platform/upstream/tensorflow.git
6 years agoUpdate comment clarifying continuous eval behavior.
A. Unique TensorFlower [Tue, 8 May 2018 15:57:45 +0000 (08:57 -0700)]
Update comment clarifying continuous eval behavior.

PiperOrigin-RevId: 195826025

6 years ago[TF:XLA] Fix NaN in StatelessRandomNormal if the underlying uniform distribution...
Peter Hawkins [Tue, 8 May 2018 15:07:08 +0000 (08:07 -0700)]
[TF:XLA] Fix NaN in StatelessRandomNormal if the underlying uniform distribution returned -1.

PiperOrigin-RevId: 195819645

6 years agoAutomated g4 rollback of changelist 195723288
A. Unique TensorFlower [Tue, 8 May 2018 15:04:07 +0000 (08:04 -0700)]
Automated g4 rollback of changelist 195723288

PiperOrigin-RevId: 195819297

6 years agoAdd missing #include for OpResponse. This class currently happens to be forward
A. Unique TensorFlower [Tue, 8 May 2018 14:57:12 +0000 (07:57 -0700)]
Add missing #include for OpResponse. This class currently happens to be forward
declared by xla.proto.h, but that proto doesn't actually need this type
anywhere and we are working on removing such unneeded forward declarations.

PiperOrigin-RevId: 195818397

6 years agoProfileHandler: Remove unnecessary interface method.
Asim Shankar [Tue, 8 May 2018 14:28:43 +0000 (07:28 -0700)]
ProfileHandler: Remove unnecessary interface method.
PiperOrigin-RevId: 195815565

6 years agoFix a test expectation.
A. Unique TensorFlower [Tue, 8 May 2018 10:26:06 +0000 (03:26 -0700)]
Fix a test expectation.

PiperOrigin-RevId: 195796348

6 years agoAutomated g4 rollback of changelist 195748721
A. Unique TensorFlower [Tue, 8 May 2018 09:11:52 +0000 (02:11 -0700)]
Automated g4 rollback of changelist 195748721

PiperOrigin-RevId: 195790581

6 years agoTemporarily disable concat rewrite.
A. Unique TensorFlower [Tue, 8 May 2018 02:56:26 +0000 (19:56 -0700)]
Temporarily disable concat rewrite.

PiperOrigin-RevId: 195762860

6 years ago[tf.data] Move tensorflow::dataset::MakeIteratorContext to core/framework
Brennan Saeta [Tue, 8 May 2018 01:31:47 +0000 (18:31 -0700)]
[tf.data] Move tensorflow::dataset::MakeIteratorContext to core/framework

PiperOrigin-RevId: 195756342

6 years agoAdd logic for StridedSlice ops in ShapeRefiner::ConstantPartialShape().
Skye Wanderman-Milne [Tue, 8 May 2018 00:28:41 +0000 (17:28 -0700)]
Add logic for StridedSlice ops in ShapeRefiner::ConstantPartialShape().

This mimics the logic in tensor_util.constant_value_as_shape, allowing
the C++ shape inference code to infer more shapes than it could before.

This change also adds an optional stride argument to InferenceContext::Subshape().

PiperOrigin-RevId: 195749522

6 years agoMake conv2d_tranpose_test.py work with C API shapes enabled.
Skye Wanderman-Milne [Tue, 8 May 2018 00:24:28 +0000 (17:24 -0700)]
Make conv2d_tranpose_test.py work with C API shapes enabled.

The C API provides more accurate shape information in many cases.

PiperOrigin-RevId: 195749030

6 years agoMake eager functions runable on TPU
Igor Ganichev [Tue, 8 May 2018 00:21:53 +0000 (17:21 -0700)]
Make eager functions runable on TPU

PiperOrigin-RevId: 195748721

6 years agoRaise an error if we try to take the gradient wrt to the initial value of a loop...
Skye Wanderman-Milne [Tue, 8 May 2018 00:21:39 +0000 (17:21 -0700)]
Raise an error if we try to take the gradient wrt to the initial value of a loop variable.

Fixes #14101

PiperOrigin-RevId: 195748688

6 years agoInternal change
Blake Hechtman [Tue, 8 May 2018 00:00:27 +0000 (17:00 -0700)]
Internal change

PiperOrigin-RevId: 195745819

6 years agoAdd test with tf.cond.
Jacques Pienaar [Mon, 7 May 2018 23:59:41 +0000 (16:59 -0700)]
Add test with tf.cond.

