and matmul operators with variable filter.

PiperOrigin-RevId: 182984061

7 years agoAdd a test case to validate number of buckets after quantile generation.
A. Unique TensorFlower [Tue, 23 Jan 2018 21:12:00 +0000 (13:12 -0800)]
Add a test case to validate number of buckets after quantile generation.

Add an option to generate quantiles instead of sampled boundaries.
This option will always return exactly the number of requested boundaries,
regardless of duplicates and fewer inputs than requested boundaries.

PiperOrigin-RevId: 182980861

7 years agoGo: Update generated wrapper functions for TensorFlow ops.
A. Unique TensorFlower [Tue, 23 Jan 2018 20:52:36 +0000 (12:52 -0800)]
Go: Update generated wrapper functions for TensorFlow ops.

PiperOrigin-RevId: 182977979

7 years agoAdding tf_export decorators/calls to TensorFlow functions and constants.
Anna R [Tue, 23 Jan 2018 20:48:44 +0000 (12:48 -0800)]
Adding tf_export decorators/calls to TensorFlow functions and constants.

PiperOrigin-RevId: 182977460

7 years ago[XLA:GPU] Support BF16 data type.
A. Unique TensorFlower [Tue, 23 Jan 2018 20:48:00 +0000 (12:48 -0800)]
[XLA:GPU] Support BF16 data type.

Add an HLO pass to the GPU backend to implement BF16 operations with F32
operations.

Define macro XLA_BACKEND_SUPPORTS_BFLOAT16=1 when building tests for the GPU
backend to enable BF16 tests for GPU. Enable bfloat16_test and other BF16 tests
for GPU.

Add hlo_element_type_converter_test.

Add convolution tests and matrix multiplication tests for BF16.

PiperOrigin-RevId: 182977358

7 years agointernal change
A. Unique TensorFlower [Tue, 23 Jan 2018 20:41:07 +0000 (12:41 -0800)]
internal change

PiperOrigin-RevId: 182976445

7 years ago[TF:XLA] Bump open source llvm revision to r323181
Sanjoy Das [Tue, 23 Jan 2018 20:39:22 +0000 (12:39 -0800)]
[TF:XLA] Bump open source llvm revision to r323181

PiperOrigin-RevId: 182976206

7 years agoInternal Change
A. Unique TensorFlower [Tue, 23 Jan 2018 20:24:21 +0000 (12:24 -0800)]
Internal Change

PiperOrigin-RevId: 182974191

7 years agoUpdate ops-related pbtxt files.
A. Unique TensorFlower [Tue, 23 Jan 2018 20:19:06 +0000 (12:19 -0800)]
Update ops-related pbtxt files.

PiperOrigin-RevId: 182973494

7 years agoAllow fisher-vector multiplies on 1D vectors for CategoricalLogitsNegativeLogProbLoss.
A. Unique TensorFlower [Tue, 23 Jan 2018 20:14:52 +0000 (12:14 -0800)]
Allow fisher-vector multiplies on 1D vectors for CategoricalLogitsNegativeLogProbLoss.

PiperOrigin-RevId: 182972901

7 years agoCorrectly incref items put into the tuple to keep msan happy.
Akshay Modi [Tue, 23 Jan 2018 20:01:11 +0000 (12:01 -0800)]
Correctly incref items put into the tuple to keep msan happy.

PiperOrigin-RevId: 182970703

7 years ago[XLA] Remove LiteralTestUtil::EqualTuple / NearTuple.
Justin Lebar [Tue, 23 Jan 2018 19:46:18 +0000 (11:46 -0800)]
[XLA] Remove LiteralTestUtil::EqualTuple / NearTuple.

Instead, simply DTRT when you pass a tuple to Equal/Near.

PiperOrigin-RevId: 182968283

7 years agoAdd `tf.contrib.distributions.Autoregressive`.
Joshua V. Dillon [Tue, 23 Jan 2018 19:37:58 +0000 (11:37 -0800)]
Add `tf.contrib.distributions.Autoregressive`.

PiperOrigin-RevId: 182966827

7 years agoMake feature_column_test.py run with the C API enabled.
Skye Wanderman-Milne [Tue, 23 Jan 2018 19:23:35 +0000 (11:23 -0800)]
Make feature_column_test.py run with the C API enabled.

