platform/upstream/tensorflow.git
6 years agoAdd a utility to visualize object-based checkpoints
Allen Lavoie [Fri, 20 Apr 2018 19:40:57 +0000 (12:40 -0700)]
Add a utility to visualize object-based checkpoints

Useful for generating a warm fuzzy feeling that everything you think should be saved was saved, and for explaining what object-based checkpointing is. (Also useful on the former front will be a planned "assert that all of this Graph's trainable variables are accessible from object X" function.)

Somewhat hacky since it generates strings rather than using the pydot bindings (and so works without a pydot dependency).

PiperOrigin-RevId: 193708003

6 years agoChanged tf_to_tflite build rule.
A. Unique TensorFlower [Fri, 20 Apr 2018 19:37:39 +0000 (12:37 -0700)]
Changed tf_to_tflite build rule.

PiperOrigin-RevId: 193707628

6 years agoAdd native dilated support for conv3d and its gradients in cudnn v>=6.
Yangzihao Wang [Fri, 20 Apr 2018 19:05:11 +0000 (12:05 -0700)]
Add native dilated support for conv3d and its gradients in cudnn v>=6.

PiperOrigin-RevId: 193703316

6 years ago[TF:XLA] Now that the compiler no longer introduces implicit broadcasts, forbid them...
Dimitris Vardoulakis [Fri, 20 Apr 2018 19:02:56 +0000 (12:02 -0700)]
[TF:XLA] Now that the compiler no longer introduces implicit broadcasts, forbid them in the HLO verifier.

PiperOrigin-RevId: 193702874

6 years agoPrefix compat import with underscore in meta_graph_transform.py so that it
Anna R [Fri, 20 Apr 2018 19:01:21 +0000 (12:01 -0700)]
Prefix compat import with underscore in meta_graph_transform.py so that it
doesn't get exported as part of API: https://www.tensorflow.org/versions/r1.8/api_docs/python/tf/contrib/meta_graph_transform/meta_graph_transform

PiperOrigin-RevId: 193702570

6 years ago[XLA] Fix a bug in ToProto: don't add gather attributes twice.
A. Unique TensorFlower [Fri, 20 Apr 2018 18:43:48 +0000 (11:43 -0700)]
[XLA] Fix a bug in ToProto: don't add gather attributes twice.

PiperOrigin-RevId: 193699745

6 years agoMove the guts of TFE_Op into EagerOperation
Akshay Modi [Fri, 20 Apr 2018 18:34:55 +0000 (11:34 -0700)]
Move the guts of TFE_Op into EagerOperation

PiperOrigin-RevId: 193698320

6 years agoAllow creating tensors from numpy arrays, and other various constants - try #2
Akshay Modi [Fri, 20 Apr 2018 18:24:53 +0000 (11:24 -0700)]
Allow creating tensors from numpy arrays, and other various constants - try #2

Allow type-inference from a different input tensor, similar to args_to_matching_eager.

- Update TFE_Py_TensorShapeSlice to take tuples.
- Update int values to allow int/long in py2
END_PUBLIC

BEGIN_PUBLIC
Automated g4 rollback of changelist 192184809

PiperOrigin-RevId: 193696790

6 years agoTemporarily set cudnn Rnn math precision to fp32.
James Qin [Fri, 20 Apr 2018 18:23:29 +0000 (11:23 -0700)]
Temporarily set cudnn Rnn math precision to fp32.

Problem:
When calling cudnnGetRNNLinLayerMatrixParams(), return error CUDNN_STATUS_BAD_PARAM if:

* RNN descriptor set math precision = CUDNN_DATA_FLOAT
* input descriptor dataType = CUDNN_DATA_HALF
* weight descriptor dataType= CUDNN_DATA_HALF

If updating Rnn descriptor math precision to CUDNN_DATA_HALF, then no error.

cudnn 7.1.4 will fix the problem.