PiperOrigin-RevId: 195745718

6 years agoDelete kTransposeDot (it is no longer in use)
Sanjoy Das [Mon, 7 May 2018 23:55:10 +0000 (16:55 -0700)]
Delete kTransposeDot (it is no longer in use)

PiperOrigin-RevId: 195745124

6 years agoFast-path to VarHandleOp
Alexandre Passos [Mon, 7 May 2018 23:49:44 +0000 (16:49 -0700)]
Fast-path to VarHandleOp

PiperOrigin-RevId: 195744374

6 years agoAdd TFX section. Add Ecosystem page and dropdown menu.
Billy Lamberta [Mon, 7 May 2018 23:38:02 +0000 (16:38 -0700)]
Add TFX section. Add Ecosystem page and dropdown menu.

PiperOrigin-RevId: 195742728

6 years agoReorder executor NodeItem variable length data section so
A. Unique TensorFlower [Mon, 7 May 2018 23:31:07 +0000 (16:31 -0700)]
Reorder executor NodeItem variable length data section so
that all multi-byte aligned types precede all byte-aligned
types so that alignment is satisfied without padding.

PiperOrigin-RevId: 195741712

6 years agoShapeRefiner fix: some variant-type tensors have handle data.
Skye Wanderman-Milne [Mon, 7 May 2018 23:16:32 +0000 (16:16 -0700)]
ShapeRefiner fix: some variant-type tensors have handle data.

ShapeRefiner::AddNode() would only propagate handle data for
DT_RESOURCE tensors, but not DT_VARIANT. The Python shape inference
logic in common_shapes.py handled this correct, which is why we didn't
notice this earlier. In particular, list ops use DT_VARIANT with
handle data.
PiperOrigin-RevId: 195739586

6 years agoRefactor TensorArray to avoid copies and memory allocations when executing eagerly.
Akshay Agrawal [Mon, 7 May 2018 23:16:24 +0000 (16:16 -0700)]
Refactor TensorArray to avoid copies and memory allocations when executing eagerly.

With this change, writes to TensorArrays when eager execution is enabled take O(1) time instead of O(n). Additionally, whereas writing to a TensorArray when constructing a graph results in allocating a new Python TensorArray object, writing to a TensorArray with eager enabled no longer performs that allocation (graph construction uses these allocations to ensure correctness of control flow and gradients, but this isn't necessary when executing eagerly). Finally, this change also removes the artificial write-once semantics of TensorArrays when executing eagerly.

PiperOrigin-RevId: 195739572

6 years ago[XLA] Make post order a possible schedule as it sometimes uses less memory than
Blake Hechtman [Mon, 7 May 2018 22:58:29 +0000 (15:58 -0700)]
[XLA] Make post order a possible schedule as it sometimes uses less memory than
the DFS or list scheduler and it is very simple.

PiperOrigin-RevId: 195736916

6 years ago[Remote functions] Only set the default runner *after* resolving the remote FLR.
Derek Murray [Mon, 7 May 2018 22:51:05 +0000 (15:51 -0700)]
[Remote functions] Only set the default runner *after* resolving the remote FLR.

Previously, if the `runner` was not specified for a function
execution, we would immediately set it to the default runner of the
*local* FLR, even if the function was to be executed remotely. This
change postpones the resolution of the default runner until after the
function invocation has been routed to the FLR that will actually
execute it.

As a result, we avoid the pathological case where a GPU device using a
private threadpool (TF_GPU_THREAD_MODE=gpu_private) ends up running
all of the ops for the CPU-side input pipeline on the private
threadpool.

PiperOrigin-RevId: 195735734

6 years agoAdd EvaluateNodes to tests: RemoveIdentityTransposesMultipleOutputs, RemoveTransposes...
A. Unique TensorFlower [Mon, 7 May 2018 22:47:57 +0000 (15:47 -0700)]
Add EvaluateNodes to tests: RemoveIdentityTransposesMultipleOutputs, RemoveTransposesWithControlDependency, CombineBitcasts, CombineAndRemoveBitcasts, RemoveRedundantCast

PiperOrigin-RevId: 195735234

6 years ago[XLA] Fix a "we're we're" in the operation semantics.
Michael Kuperstein [Mon, 7 May 2018 22:41:52 +0000 (15:41 -0700)]
[XLA] Fix a "we're we're" in the operation semantics.

PiperOrigin-RevId: 195734316

6 years agoAdd support for select (via tf.where) support to tflite.
A. Unique TensorFlower [Mon, 7 May 2018 22:41:22 +0000 (15:41 -0700)]
Add support for select (via tf.where) support to tflite.