I only enabled running the C API with test classes that other testing
found failures in to avoid increasing the test time unnecessarily.

PiperOrigin-RevId: 182964349

7 years agoMoves batch ops to core and exposes tf.contrib.batching
Alexandre Passos [Tue, 23 Jan 2018 19:20:29 +0000 (11:20 -0800)]
Moves batch ops to core and exposes tf.contrib.batching

PiperOrigin-RevId: 182963906

7 years agoRun TF opensource presubmits with sanybridge CPU optimizations.
Gunhan Gulsoy [Tue, 23 Jan 2018 19:09:26 +0000 (11:09 -0800)]
Run TF opensource presubmits with sanybridge CPU optimizations.

PiperOrigin-RevId: 182961789

7 years agoTFE usability: Implement `ndim` for `EagerTensor`+NumPy compatibility
Akshay Agrawal [Tue, 23 Jan 2018 19:05:50 +0000 (11:05 -0800)]
TFE usability: Implement `ndim` for `EagerTensor`+NumPy compatibility

Third party libraries like matplotlib often access `ndim`.

PiperOrigin-RevId: 182961097

7 years agoRemoved unecessary dependency
Benoit Steiner [Tue, 23 Jan 2018 18:59:00 +0000 (10:59 -0800)]
Removed unecessary dependency

PiperOrigin-RevId: 182959711

7 years agoIncrease shard count of tensorflow/contrib/learn:linear_test to avoid test timeouts.
A. Unique TensorFlower [Tue, 23 Jan 2018 18:19:19 +0000 (10:19 -0800)]
Increase shard count of tensorflow/contrib/learn:linear_test to avoid test timeouts.

The test tensorflow/contrib/learn:linear_test has been getting flaky timeouts t
a low rate for some time, but they increased greatly under the address
sanitizer due to a recent change there.

This CL increases the test's shard count to bring the time for each shard
well under the 5 minute timeout.

PiperOrigin-RevId: 182953078

7 years agoThe link translator expects "@{" for API links.
Mark Daoust [Tue, 23 Jan 2018 17:46:25 +0000 (09:46 -0800)]
The link translator expects "@{" for API links.

This is currently not rendering:

https://www.tensorflow.org/api_docs/python/tf/contrib/tpu/TPUEstimatorSpec

PiperOrigin-RevId: 182947562

7 years agoCorrect handling of dtype for Categorical sampling.
A. Unique TensorFlower [Tue, 23 Jan 2018 17:18:11 +0000 (09:18 -0800)]
Correct handling of dtype for Categorical sampling.

PiperOrigin-RevId: 182943806

7 years agoTurn off address sanitizer checking for tensorflow/python/kernel_tests:sparse_ops_test
A. Unique TensorFlower [Tue, 23 Jan 2018 08:10:52 +0000 (00:10 -0800)]
Turn off address sanitizer checking for tensorflow/python/kernel_tests:sparse_ops_test

It getting too many timeouts.

PiperOrigin-RevId: 182892876

7 years ago[XLA] Inline definitions of NativeToPrimitiveType<T>.
Justin Lebar [Tue, 23 Jan 2018 07:29:57 +0000 (23:29 -0800)]
[XLA] Inline definitions of NativeToPrimitiveType<T>.

These functions are hot in the evaluator, and that's silly, because
they're pure constants.  Just inline them.

PiperOrigin-RevId: 182889992

7 years ago[XLA] Speed up HloEvaluator's convolution routine.
Justin Lebar [Tue, 23 Jan 2018 06:56:04 +0000 (22:56 -0800)]
[XLA] Speed up HloEvaluator's convolution routine.

Skip two integer divisions when the window has no dilation.  This is
good for a ~30% speedup on my machine.

PiperOrigin-RevId: 182887749

7 years agoPuts the iterations_per_loop, which controls how many steps the TPU system to
Jianwei Xie [Tue, 23 Jan 2018 06:06:10 +0000 (22:06 -0800)]
Puts the iterations_per_loop, which controls how many steps the TPU system to
run for each training/evaluatoin loop, into local variable collection.