PiperOrigin-RevId: 193696566

6 years agoSpeed up safe_strtod and safe_strtof functions by using double-conversion library
akindyakov [Fri, 20 Apr 2018 18:23:15 +0000 (11:23 -0700)]
Speed up safe_strtod and safe_strtof functions by using double-conversion library
Closes #12102.

PiperOrigin-RevId: 193696537

6 years agoUse CreateWorkerSession and DeleteWorkerSession for all distributed sessions.
Derek Murray [Fri, 20 Apr 2018 18:13:16 +0000 (11:13 -0700)]
Use CreateWorkerSession and DeleteWorkerSession for all distributed sessions.

This change adds a phase to the session creation protocol: the master now contacts all workers to register a session handle and create a "WorkerSession" on each worker before it first registers or runs a graph on any worker. Subsequent requests to a worker ensure that the worker has the session handle registered before performing the request, and an AbortedError is raised if the worker has not (e.g. because it restarted after a failure).

As a result, more failure cases are covered by the high-level APIs (tf.estimator, Slim, etc.) that recreate the session on receiving an AbortedError. Previously, there was a possible race condition in which a PS task could restart between variable initialization and the first step, leading to a FailedPreconditionError ("Attempting to use uninitialized value") that would not be handled by the high-level APIs.

PiperOrigin-RevId: 193694958

6 years ago[XLA:CPU] Use Eigen for F64 dot operations
Sanjoy Das [Fri, 20 Apr 2018 18:10:56 +0000 (11:10 -0700)]
[XLA:CPU] Use Eigen for F64 dot operations

PiperOrigin-RevId: 193694613

6 years agoDisable constant folding and arithmetic optimizations for functions.
A. Unique TensorFlower [Fri, 20 Apr 2018 17:34:49 +0000 (10:34 -0700)]
Disable constant folding and arithmetic optimizations for functions.

PiperOrigin-RevId: 193688466

6 years agoAdding Shape inference functions to outfeed enqueue ops.
Jacques Pienaar [Fri, 20 Apr 2018 17:32:24 +0000 (10:32 -0700)]
Adding Shape inference functions to outfeed enqueue ops.

PiperOrigin-RevId: 193688099

6 years agoChange the TF record reader to use 16MB buffering by default in order to improve...
A. Unique TensorFlower [Fri, 20 Apr 2018 17:16:03 +0000 (10:16 -0700)]
Change the TF record reader to use 16MB buffering by default in order to improve performance.

PiperOrigin-RevId: 193685521

6 years agoCorrect error in "Adding An Op" docs.
James Keeling [Fri, 20 Apr 2018 16:54:50 +0000 (09:54 -0700)]
Correct error in "Adding An Op" docs.

The macro `REGISTER_KERNEL_BUILDER` always declared a functor specialized on floats, instead of the type actually passed into the macro.

PiperOrigin-RevId: 193682519

6 years agoDon't delete inbound_nodes and outbound_nodes, these no longer exist.
Mark Daoust [Fri, 20 Apr 2018 16:31:08 +0000 (09:31 -0700)]
Don't delete inbound_nodes and outbound_nodes, these no longer exist.

PiperOrigin-RevId: 193679512

6 years agoFix msan error in MapAndBatchDataset.
Saurabh Saxena [Fri, 20 Apr 2018 16:23:52 +0000 (09:23 -0700)]
Fix msan error in MapAndBatchDataset.
While checkpointing tensors in BatchResult.output save only the initialized slice.
If the final batch is short, the entire batch tensor may not be initialized.

PiperOrigin-RevId: 193678679

6 years ago[TF:XLA] Bump open source llvm revision to r330313
Sanjoy Das [Fri, 20 Apr 2018 16:20:36 +0000 (09:20 -0700)]
[TF:XLA] Bump open source llvm revision to r330313

PiperOrigin-RevId: 193678317

6 years agoEliminate the guard around Winograd non-fused convolutions with cudnn7.
A. Unique TensorFlower [Fri, 20 Apr 2018 15:08:01 +0000 (08:08 -0700)]
Eliminate the guard around Winograd non-fused convolutions with cudnn7.