PiperOrigin-RevId: 195734246

6 years agoInternal Change.
Michael Case [Mon, 7 May 2018 22:24:02 +0000 (15:24 -0700)]
Internal Change.

PiperOrigin-RevId: 195731675

6 years agoFix TypeError in update_version.py
Michael Case [Mon, 7 May 2018 22:20:49 +0000 (15:20 -0700)]
Fix TypeError in update_version.py

PiperOrigin-RevId: 195731183

6 years agoAdd support for tf.data.Dataset iterators in model training/eval methods in eager...
Pavithra Vijay [Mon, 7 May 2018 22:16:59 +0000 (15:16 -0700)]
Add support for tf.data.Dataset iterators in model training/eval methods in eager-mode

PiperOrigin-RevId: 195730534

6 years agoReplace references to TensorInfo with XlaTensor
Igor Ganichev [Mon, 7 May 2018 22:15:01 +0000 (15:15 -0700)]
Replace references to TensorInfo with XlaTensor

PiperOrigin-RevId: 195730139

6 years agoDisable automated testing of tensorflow/compiler/tests:extract_image_patches_op_test_...
A. Unique TensorFlower [Mon, 7 May 2018 21:55:26 +0000 (14:55 -0700)]
Disable automated testing of tensorflow/compiler/tests:extract_image_patches_op_test_cpu_ondemand

A recent change has made this test flaky.

PiperOrigin-RevId: 195726647

6 years agoAllow output has a different shape from input in the image.transform (#17011).
A. Unique TensorFlower [Mon, 7 May 2018 21:34:11 +0000 (14:34 -0700)]
Allow output has a different shape from input in the image.transform (#17011).

PiperOrigin-RevId: 195723288

6 years agoFix resource variable in cond gradient.
Alexandre Passos [Mon, 7 May 2018 21:29:03 +0000 (14:29 -0700)]
Fix resource variable in cond gradient.

PiperOrigin-RevId: 195722449

6 years ago[XLA] Shard compilation of HloEvaluator.
Justin Lebar [Mon, 7 May 2018 21:23:19 +0000 (14:23 -0700)]
[XLA] Shard compilation of HloEvaluator.

PiperOrigin-RevId: 195721404

6 years agoMove PadV2Options to the end, in order to maintain schema compatibility.
A. Unique TensorFlower [Mon, 7 May 2018 21:15:37 +0000 (14:15 -0700)]
Move PadV2Options to the end, in order to maintain schema compatibility.

PiperOrigin-RevId: 195720133

6 years agoUse 64bit aggregation for gradients and hessians since the 32 bit version is numerica...
A. Unique TensorFlower [Mon, 7 May 2018 21:13:23 +0000 (14:13 -0700)]
Use 64bit aggregation for gradients and hessians since the 32 bit version is numerically unstable for large minibatches.

PiperOrigin-RevId: 195719795

6 years agoChange deprecation_version field in api_def.proto to a string.
Anna R [Mon, 7 May 2018 21:03:15 +0000 (14:03 -0700)]
Change deprecation_version field in api_def.proto to a string.

PiperOrigin-RevId: 195718061

6 years ago[TF:XLA] Bump open source llvm revision to r331624
Sanjoy Das [Mon, 7 May 2018 21:00:09 +0000 (14:00 -0700)]
[TF:XLA] Bump open source llvm revision to r331624

PiperOrigin-RevId: 195717497

6 years agoFixes for accessing variables with a MirroredStrategy in a
A. Unique TensorFlower [Mon, 7 May 2018 20:24:12 +0000 (13:24 -0700)]
Fixes for accessing variables with a MirroredStrategy in a
cross-tower context:

* only provide read-only access to variables via get()

* don't fail if use the variable isn't copied to the current device in
  get()

* make _as_graph_element() return the aggregate value for tower-local
  variables (instead of the incorrect previous behavior of returning
  the primary)

PiperOrigin-RevId: 195711474

6 years agoSpecialize functions only once per unique context.
A. Unique TensorFlower [Mon, 7 May 2018 20:18:33 +0000 (13:18 -0700)]
Specialize functions only once per unique context.

PiperOrigin-RevId: 195710562

6 years agoRename HloDotWithContractDimsMatcher to HloDotWithContractingDimsMatcher
Sanjoy Das [Mon, 7 May 2018 19:48:38 +0000 (12:48 -0700)]
Rename HloDotWithContractDimsMatcher to HloDotWithContractingDimsMatcher

This is a typo I introduced in cr/195514907.