PiperOrigin-RevId: 182884955

7 years ago[XLA] Make kokoro happy in prng_test
Yunxing Dai [Tue, 23 Jan 2018 04:43:06 +0000 (20:43 -0800)]
[XLA] Make kokoro happy in prng_test

Open source builds don't support bit_cast yet, use reinterpret cast instead.
RELNOTES: n/a
PiperOrigin-RevId: 182879589

7 years agoAdding tf_export decorators/calls to TensorFlow functions and constants.
Anna R [Tue, 23 Jan 2018 04:07:42 +0000 (20:07 -0800)]
Adding tf_export decorators/calls to TensorFlow functions and constants.

PiperOrigin-RevId: 182877401

7 years agoGo: Update generated wrapper functions for TensorFlow ops.
A. Unique TensorFlower [Tue, 23 Jan 2018 02:45:56 +0000 (18:45 -0800)]
Go: Update generated wrapper functions for TensorFlow ops.

PiperOrigin-RevId: 182871435

7 years agoUpdate ops-related pbtxt files.
A. Unique TensorFlower [Tue, 23 Jan 2018 02:18:46 +0000 (18:18 -0800)]
Update ops-related pbtxt files.

PiperOrigin-RevId: 182868865

7 years ago_DefinedFunction.name should be a string.
Skye Wanderman-Milne [Tue, 23 Jan 2018 01:43:06 +0000 (17:43 -0800)]
_DefinedFunction.name should be a string.

With the C API enabled and python2, it was being set to a unicode
object.

PiperOrigin-RevId: 182865087

7 years agoReject retried RecvTensor requests.
Jeremy Lau [Tue, 23 Jan 2018 01:26:43 +0000 (17:26 -0800)]
Reject retried RecvTensor requests.

Retried RecvTensorRequests are problematic because a RecvTensor with no
corresponding sender will wait forever, and the tensor may have been delivered
to a previous retry.

This change adds a unique request_id to each RecvTensor request, and we check
these request_ids against a set of recent request_ids. If a request_id is in the
recent set, we reject the RecvTensor request.

PiperOrigin-RevId: 182863245

7 years agoAdding tf_export decorators/calls to TensorFlow functions and constants.
Anna R [Tue, 23 Jan 2018 01:16:44 +0000 (17:16 -0800)]
Adding tf_export decorators/calls to TensorFlow functions and constants.

PiperOrigin-RevId: 182862075

7 years agoOps to get and set individual elements of TensorLists, and their gradients.
Alexandre Passos [Tue, 23 Jan 2018 01:12:18 +0000 (17:12 -0800)]
Ops to get and set individual elements of TensorLists, and their gradients.

PiperOrigin-RevId: 182861559

7 years agoAdd SeparableConv1D layer and refactor SeparableConv2D to maximize code reuse and...
Francois Chollet [Tue, 23 Jan 2018 01:10:35 +0000 (17:10 -0800)]
Add SeparableConv1D layer and refactor SeparableConv2D to maximize code reuse and maintenability.

PiperOrigin-RevId: 182861372

7 years agoReplaced LOGs with NSLogs.
A. Unique TensorFlower [Tue, 23 Jan 2018 00:44:22 +0000 (16:44 -0800)]
Replaced LOGs with NSLogs.
Also inserted exit(-1) after LOG(FATAL)s.
Removed unused definition of CHECK macro.

PiperOrigin-RevId: 182857784

7 years agoAdds unit tests for squeeze, to test the case when all dims are 1.
A. Unique TensorFlower [Tue, 23 Jan 2018 00:41:37 +0000 (16:41 -0800)]
Adds unit tests for squeeze, to test the case when all dims are 1.

PiperOrigin-RevId: 182857374

7 years agoAdd functionality to include an optional additional header to be sent on every reques...
Frank Chen [Tue, 23 Jan 2018 00:28:47 +0000 (16:28 -0800)]
Add functionality to include an optional additional header to be sent on every request when using the GCS file system.

PiperOrigin-RevId: 182855525

7 years agoAdd file name and line number to statuses generated by OP_REQUIRES macros.
Jeremy Lau [Mon, 22 Jan 2018 23:54:33 +0000 (15:54 -0800)]
Add file name and line number to statuses generated by OP_REQUIRES macros.