PiperOrigin-RevId: 193669636

6 years agointernal change
A. Unique TensorFlower [Fri, 20 Apr 2018 13:36:56 +0000 (06:36 -0700)]
internal change

PiperOrigin-RevId: 193659701

6 years ago[TF:XLA] Implement ClipByValue.
Peter Hawkins [Fri, 20 Apr 2018 10:29:31 +0000 (03:29 -0700)]
[TF:XLA] Implement ClipByValue.

PiperOrigin-RevId: 193646890

6 years ago[TF:XLA] Factor out the handling of while instructions to make HloVerifier::Run shorter.
Dimitris Vardoulakis [Fri, 20 Apr 2018 06:35:04 +0000 (23:35 -0700)]
[TF:XLA] Factor out the handling of while instructions to make HloVerifier::Run shorter.

PiperOrigin-RevId: 193626864

6 years agoAutomated g4 rollback of changelist 193602050
Asim Shankar [Fri, 20 Apr 2018 06:08:53 +0000 (23:08 -0700)]
Automated g4 rollback of changelist 193602050

PiperOrigin-RevId: 193625346

6 years ago[XLA] Rework the local XLA client's Shape class with separate array and tuple shape...
Roy Frostig [Fri, 20 Apr 2018 06:01:07 +0000 (23:01 -0700)]
[XLA] Rework the local XLA client's Shape class with separate array and tuple shape constructors.

PiperOrigin-RevId: 193624591

6 years agoSkip tests with no_oss tag in XLA builds.
Yifei Feng [Fri, 20 Apr 2018 04:33:41 +0000 (21:33 -0700)]
Skip tests with no_oss tag in XLA builds.

PiperOrigin-RevId: 193619344

6 years agoAlways include the local worker in the list of filtered targets.
Derek Murray [Fri, 20 Apr 2018 03:00:21 +0000 (20:00 -0700)]
Always include the local worker in the list of filtered targets.

It is currently legal to specify a device filter that doesn't include the local worker.
In that case, the MasterSession includes all local devices regardless of the filter.
This change extends this behavior to the list of filtered workers, which will be crucial for backwards compatibility when we enable CreateWorkerSession for all MasterSessions, because we need to call CreateWorkerSession on all potential workers.

PiperOrigin-RevId: 193613313

6 years ago[XLA] Remove default argument on virtual function DeviceMemoryAllocator::Allocate().
Justin Lebar [Fri, 20 Apr 2018 02:27:34 +0000 (19:27 -0700)]
[XLA] Remove default argument on virtual function DeviceMemoryAllocator::Allocate().

Default args on virtual functions are disallowed by the Google style
guide, for good reason.  They have the extremely surprising behavior
that the defaults you get when calling a function on a pointer depend
not on the underlying type of the object, but on whatever is the
semantic type of the pointer!

PiperOrigin-RevId: 193611213

6 years agoSupport variable parameter structure in TPU distribution strategy.
Igor Saprykin [Fri, 20 Apr 2018 02:11:37 +0000 (19:11 -0700)]
Support variable parameter structure in TPU distribution strategy.

TPUStrategy is added to a few more tests.

There appears to be an issue with the batch norm test in minimize_loss_test where the moving averages stay at 0.  I'm trying to resolve that separately as the next CL.

PiperOrigin-RevId: 193610264

6 years agoAutomated g4 rollback of changelist 193593761
Justin Lebar [Fri, 20 Apr 2018 02:03:03 +0000 (19:03 -0700)]
Automated g4 rollback of changelist 193593761

PiperOrigin-RevId: 193609407

6 years agoInternal change.
Anna R [Fri, 20 Apr 2018 01:45:47 +0000 (18:45 -0700)]
Internal change.