PiperOrigin-RevId: 195706006

6 years agoMake test tensorflow/python/keras:resnet50_test be size "medium"
A. Unique TensorFlower [Mon, 7 May 2018 19:41:07 +0000 (12:41 -0700)]
Make test tensorflow/python/keras:resnet50_test be size "medium"

This test sometimes runs longer than 60s, and has been getting flaky timeouts
as a result.  With a longer timeout, it succeeds reliably.

PiperOrigin-RevId: 195704998

6 years agoAdding Greater/GreaterEqual/LessEqual ops to complement Less.
A. Unique TensorFlower [Mon, 7 May 2018 19:37:36 +0000 (12:37 -0700)]
Adding Greater/GreaterEqual/LessEqual ops to complement Less.

PiperOrigin-RevId: 195704492

6 years agoAdd 'optonly' directive to linear_operator_circulant tests.
Ian Langmore [Mon, 7 May 2018 19:17:02 +0000 (12:17 -0700)]
Add 'optonly' directive to linear_operator_circulant tests.

PiperOrigin-RevId: 195701399

6 years ago[TF:XLA:GPU] Allow the use of linear address when there are size one dimensions
Bixia Zheng [Mon, 7 May 2018 19:15:52 +0000 (12:15 -0700)]
[TF:XLA:GPU] Allow the use of linear address when there are size one dimensions
in a tensor.

The current implementation of EmitArrayElementAddress incorrectly concludes
that having a size one dimension in a tensor indicates broadcasting is needed
and the linear address can't be used to access the tensor. We fix this by
leaving LinearValidOnShape to decide whether the linear address can be used to
access the tensor. This enables the vectorization of loads/stores in unrolled
elementwise op kernels when other criteria are met.

Add a test case.

PiperOrigin-RevId: 195701194

6 years ago[XLA] Add FusionKind matcher to pattern_matcher.h.
Justin Lebar [Mon, 7 May 2018 19:10:15 +0000 (12:10 -0700)]
[XLA] Add FusionKind matcher to pattern_matcher.h.

PiperOrigin-RevId: 195700319

6 years ago[tf.data] Patch to unref iterator_resource in DeserializeIteratorOp.
Shivani Agrawal [Mon, 7 May 2018 19:03:20 +0000 (12:03 -0700)]
[tf.data] Patch to unref iterator_resource in DeserializeIteratorOp.

PiperOrigin-RevId: 195698980

6 years agoDisable autograph cfg_test in windows.
Gunhan Gulsoy [Mon, 7 May 2018 18:52:46 +0000 (11:52 -0700)]
Disable autograph cfg_test in windows.

PiperOrigin-RevId: 195697446

6 years agoRegister bool scatter_update for resource variables
Alexandre Passos [Mon, 7 May 2018 18:49:08 +0000 (11:49 -0700)]
Register bool scatter_update for resource variables
Fixes #17784

PiperOrigin-RevId: 195696915

6 years agoGeneralize the input to TPU distribution strategy. Add cross-shard-replica sum.
Igor Saprykin [Mon, 7 May 2018 18:48:31 +0000 (11:48 -0700)]
Generalize the input to TPU distribution strategy.  Add cross-shard-replica sum.

TPUStrategy passes tests in minimize_loss_test.  That caused me to add a capability to have `iterations x cores` inputs of any structure.  I also resolved a big number of small issues and uncovered more things to resolve that are documented as todos.

PiperOrigin-RevId: 195696833

6 years agoRelease notes for TensorFlow Lite.
Yu-Cheng Ling [Mon, 7 May 2018 18:27:02 +0000 (11:27 -0700)]
Release notes for TensorFlow Lite.

PiperOrigin-RevId: 195693362

6 years agoExtracts PartialConcatConstFolding into a method.
A. Unique TensorFlower [Mon, 7 May 2018 18:11:28 +0000 (11:11 -0700)]
Extracts PartialConcatConstFolding into a method.

PiperOrigin-RevId: 195690333

6 years agoAdd tests for broadcasting KL divergence calculations.
A. Unique TensorFlower [Mon, 7 May 2018 18:09:47 +0000 (11:09 -0700)]
Add tests for broadcasting KL divergence calculations.

PiperOrigin-RevId: 195690035

6 years agoReplaced calls to tensorflow::StringPiece::ToString with std::string conversions.
A. Unique TensorFlower [Mon, 7 May 2018 18:05:56 +0000 (11:05 -0700)]
Replaced calls to tensorflow::StringPiece::ToString with std::string conversions.
That is, instances of sp.ToString() are replaced with std::string(sp).