PiperOrigin-RevId: 182850486

7 years ago[XLA][BF16] Add tests to demonstrate the brokeness of BF16 RNG functions in LLVM...
Yunxing Dai [Mon, 22 Jan 2018 23:10:08 +0000 (15:10 -0800)]
[XLA][BF16] Add tests to demonstrate the brokeness of BF16 RNG functions in LLVM backends.
- Added prng tests to demonstrate the brokeness of BF16 RNG functions in LLVM backends.

PiperOrigin-RevId: 182843680

7 years agoAdding tf_export decorators/calls to TensorFlow functions and constants.
Anna R [Mon, 22 Jan 2018 22:59:19 +0000 (14:59 -0800)]
Adding tf_export decorators/calls to TensorFlow functions and constants.

PiperOrigin-RevId: 182841644

7 years agoThread-safety for the TFE tape code.
Alexandre Passos [Mon, 22 Jan 2018 22:21:16 +0000 (14:21 -0800)]
Thread-safety for the TFE tape code.

In multithreaded python any call to a python API function can release the GIL,
so we should not assume our data structures haven't changed across calls to
the python API.

PiperOrigin-RevId: 182835555

7 years agoTrack swapping candidates directly in memory instead of annotating them in the graph
Benoit Steiner [Mon, 22 Jan 2018 21:58:08 +0000 (13:58 -0800)]
Track swapping candidates directly in memory instead of annotating them in the graph

PiperOrigin-RevId: 182831471

7 years agoBuildifier cleanup.
Yifei Feng [Mon, 22 Jan 2018 21:42:50 +0000 (13:42 -0800)]
Buildifier cleanup.

PiperOrigin-RevId: 182829120

7 years ago[TF:XLA] Bump open source llvm revision to r323081
Sanjoy Das [Mon, 22 Jan 2018 21:21:46 +0000 (13:21 -0800)]
[TF:XLA] Bump open source llvm revision to r323081

PiperOrigin-RevId: 182825537

7 years agoAllowing custom creators for variable_scope.variable()
Alexandre Passos [Mon, 22 Jan 2018 21:00:54 +0000 (13:00 -0800)]
Allowing custom creators for variable_scope.variable()

PiperOrigin-RevId: 182822430

7 years agoExtend the API with a pair of decorators that can convert functions inline.
A. Unique TensorFlower [Mon, 22 Jan 2018 19:27:33 +0000 (11:27 -0800)]
Extend the API with a pair of decorators that can convert functions inline.
Add logic to detect these decorators and treat them appropriately (e.g. when converting recursively, do all decorated functions as they are).

PiperOrigin-RevId: 182808673

7 years agoAdd new src targets for C API and native TF
A. Unique TensorFlower [Mon, 22 Jan 2018 19:18:44 +0000 (11:18 -0800)]
Add new src targets for C API and native TF

PiperOrigin-RevId: 182807258

7 years agoAdd support for tf.Example to saved_model_cli.
Yifei Feng [Mon, 22 Jan 2018 19:07:28 +0000 (11:07 -0800)]
Add support for tf.Example to saved_model_cli.
You can now pass tf.Example as inputs to with option --input_examples, for example:
--input_examples 'input_key=[{"age":[25],"education":["11th"]}]'

PiperOrigin-RevId: 182805401

7 years agotf.keras recurrent Layers: Fix doc strings on default activations
Shanqing Cai [Mon, 22 Jan 2018 19:02:40 +0000 (11:02 -0800)]
tf.keras recurrent Layers: Fix doc strings on default activations

* The default activation for input is `tanh`, not `linear`, as
  the previous doc string might mislead readers to think.
* Also clarify that the default activation for recurrent is
  `hard_sigmoid`.

PiperOrigin-RevId: 182804493

7 years agoFix TFLite Conv reference implementation
Yu-Cheng Ling [Mon, 22 Jan 2018 18:48:20 +0000 (10:48 -0800)]
Fix TFLite Conv reference implementation

PiperOrigin-RevId: 182802058

7 years agoAllow whitespace after commas in hparam strings.
A. Unique TensorFlower [Mon, 22 Jan 2018 18:36:05 +0000 (10:36 -0800)]
Allow whitespace after commas in hparam strings.

E.g., allow 'a=b, c=d' or 'a=b,\nc=d' in addition to just 'a=b,c=d'.