PiperOrigin-RevId: 193608140

6 years ago[XLA] Fix a bug in the name_uniquer.
A. Unique TensorFlower [Fri, 20 Apr 2018 01:29:00 +0000 (18:29 -0700)]
[XLA] Fix a bug in the name_uniquer.

The problem happens because the name_uniquer stripped away the numeric suffix if it <=0. The solution is, if there was a numeric suffix, the result should also have a numeric suffix.

PiperOrigin-RevId: 193606838

6 years agoOptimize Graph function library.
A. Unique TensorFlower [Fri, 20 Apr 2018 01:17:02 +0000 (18:17 -0700)]
Optimize Graph function library.

PiperOrigin-RevId: 193605910

6 years agoRespect any device filters in {Create,Delete}WorkerSessions().
Derek Murray [Fri, 20 Apr 2018 01:12:57 +0000 (18:12 -0700)]
Respect any device filters in {Create,Delete}WorkerSessions().

This is another step towards enabling us to turn on explicit worker
sessions for all master sessions.

PiperOrigin-RevId: 193605565

6 years agointernal
Lukasz Kaiser [Fri, 20 Apr 2018 00:39:09 +0000 (17:39 -0700)]
internal
END_PUBLIC

BEGIN_PUBLIC
Automated g4 rollback of changelist 193571934

PiperOrigin-RevId: 193602050

6 years agoInternal testing changes
Austin Anderson [Fri, 20 Apr 2018 00:30:49 +0000 (17:30 -0700)]
Internal testing changes

PiperOrigin-RevId: 193601134

6 years agoUpdate DecodeProtoOp so that it returns explicitly specified default values for missi...
A. Unique TensorFlower [Fri, 20 Apr 2018 00:27:04 +0000 (17:27 -0700)]
Update DecodeProtoOp so that it returns explicitly specified default values for missing fields.

PiperOrigin-RevId: 193600735

6 years agoAdd support for non-Tensor args in recompute_grad
A. Unique TensorFlower [Fri, 20 Apr 2018 00:26:41 +0000 (17:26 -0700)]
Add support for non-Tensor args in recompute_grad

Previously, the function decorated by recompute_grad had to have a signature that contained only positional arguments, and all those arguments had to be Tensors. Most "layers" users define however have non-Tensor arguments (for example, various hyperparameters) and often have keyword arguments as well. This change allows a user to use whatever function signature they wish while being explicit about which arguments are Tensors.

PiperOrigin-RevId: 193600682

6 years agoMinor code refactoring.
Jiri Simsa [Fri, 20 Apr 2018 00:21:50 +0000 (17:21 -0700)]
Minor code refactoring.

PiperOrigin-RevId: 193600173

6 years ago[XLA] De-unique_ptr-ify ShapedBuffer and ScopedShapedBuffer.
Justin Lebar [Fri, 20 Apr 2018 00:18:10 +0000 (17:18 -0700)]
[XLA] De-unique_ptr-ify ShapedBuffer and ScopedShapedBuffer.

These are already notionally equivalent to T* and unique_ptr<T>, so
having a unique_ptr of a {Scoped,}ShapedBuffer is pretty redundant.

Also clean up the ScopedShapedBuffer API a bit.

PiperOrigin-RevId: 193599773

6 years agoRemoved deprecated methods from tensorflow::StringPiece.
A. Unique TensorFlower [Fri, 20 Apr 2018 00:17:05 +0000 (17:17 -0700)]
Removed deprecated methods from tensorflow::StringPiece.

This will allow tensorflow::StringPiece to be more easily replaced with absl::string_view as absl::string_view does not contain those methods.

PiperOrigin-RevId: 193599651

6 years agoRun EvaluateNodes for ModelPruner test except for NoPruning.
A. Unique TensorFlower [Thu, 19 Apr 2018 23:53:14 +0000 (16:53 -0700)]
Run EvaluateNodes for ModelPruner test except for NoPruning.