This will allow tensorflow::StringPiece::ToString to be removed, which is necessary before it can be replaced with absl::string_view.

PiperOrigin-RevId: 195689392

6 years agoExtend block sparsity support for TPUs
A. Unique TensorFlower [Mon, 7 May 2018 17:49:26 +0000 (10:49 -0700)]
Extend block sparsity support for TPUs

PiperOrigin-RevId: 195685740

6 years agoAdd EvaluateNodes to HoistFactorDiv test.
A. Unique TensorFlower [Mon, 7 May 2018 17:47:35 +0000 (10:47 -0700)]
Add EvaluateNodes to HoistFactorDiv test.

PiperOrigin-RevId: 195685340

6 years agoRemoving quotations to fix the broken link found on https://www.tensorflow.org/progra...
A. Unique TensorFlower [Mon, 7 May 2018 17:42:13 +0000 (10:42 -0700)]
Removing quotations to fix the broken link found on https://tensorflow.org/programmers_guide/embedding

The link at "Follow this link to see a fun example of thumbnail images in the Embedding Projector." It should go to https://www.tensorflow.org/images/embedding-mnist.mp4 but instead goes to the TF index page.

PiperOrigin-RevId: 195684456

6 years agoInternal Change
Abhijit Karmarkar [Mon, 7 May 2018 17:27:50 +0000 (10:27 -0700)]
Internal Change

PiperOrigin-RevId: 195681946

6 years agoControl flow graph with forward and backward analysis
A. Unique TensorFlower [Mon, 7 May 2018 13:27:43 +0000 (06:27 -0700)]
Control flow graph with forward and backward analysis

PiperOrigin-RevId: 195654450

6 years agoAutomated g4 rollback of changelist 195638795
A. Unique TensorFlower [Mon, 7 May 2018 11:25:03 +0000 (04:25 -0700)]
Automated g4 rollback of changelist 195638795

PiperOrigin-RevId: 195645734

6 years agoImprove fusion logic of (a dot b) * alpha
A. Unique TensorFlower [Mon, 7 May 2018 10:00:34 +0000 (03:00 -0700)]
Improve fusion logic of (a dot b) * alpha

The previous approach didn't work because a multiplication by a scalar value
will be changed into an explicit broadcast.
Another issue that is fixed in this CL is retrieving the constant value from
the literal. This depends on the PrimitiveType, before we always assumed it to be double.
Also when checking ImplementedAsGemm() we should not call it recursively, but instead just the check related to kDot.
Finally add an execution test and adjust the fusion logic test.

PiperOrigin-RevId: 195638795

6 years agoRemove unused threadpool from stream executor.
A. Unique TensorFlower [Mon, 7 May 2018 08:57:29 +0000 (01:57 -0700)]
Remove unused threadpool from stream executor.

PiperOrigin-RevId: 195632175

6 years agoGracefully handle workers without heartbeat support enabled.
Russell Power [Sun, 6 May 2018 01:55:28 +0000 (18:55 -0700)]
Gracefully handle workers without heartbeat support enabled.

PiperOrigin-RevId: 195560525

6 years ago[XLA:GPU] Zero out input buffers before running cudnn conv autotune.
Justin Lebar [Sat, 5 May 2018 19:34:32 +0000 (12:34 -0700)]
[XLA:GPU] Zero out input buffers before running cudnn conv autotune.

We don't need a corresponding change in gemm_thunk.cc because for gemms,
we do our autotune at runtime, at which point we have some real data in
our input/output buffers.

PiperOrigin-RevId: 195548896

6 years ago[TPU] Add option to only compile a replicated graph.
Jacques Pienaar [Sat, 5 May 2018 19:02:32 +0000 (12:02 -0700)]
[TPU] Add option to only compile a replicated graph.

Useful when wanting to compile a computation but not run it. Returns a serialized CompilationResult string with the error message.

PiperOrigin-RevId: 195547847

6 years agoAllow benchmark model graph to be specified in text proto format.
Shashi Shekhar [Sat, 5 May 2018 18:55:53 +0000 (11:55 -0700)]
Allow benchmark model graph to be specified in text proto format.

PiperOrigin-RevId: 195547670

6 years ago[XLA] Always be willing to duplicate widening kConvert instructions during fusion.
Justin Lebar [Sat, 5 May 2018 18:24:06 +0000 (11:24 -0700)]
[XLA] Always be willing to duplicate widening kConvert instructions during fusion.