PiperOrigin-RevId: 182800230

7 years agoInternal change
A. Unique TensorFlower [Mon, 22 Jan 2018 18:19:07 +0000 (10:19 -0800)]
Internal change

PiperOrigin-RevId: 182797605

7 years agomnist and resnet eager examples avoiding optimizer.minimize
Alexandre Passos [Mon, 22 Jan 2018 17:47:55 +0000 (09:47 -0800)]
mnist and resnet eager examples avoiding optimizer.minimize

PiperOrigin-RevId: 182792844

7 years agoReducing op overhead for eager.
Akshay Modi [Mon, 22 Jan 2018 17:39:44 +0000 (09:39 -0800)]
Reducing op overhead for eager.

Benchmarks:
(new) tfe_py_fastpath_execute_matmul: 7.86213080088 walltime
(old) gen_math_ops_matmul: 11.2566947937 walltime

(The slowdown is due to adding the record_gradient callback)

This will be a 3-step process:

1. (This CL) Add a function that allows execution of an op on the fastpath.
2. Update python generation code to create 2 new python functions
(_op/_op_gradient_callback and _op_slowpath_fallback):
   -- first function (_op) checks if it is in graph mode, if so, do the normal
      thing, else call out to the function added in step 1.
   -- second function does the else part similar to today (calling out to
   args_to_matching_eager etc.)
3. Rename the first function generated above to be the canonical _op function.

PiperOrigin-RevId: 182791741

7 years agoMake array_ops_test.py work with the C API enabled.
Skye Wanderman-Milne [Mon, 22 Jan 2018 16:31:23 +0000 (08:31 -0800)]
Make array_ops_test.py work with the C API enabled.

I only enabled the test class that other testing identified needed changes.

PiperOrigin-RevId: 182783262

7 years agoImplement native support of CycleGAN in tf.contrib.gan
A. Unique TensorFlower [Mon, 22 Jan 2018 16:19:56 +0000 (08:19 -0800)]
Implement native support of CycleGAN in tf.contrib.gan

PiperOrigin-RevId: 182782072

7 years agoInternal cleanup
A. Unique TensorFlower [Mon, 22 Jan 2018 15:47:48 +0000 (07:47 -0800)]
Internal cleanup

PiperOrigin-RevId: 182777782

7 years agoReplace "Google Container Engine" by "Google Kubernetes Engine" in README.md.
A. Unique TensorFlower [Mon, 22 Jan 2018 14:39:21 +0000 (06:39 -0800)]
Replace "Google Container Engine" by "Google Kubernetes Engine" in README.md.

PiperOrigin-RevId: 182771284

7 years agoFix a bug in the associative & commutative rewrite in constant folding:
A. Unique TensorFlower [Sun, 21 Jan 2018 17:22:43 +0000 (09:22 -0800)]
Fix a bug in the associative & commutative rewrite in constant folding:
I finally tracked it down to the case where we do the transformation:

    //
    //                      +                +       = parent
    //                     / \              / \
    //                    C   +    -- >    X   +     = children
    //                       / \              / \
    //                      X   Y            C   Y   = leaves

when there exists a control dependency from the child '+' to C. This control dependency creates a cycle in the graph when X and C are simply swapped. We can fix it by moving the control dependency and anchoring it on Y instead (We still need it in case we are inside a frame).

PiperOrigin-RevId: 182705910

7 years ago[TF:XLA] Bump open source llvm revision to r323005
Benjamin Kramer [Sun, 21 Jan 2018 16:03:46 +0000 (08:03 -0800)]
[TF:XLA] Bump open source llvm revision to r323005

PiperOrigin-RevId: 182703514

7 years agoImprove error message in create_python_api.py. The new message would print
Anna R [Sun, 21 Jan 2018 04:34:12 +0000 (20:34 -0800)]
Improve error message in create_python_api.py. The new message would print
exact filenames that need to be added to the
tensorflow/tools/api/generator/BUILD file.

PiperOrigin-RevId: 182679235

7 years ago[XLA:GPU] Warn on more broken CUDA versions.
Justin Lebar [Sun, 21 Jan 2018 02:13:15 +0000 (18:13 -0800)]
[XLA:GPU] Warn on more broken CUDA versions.