PiperOrigin-RevId: 193596812

6 years agoImprove error messages for LiteralTestUtil::Near.
Mark Heffernan [Thu, 19 Apr 2018 23:35:40 +0000 (16:35 -0700)]
Improve error messages for LiteralTestUtil::Near.
Previously error messages for mismatches were difficult to read with much of the space taken by useless stack traces. This CL cleans up the message considerably and adds additional information including statistics about the values and mismatches.

PiperOrigin-RevId: 193594593

6 years agoCleaned up the handling of merge nodes
Benoit Steiner [Thu, 19 Apr 2018 23:30:26 +0000 (16:30 -0700)]
Cleaned up the handling of merge nodes

PiperOrigin-RevId: 193593810

6 years agoMake xla/service:cpu_plugin depend on the StreamExecutor host platform.
Justin Lebar [Thu, 19 Apr 2018 23:30:02 +0000 (16:30 -0700)]
Make xla/service:cpu_plugin depend on the StreamExecutor host platform.

PiperOrigin-RevId: 193593761

6 years agoAutomated g4 rollback of changelist 193564222
A. Unique TensorFlower [Thu, 19 Apr 2018 22:56:17 +0000 (15:56 -0700)]
Automated g4 rollback of changelist 193564222

PiperOrigin-RevId: 193588935

6 years agoInternal-only change.
Justin Lebar [Thu, 19 Apr 2018 22:55:53 +0000 (15:55 -0700)]
Internal-only change.

PiperOrigin-RevId: 193588868

6 years agoSimplifying cols_to_vars update
Rohan Jain [Thu, 19 Apr 2018 22:32:37 +0000 (15:32 -0700)]
Simplifying cols_to_vars update

PiperOrigin-RevId: 193585237

6 years agoPrevent a bool field from being accessed when uninitialized.
Derek Murray [Thu, 19 Apr 2018 22:29:21 +0000 (15:29 -0700)]
Prevent a bool field from being accessed when uninitialized.

PiperOrigin-RevId: 193584746

6 years agoConvert a local variable and mutex to a struct so GUARDED_BY annotation works correctly.
A. Unique TensorFlower [Thu, 19 Apr 2018 22:27:19 +0000 (15:27 -0700)]
Convert a local variable and mutex to a struct so GUARDED_BY annotation works correctly.

PiperOrigin-RevId: 193584438

6 years agoOnly generate floating points that are fractions like n / 256, since they are RGB...
A. Unique TensorFlower [Thu, 19 Apr 2018 22:09:58 +0000 (15:09 -0700)]
Only generate floating points that are fractions like n / 256, since they are RGB pixels. This fixes RGBToHSVTest.testBatch on low-precision dtypes like bfloat16.

PiperOrigin-RevId: 193581652

6 years agodisabling opensource testing for failing xla test
Olivia Nordquist [Thu, 19 Apr 2018 21:59:25 +0000 (14:59 -0700)]
disabling opensource testing for failing xla test

PiperOrigin-RevId: 193579805

6 years agoReplaced calls to deprecated tensorflow::StringPiece methods with their
A. Unique TensorFlower [Thu, 19 Apr 2018 21:35:27 +0000 (14:35 -0700)]
Replaced calls to deprecated tensorflow::StringPiece methods with their
tensorflow::str_util equivalents.

This will allow the deprecated methods to be removed.

PiperOrigin-RevId: 193575992

6 years agoThe HLO element type converter must remove side effecting instructions like Rng
Sanjoy Das [Thu, 19 Apr 2018 21:30:27 +0000 (14:30 -0700)]
The HLO element type converter must remove side effecting instructions like Rng

The CPU backend does not know how to lower bf16 typed RNG nodes so even unused
instances of these can't remain in the HLO IR.
HloComputation::ReplaceInstruction keeps these Rng nodes around since it doesn't
remove side effecting nodes.
PiperOrigin-RevId: 193575183

6 years agoRecordReader: Simplify interface contract and implementation.
Asim Shankar [Thu, 19 Apr 2018 21:10:01 +0000 (14:10 -0700)]
RecordReader: Simplify interface contract and implementation.