This has the effect of pushing widening kConvert HLOs into consumers.
This is what we want, because it means that the producer writes the
narrower type (e.g. f16) and the consumer reads it and internally
upcasts to the wider type (e.g. f32).  This lets the producer and
consumer both run faster, because they have to touch less memory.

PiperOrigin-RevId: 195546910

6 years agoPart 2 of Swift<->TF sends/recvs: support receiving tensors in TF from
Mingsheng Hong [Sat, 5 May 2018 14:10:58 +0000 (07:10 -0700)]
Part 2 of Swift<->TF sends/recvs: support receiving tensors in TF from
Swift via direct session.

The changes are:

1. Added a TF experimental C API for Swift host to enqueue a tensor for sending
  to TF. Again, the C APIs can be removed once the Fifo-queue based design
  proves stable later.

2. TFLowerGraph is extended to generate Fifo related nodes for TF to receive
  tensors. This is similar to the extension for TF to send tensors.

3. TFPartition is extended to support host send (createHostSend()), which does
  the tensor send via a new protocol method TensorSendableReceivable.sendToDevice().

The main complexity is in sending a scalar, where a new protocol method
TensorSendableReceivable.createScalarTensor() is called to first create a tensor
out of it, and then send it over to TF.

Also removed code for protocol conformance on AccelerableByTensorFlow. Instead
have compiler look up that conformance from the SILFunction on sending/receiving
tensors.

AccelerableByTensorFlow could be removed from the compiler-known protocol list
now, but we'll defer that till things can stabilized more (in the past this
protocol has been added to and removed from the list at different times).

PiperOrigin-RevId: 195539436

6 years agoRemove uses of the kTransposeDot fusion
Sanjoy Das [Sat, 5 May 2018 05:04:20 +0000 (22:04 -0700)]
Remove uses of the kTransposeDot fusion

I didn't remove the enum itself, but after this change removing the enum should
be a simple NFC change (famous last words!).

This will make it easier to implement BatchDot on CPU.

The change removes usages of kTransposeDot by:

 - Teaching TransposeFolding to "fuse" transposes into dots by flipping the
   lhs_contracting_dims/rhs_contracting_dims fields.

 - Replacing the notion of transpose_lhs/transpose_rhs in the IR emitters with
   "has a non-canonical LHS contraction dimension"/"has a non-canonical RHS
   contraction dimension" where the canonical LHS and RHS contraction dims [0]
   are 1 and 0.

Some tests were getting away with creating Dot instructions with their
dimensions numbers unset.  I've fixed these to create canonical dot operations
instead.

It is possible (but hard to tell without trying) that some of the IR emission
logic and Eigen runtime calls can now be simplified further.  For instance,
instead of passing in a `transpose_lhs` and `transpose_rhs` to the Eigen GEMM
routines, we could instead pass in the LHS and RHS contraction dimensions
directly.

[0] See HloInstruction::CreateCanonicalDot.

PiperOrigin-RevId: 195514907

6 years agoFix landscape layout.
Shashi Shekhar [Sat, 5 May 2018 02:18:44 +0000 (19:18 -0700)]
Fix landscape layout.

PiperOrigin-RevId: 195506194

6 years agoOVIC Benchmarker App (current without the functionality to bind to a CPU).
A. Unique TensorFlower [Sat, 5 May 2018 01:49:08 +0000 (18:49 -0700)]
OVIC Benchmarker App (current without the functionality to bind to a CPU).

PiperOrigin-RevId: 195503895

6 years agoadd support for PadV2
A. Unique TensorFlower [Sat, 5 May 2018 01:49:08 +0000 (18:49 -0700)]
add support for PadV2

PiperOrigin-RevId: 195503894

6 years agoCheckpointable: A small utility for exempting objects from __setattr__ tracking
Allen Lavoie [Sat, 5 May 2018 01:25:18 +0000 (18:25 -0700)]
Checkpointable: A small utility for exempting objects from __setattr__ tracking

Exposes it as tf.contrib.checkpoint.NoDependency. Objects wrapped in a
NoDependency object get unwrapped in __setattr__ and not tracked.

Removes the _save_counter dependency from tf.train.Checkpoint (the save counter
is still tracked as "save_counter" and always has been, so this is a
backwards-compatible dependency removal).

PiperOrigin-RevId: 195502562

6 years agoGuaranteeConst is a NoOp for the op_level_cost_estiamtor.
Max Galkin [Sat, 5 May 2018 01:17:45 +0000 (18:17 -0700)]
GuaranteeConst is a NoOp for the op_level_cost_estiamtor.