We've received additional data on which CUDA versions contain a ptx
assembler bug that affects XLA.  Warn if you're using any of these.

PiperOrigin-RevId: 182675055

7 years agoUpdate XLA for LLVM r323001
Benjamin Kramer [Sat, 20 Jan 2018 18:50:16 +0000 (10:50 -0800)]
Update XLA for LLVM r323001

This will require an LLVM version bump

PiperOrigin-RevId: 182661291

7 years agoAdded BeamRoot to manage instances of BeamEntry for CTCBeamSearchDecoder
A. Unique TensorFlower [Sat, 20 Jan 2018 04:52:19 +0000 (20:52 -0800)]
Added BeamRoot to manage instances of BeamEntry for CTCBeamSearchDecoder
to avoid recursive destructor call.

PiperOrigin-RevId: 182625992

7 years agoTurn the op_performance_data proto lib into a header only library by default
Benoit Steiner [Sat, 20 Jan 2018 03:13:43 +0000 (19:13 -0800)]
Turn the op_performance_data proto lib into a header only library by default

PiperOrigin-RevId: 182621348

7 years agoAdd check for valid outfeed shape on outfeed creation and in HloVerifier.
Mark Heffernan [Sat, 20 Jan 2018 02:05:19 +0000 (18:05 -0800)]
Add check for valid outfeed shape on outfeed creation and in HloVerifier.

PiperOrigin-RevId: 182616345

7 years ago[XLA] Update R4 reduce window test cases.
A. Unique TensorFlower [Sat, 20 Jan 2018 01:18:25 +0000 (17:18 -0800)]
[XLA] Update R4 reduce window test cases.

PiperOrigin-RevId: 182612106

7 years agoAdd clif rules for function.proto.
Jacques Pienaar [Sat, 20 Jan 2018 00:55:11 +0000 (16:55 -0800)]
Add clif rules for function.proto.

PiperOrigin-RevId: 182609604

7 years agoMake sure the same rewrite is not performed multiple times in ArithmeticOptimizer...
A. Unique TensorFlower [Sat, 20 Jan 2018 00:54:39 +0000 (16:54 -0800)]
Make sure the same rewrite is not performed multiple times in ArithmeticOptimizer, and that added nodes are unique.
A couple of minor cleanups.

PiperOrigin-RevId: 182609552

7 years agoReport incompatible shapes in case of check failure.
Jacques Pienaar [Fri, 19 Jan 2018 23:42:39 +0000 (15:42 -0800)]
Report incompatible shapes in case of check failure.

PiperOrigin-RevId: 182600051

7 years agoAutomated g4 rollback of changelist 182559231
A. Unique TensorFlower [Fri, 19 Jan 2018 23:31:24 +0000 (15:31 -0800)]
Automated g4 rollback of changelist 182559231

PiperOrigin-RevId: 182598631

7 years ago[XLA] Add xla_dump_prepass_hlo_proto_to debug flag
Nick Desaulniers [Fri, 19 Jan 2018 23:01:35 +0000 (15:01 -0800)]
[XLA] Add xla_dump_prepass_hlo_proto_to debug flag

Add xla_dump_prepass_hlo_proto_to debug flag for dumping HLO protos *BEFORE*
running HLO passes, which will perform backend specific optimizations.

PiperOrigin-RevId: 182594324

7 years agoTFLite export optional tensor as -1 id.
Zhixian Yan [Fri, 19 Jan 2018 22:51:11 +0000 (14:51 -0800)]
TFLite export optional tensor as -1 id.

PiperOrigin-RevId: 182592943

7 years ago[tf.data] Minor fix for saveable MapDataset and ParallelMapDataset.
Shivani Agrawal [Fri, 19 Jan 2018 22:38:45 +0000 (14:38 -0800)]
[tf.data] Minor fix for saveable MapDataset and ParallelMapDataset.

PiperOrigin-RevId: 182591174

7 years agoPublish TPU Estimators Docs on github
A. Unique TensorFlower [Fri, 19 Jan 2018 22:05:20 +0000 (14:05 -0800)]
Publish TPU  Estimators Docs on github

PiperOrigin-RevId: 182585909

7 years agoFilling in some holes in C64 support with Literals.
Brian Patton [Fri, 19 Jan 2018 22:03:27 +0000 (14:03 -0800)]
Filling in some holes in C64 support with Literals.