Prior to this change, RecordReader had the following contract:
- Records can be read in any order, EXCEPT if compression or buffering was
  enabled.
- If the underlying file is being concurrently written to
  then calls to ReadRecord() may fail (because of an incomplete
  record near the end of a file), but a retry may succeed (once the
  record is written), EXCEPT if compression or buffering is enabled
  (in which case the failure will be terminal).

  This "retry-may-succeed" behavior is relied upon by tensorboard
  (https://github.com/tensorflow/tensorboard/blob/1.7/tensorboard/backend/event_processing/event_file_loader.py#L55)
  where one process (typically the model training process) is writing
  tf.summary events to an event file and another process (tensorboard)
  is concurrently reading it.

With this change, the intent is to remove the EXCEPTions and have the
same behavior irrespective of compression/buffering.

Additionally, fix a memory leak when ZlibInputStream::Reset() is invoked.

PiperOrigin-RevId: 193571934

6 years agoMake flat_transforms_to_matrices and matrices_to_flat_transforms public available.
A. Unique TensorFlower [Thu, 19 Apr 2018 21:05:06 +0000 (14:05 -0700)]
Make flat_transforms_to_matrices and matrices_to_flat_transforms public available.

PiperOrigin-RevId: 193571089

6 years agotools/lib_package: Fix typo in README
Asim Shankar [Thu, 19 Apr 2018 20:38:43 +0000 (13:38 -0700)]
tools/lib_package: Fix typo in README

PiperOrigin-RevId: 193566850

6 years agoAllow output has a different shape from input in the image.transform (#17011).
A. Unique TensorFlower [Thu, 19 Apr 2018 20:21:25 +0000 (13:21 -0700)]
Allow output has a different shape from input in the image.transform (#17011).

PiperOrigin-RevId: 193564222

6 years agoSupport various shapes in TPU DistributionStrategy.
Igor Saprykin [Thu, 19 Apr 2018 20:19:27 +0000 (13:19 -0700)]
Support various shapes in TPU DistributionStrategy.

PiperOrigin-RevId: 193563912

6 years agoCollective Ops Part 4
A. Unique TensorFlower [Thu, 19 Apr 2018 20:09:07 +0000 (13:09 -0700)]
Collective Ops Part 4

Add Broadcaster.
A few minor adjustments to CollectiveParams and RMA.

This change is part of a series of changes introducing infrastructure
for collective ops and initial implementations of reduction and broadcast.

PiperOrigin-RevId: 193562391

6 years agoMake PmfToQuantizedCdf op to make adjustments if the sum of quantized pmf is
Sung Jin Hwang [Thu, 19 Apr 2018 19:44:21 +0000 (12:44 -0700)]
Make PmfToQuantizedCdf op to make adjustments if the sum of quantized pmf is
less than 2**precision.

Prior to the change, the op did nothing when the sum of quantized pmf was less
than 2**precision. While the produced CDF was valid for range coders,
adjustments to CDF could be made to achieve better compression rate.

PiperOrigin-RevId: 193558740

6 years agoFix bug in ring_reducer.cc abort handling.
A. Unique TensorFlower [Thu, 19 Apr 2018 19:32:52 +0000 (12:32 -0700)]
Fix bug in ring_reducer.cc abort handling.

PiperOrigin-RevId: 193557334

6 years agoAdd support for Dataset Iterators in Model training/eval methods in graph mode.
Francois Chollet [Thu, 19 Apr 2018 18:58:04 +0000 (11:58 -0700)]
Add support for Dataset Iterators in Model training/eval methods in graph mode.

PiperOrigin-RevId: 193552275

6 years agoInternal cleanup.
A. Unique TensorFlower [Thu, 19 Apr 2018 18:57:48 +0000 (11:57 -0700)]
Internal cleanup.