PiperOrigin-RevId: 195501990

6 years agoInternal Change
A. Unique TensorFlower [Sat, 5 May 2018 01:07:41 +0000 (18:07 -0700)]
Internal Change

PiperOrigin-RevId: 195501342

6 years agoGo: Update generated wrapper functions for TensorFlow ops.
A. Unique TensorFlower [Sat, 5 May 2018 00:46:13 +0000 (17:46 -0700)]
Go: Update generated wrapper functions for TensorFlow ops.
PiperOrigin-RevId: 195499636

6 years agoUpdate ops-related pbtxt files.
A. Unique TensorFlower [Sat, 5 May 2018 00:18:35 +0000 (17:18 -0700)]
Update ops-related pbtxt files.

PiperOrigin-RevId: 195497084

6 years ago[tf.data] Adding `num_parallel_calls` to `map_and_batch`.
Jiri Simsa [Sat, 5 May 2018 00:03:52 +0000 (17:03 -0700)]
[tf.data] Adding `num_parallel_calls` to `map_and_batch`.

PiperOrigin-RevId: 195495206

6 years agoAdd infrastructure for a backend-specific configuration for each op. This is intentio...
Bjarke Hammersholt Roune [Fri, 4 May 2018 23:51:06 +0000 (16:51 -0700)]
Add infrastructure for a backend-specific configuration for each op. This is intentionally not exposed in ComputationBuilder and is not intended for use or to be set at all prior to the last backend-specific part of compilation.

PiperOrigin-RevId: 195493500

6 years agoExpose read-only versions of tensors in tflite.
A. Unique TensorFlower [Fri, 4 May 2018 23:37:27 +0000 (16:37 -0700)]
Expose read-only versions of tensors in tflite.

PiperOrigin-RevId: 195491701

6 years agoAdd the ability to export separate SavedModels for train and eval mode to Estimator...
Karmel Allison [Fri, 4 May 2018 23:01:02 +0000 (16:01 -0700)]
Add the ability to export separate SavedModels for train and eval mode to Estimator with two new methods, available in tf.contrib: export_all_saved_models and export_saved_model_for_mode.

PiperOrigin-RevId: 195485922

6 years ago[XLA] Print allowed attributes when the user specifies an invalid attr.
Justin Lebar [Fri, 4 May 2018 22:40:07 +0000 (15:40 -0700)]
[XLA] Print allowed attributes when the user specifies an invalid attr.

PiperOrigin-RevId: 195482974

6 years agoIdentify and prune nodes that can never be executed
Benoit Steiner [Fri, 4 May 2018 22:14:00 +0000 (15:14 -0700)]
Identify and prune nodes that can never be executed

PiperOrigin-RevId: 195478951

6 years agoFix build failure for macos py3
James Qin [Fri, 4 May 2018 21:53:58 +0000 (14:53 -0700)]
Fix build failure for macos py3

PiperOrigin-RevId: 195475780

6 years ago[XLA:GPU] Mark floating-point division as an inexpensive op.
Justin Lebar [Fri, 4 May 2018 21:40:02 +0000 (14:40 -0700)]
[XLA:GPU] Mark floating-point division as an inexpensive op.

"Expensive" really means "so expensive you'd choose not to fuse in order
to avoid doing it twice".  FP division definitely isn't that expensive.

PiperOrigin-RevId: 195473524

6 years agoSome fixes to support another TF graph:
A. Unique TensorFlower [Fri, 4 May 2018 19:28:42 +0000 (12:28 -0700)]
Some fixes to support another TF graph:
1. Fix ResolveBatchNormalization to avoid deleting arrays that may still be
used.
2. Correctly count the number of ops using a given array, even when some ops
use the same array as more than one of their inputs.
3. In PropagateFixedSizes for Concatenation ops, when resolving a -1 wildcard
to a fixed value, we were doing so in a local 'axis' variable without actually
updating op->axis! The resulting -1 value still in op->axis tripped runtime code,
causing the concatenation to misbehave during inference.

PiperOrigin-RevId: 195454037

6 years ago[XLA] Redesign: migrate the SWIG wrapped xla client.
A. Unique TensorFlower [Fri, 4 May 2018 18:40:20 +0000 (11:40 -0700)]
[XLA] Redesign: migrate the SWIG wrapped xla client.
Added LocalOp that wraps XlaOp, so that it's fully visible to swig.