PiperOrigin-RevId: 182585597

7 years ago[XLA] Simplify trivial pad/reduce-window combos into broadcasts.
Chris Leary [Fri, 19 Jan 2018 22:01:29 +0000 (14:01 -0800)]
[XLA] Simplify trivial pad/reduce-window combos into broadcasts.

PiperOrigin-RevId: 182585236

7 years agoUpdate pin for bazel-toolchains to latest version
A. Unique TensorFlower [Fri, 19 Jan 2018 21:05:53 +0000 (13:05 -0800)]
Update pin for bazel-toolchains to latest version

https://github.com/bazelbuild/bazel-toolchains/releases/tag/f3b0970

PiperOrigin-RevId: 182576952

7 years ago[TF:XLA] Improve the error message when ComputeConstant fails.
Peter Hawkins [Fri, 19 Jan 2018 20:09:13 +0000 (12:09 -0800)]
[TF:XLA] Improve the error message when ComputeConstant fails.

Make comment in computation_builder.h more precise.

PiperOrigin-RevId: 182569583

7 years agoAdd support for time_major shape format to the sequential RNN Op in TF Lite.
A. Unique TensorFlower [Fri, 19 Jan 2018 20:08:47 +0000 (12:08 -0800)]
Add support for time_major shape format to the sequential RNN Op in TF Lite.

This option, if set, changes the shape format of the inputs and outputs to
[max_time, batch_size, depth]. If false, it uses [batch_size, max_time, depth].
By default, it is set to false.

PiperOrigin-RevId: 182569507

7 years agoHandling resource captures when defining functions.
Alexandre Passos [Fri, 19 Jan 2018 20:03:50 +0000 (12:03 -0800)]
Handling resource captures when defining functions.

PiperOrigin-RevId: 182568701

7 years agoMove layers_dense_variational_impl.py to layers_dense_variational.py.
Dustin Tran [Fri, 19 Jan 2018 20:00:33 +0000 (12:00 -0800)]
Move layers_dense_variational_impl.py to layers_dense_variational.py.

PiperOrigin-RevId: 182568201

7 years agoChange deprecated keep_dims parameter to keepdims.
A. Unique TensorFlower [Fri, 19 Jan 2018 19:59:42 +0000 (11:59 -0800)]
Change deprecated keep_dims parameter to keepdims.

PiperOrigin-RevId: 182568093

7 years ago [tpu:profiler] Do not generate output files from other cloud tools if no trace...
A. Unique TensorFlower [Fri, 19 Jan 2018 19:49:29 +0000 (11:49 -0800)]
   [tpu:profiler] Do not generate output files from other cloud tools if no trace is collected.

PiperOrigin-RevId: 182566765

7 years ago[TF:XLA] Bump open source llvm revision to r322942
Sanjoy Das [Fri, 19 Jan 2018 19:46:55 +0000 (11:46 -0800)]
[TF:XLA] Bump open source llvm revision to r322942

PiperOrigin-RevId: 182566386

7 years agoAdd convolutional Flipout layers.
Dustin Tran [Fri, 19 Jan 2018 19:04:26 +0000 (11:04 -0800)]
Add convolutional Flipout layers.

PiperOrigin-RevId: 182559534

7 years agoCorrect handling of dtype for multinomial sampling.
A. Unique TensorFlower [Fri, 19 Jan 2018 19:02:50 +0000 (11:02 -0800)]
Correct handling of dtype for multinomial sampling.

PiperOrigin-RevId: 182559231

7 years agoAdd option for loop unrolling in MaskedAutoregressiveFlow bijector.
A. Unique TensorFlower [Fri, 19 Jan 2018 19:01:45 +0000 (11:01 -0800)]
Add option for loop unrolling in MaskedAutoregressiveFlow bijector.

PiperOrigin-RevId: 182559030

7 years agoFixes issue with using lookup ops in datasets from eager.
Alexandre Passos [Fri, 19 Jan 2018 18:36:05 +0000 (10:36 -0800)]
Fixes issue with using lookup ops in datasets from eager.

Addresses issue #16160

PiperOrigin-RevId: 182554969