PiperOrigin-RevId: 193552240

6 years agoAvoid looking up the shape functions multiple times
Benoit Steiner [Thu, 19 Apr 2018 18:57:36 +0000 (11:57 -0700)]
Avoid looking up the shape functions multiple times
Improved the handling of fed nodes

PiperOrigin-RevId: 193552210

6 years agoChange estimator to only log non-binary eval metrics, because logging binary metrics...
A. Unique TensorFlower [Thu, 19 Apr 2018 18:55:46 +0000 (11:55 -0700)]
Change estimator to only log non-binary eval metrics, because logging binary metrics such as images will lead to crash.

PiperOrigin-RevId: 193551927

6 years agoImplements linear_model using _LinearModel. Added support for cols_to_vars in _Linear...
Rohan Jain [Thu, 19 Apr 2018 18:52:10 +0000 (11:52 -0700)]
Implements linear_model using _LinearModel. Added support for cols_to_vars in _LinearModel in order to make this possible. Also, made some fixes so that variable names come out the same as before.

PiperOrigin-RevId: 193551353

6 years agoGo: Update generated wrapper functions for TensorFlow ops.
A. Unique TensorFlower [Thu, 19 Apr 2018 18:47:28 +0000 (11:47 -0700)]
Go: Update generated wrapper functions for TensorFlow ops.
PiperOrigin-RevId: 193550428

6 years agoUpdate ops-related pbtxt files.
A. Unique TensorFlower [Thu, 19 Apr 2018 18:21:21 +0000 (11:21 -0700)]
Update ops-related pbtxt files.

PiperOrigin-RevId: 193546050

6 years agoUpdating tests in constant_folding_test.cc so that they Evaluate the optimized and...
A. Unique TensorFlower [Thu, 19 Apr 2018 18:08:08 +0000 (11:08 -0700)]
Updating tests in constant_folding_test.cc so that they Evaluate the optimized and original graph and check if their outputs are the same.

PiperOrigin-RevId: 193543478

6 years agoFix open source BUILD bugs for cloud profiler. Increment
A. Unique TensorFlower [Thu, 19 Apr 2018 17:59:08 +0000 (10:59 -0700)]
Fix open source BUILD bugs for cloud profiler. Increment
version for releasing cloud_tpu_profiler 1.6 with pod profiling support.

PiperOrigin-RevId: 193541692

6 years agoAllow non-isolated worker sessions to borrow `WorkerEnv::device_mgr`.
Derek Murray [Thu, 19 Apr 2018 17:57:55 +0000 (10:57 -0700)]
Allow non-isolated worker sessions to borrow `WorkerEnv::device_mgr`.

Without this change, a shared resource (e.g. an Iterator) could not be
created in one session `s1`, and used in a later session `s2` after
`s1` was closed, because the iterator might indirectly capture devices
from the previous session, and use them after they are freed when the
`WorkerSession` was deleted.

The current change only affects the singleton "legacy" WorkerSession,
which is never deleted, but this is necessary to switch all sessions
to use separate WorkerSession objects.

PiperOrigin-RevId: 193541426

6 years agoFix undefined signed integer overflow by performing addition more carefully.
A. Unique TensorFlower [Thu, 19 Apr 2018 17:33:42 +0000 (10:33 -0700)]
Fix undefined signed integer overflow by performing addition more carefully.

PiperOrigin-RevId: 193537461

6 years agoAdd metric names to model.metrics_names in compile for keras models run in eager...
Anjali Sridhar [Thu, 19 Apr 2018 17:26:54 +0000 (10:26 -0700)]
Add metric names to model.metrics_names in compile for keras models run in eager execution. This prevents us from dropping metrics when we run model.evaluate.

PiperOrigin-RevId: 193536341

6 years agodistribution_util moved into its own BUILD target, so linear_operator can depend...
Ian Langmore [Thu, 19 Apr 2018 17:20:43 +0000 (10:20 -0700)]
distribution_util moved into its own BUILD target, so linear_operator can depend on it.