PiperOrigin-RevId: 195446939

6 years ago[XLA] Cleanup client_library_test_base: move definition of CreateParameterAndTransfer...
A. Unique TensorFlower [Fri, 4 May 2018 18:40:01 +0000 (11:40 -0700)]
[XLA] Cleanup client_library_test_base: move definition of CreateParameterAndTransferLiteral to .cc file

PiperOrigin-RevId: 195446864

6 years agoAdd operations before Identity operations should be quantized.
Suharsh Sivakumar [Fri, 4 May 2018 18:17:17 +0000 (11:17 -0700)]
Add operations before Identity operations should be quantized.

Fixes #19014

PiperOrigin-RevId: 195443326

6 years agoInternal clean up: change scanf to use int64_t instead of int64
A. Unique TensorFlower [Fri, 4 May 2018 17:50:29 +0000 (10:50 -0700)]
Internal clean up: change scanf to use int64_t instead of int64

PiperOrigin-RevId: 195438212

6 years agoImprove broadcast add implementation.
A. Unique TensorFlower [Fri, 4 May 2018 17:47:38 +0000 (10:47 -0700)]
Improve broadcast add implementation.

PiperOrigin-RevId: 195437679

6 years agoTFTS: Make it easier to swap in different autoregressive models.
Allen Lavoie [Fri, 4 May 2018 17:37:42 +0000 (10:37 -0700)]
TFTS: Make it easier to swap in different autoregressive models.

Adds a very simple LSTM encoder/decoder option as an example.

ARModel's new constructor argument is a bit awkward, since Estimator's new graphs mean we need a Model factory rather than a Model (or to un-build the model?). It's still a much more pleasant way to write autoregressive models than fiddling with ARModel directly, since ARModel handles collecting all the features (and the prediction loop, etc.). Happy to hear other ideas for an API.

PiperOrigin-RevId: 195436186

6 years agoImplement neg op
Alan Chiao [Fri, 4 May 2018 17:31:01 +0000 (10:31 -0700)]
Implement neg op

PiperOrigin-RevId: 195435079

6 years agoChange RecvTensor RPC implementation to use DeviceContext::CopyDeviceTensorToCPU...
Peter Hawkins [Fri, 4 May 2018 17:18:46 +0000 (10:18 -0700)]
Change RecvTensor RPC implementation to use DeviceContext::CopyDeviceTensorToCPU rather than calling GPUUtil::CopyGPUTensorToCPU. The direct call into the GPU code is problematic for non-GPU devices.

PiperOrigin-RevId: 195433287

6 years ago[XLA] Remove template keyword on non-template methods.
Benjamin Kramer [Fri, 4 May 2018 10:43:00 +0000 (03:43 -0700)]
[XLA] Remove template keyword on non-template methods.

This is an error with clang trunk.

PiperOrigin-RevId: 195394277

6 years agoFix HloSharding::GetSubSharding to return correct array shardings
A. Unique TensorFlower [Fri, 4 May 2018 10:27:18 +0000 (03:27 -0700)]
Fix HloSharding::GetSubSharding to return correct array shardings

Previously it always returned a tuple sharding even if the specified
index was referenceing a non-tuple element.

PiperOrigin-RevId: 195393313

6 years agoDo not crash on ROOT outfeed operations.
A. Unique TensorFlower [Fri, 4 May 2018 09:22:14 +0000 (02:22 -0700)]
Do not crash on ROOT outfeed operations.

PiperOrigin-RevId: 195388075

6 years agoFixing some linter errors in TF documentation (Github > GitHub, the the > the).
A. Unique TensorFlower [Fri, 4 May 2018 09:04:33 +0000 (02:04 -0700)]
Fixing some linter errors in TF documentation (Github > GitHub, the the > the).

PiperOrigin-RevId: 195386172

6 years agoPrefer non-nested GradientTape.gradient call when only one source is passed.
Tom Hennigan [Fri, 4 May 2018 08:57:02 +0000 (01:57 -0700)]
Prefer non-nested GradientTape.gradient call when only one source is passed.

PiperOrigin-RevId: 195385406

6 years ago * Don't copy on-host and on-device shapes locally.
A. Unique TensorFlower [Fri, 4 May 2018 08:47:12 +0000 (01:47 -0700)]
  * Don't copy on-host and on-device shapes locally.
  * Use ForEachMutableElement rather than the iterators, as it is much quicker.

There is still room for improvement; ForEachMutableElement is linear in the number of nodes in the shape tree but we want to be linear in the number of nodes in the sub shape tree. But I feel this is a good enough improvement.

PiperOrigin-RevId: 195384423