PiperOrigin-RevId: 193535400

6 years agoPrototype for tf.data writer API.
Jiri Simsa [Thu, 19 Apr 2018 17:14:09 +0000 (10:14 -0700)]
Prototype for tf.data writer API.

PiperOrigin-RevId: 193534333

6 years agoAdded support for concatenation and slicing of symbolic shapes
Benoit Steiner [Thu, 19 Apr 2018 17:05:08 +0000 (10:05 -0700)]
Added support for concatenation and slicing of symbolic shapes

PiperOrigin-RevId: 193532769

6 years agoCreate a skeleton tf.contrib.checkpoint.
Allen Lavoie [Thu, 19 Apr 2018 16:54:40 +0000 (09:54 -0700)]
Create a skeleton tf.contrib.checkpoint.

My plan for this is to incubate tools for working with object-based checkpoints:
  - Tools for managing dependency graphs, e.g. checkpointable lists/dictionaries
  - Inspecting/visualizing checkpoints
  - Listing variables and gathering initializers from a Checkpointable object
    and its dependencies
  - Verifying all variables are accessible as dependencies, which should make
    converting existing graph building Saver uses easier/safer.

This CL includes none of those things, it just moves the split_dependency tool
here instead of contrib/eager.

PiperOrigin-RevId: 193531292

6 years ago[tf.data] Add checkpointing support for MapAndBatchDataset.
Saurabh Saxena [Thu, 19 Apr 2018 16:37:20 +0000 (09:37 -0700)]
[tf.data] Add checkpointing support for MapAndBatchDataset.

PiperOrigin-RevId: 193528712

6 years agoDisable dlopen error of libneuralnetworks for non-Android platforms.
A. Unique TensorFlower [Thu, 19 Apr 2018 16:34:24 +0000 (09:34 -0700)]
Disable dlopen error of libneuralnetworks for non-Android platforms.

PiperOrigin-RevId: 193528346

6 years ago[TF:XLA] Change HloTestBase::ExecuteNoHloPasses to return a literal directly.
A. Unique TensorFlower [Thu, 19 Apr 2018 16:23:35 +0000 (09:23 -0700)]
[TF:XLA] Change HloTestBase::ExecuteNoHloPasses to return a literal directly.

PiperOrigin-RevId: 193526900

6 years agoPin pip to version 9.0.3
A. Unique TensorFlower [Thu, 19 Apr 2018 16:13:47 +0000 (09:13 -0700)]
Pin pip to version 9.0.3

* This is because pip 10 is still unstable in some distros
* reference: https://github.com/pypa/pip/issues/5240

PiperOrigin-RevId: 193525542

6 years agoTeach the conditinal simplifier about sharding.
A. Unique TensorFlower [Thu, 19 Apr 2018 14:13:03 +0000 (07:13 -0700)]
Teach the conditinal simplifier about sharding.

PiperOrigin-RevId: 193510638

6 years agoUpdate the Colorbot demo to use a Keras model in addition to the Estimator.
A. Unique TensorFlower [Thu, 19 Apr 2018 13:58:34 +0000 (06:58 -0700)]
Update the Colorbot demo to use a Keras model in addition to the Estimator.

PiperOrigin-RevId: 193508874

6 years agoSupport for converting entire class hierarchies:
A. Unique TensorFlower [Thu, 19 Apr 2018 12:11:30 +0000 (05:11 -0700)]
Support for converting entire class hierarchies:
 * limit the methods being converted to those that have not been inherited from the superclass
 * include the (possibly compiled) superclass in the definition of the compiled class
 * either mark the superclass for conversion or generate an absolute aliased import line, depending on whether it's whitelisted

PiperOrigin-RevId: 193499204

6 years agoHandle corner case in Python 3: members annotated with @classmethod.
A. Unique TensorFlower [Thu, 19 Apr 2018 11:27:38 +0000 (04:27 -0700)]
Handle corner case in Python 3: members annotated with @classmethod.

PiperOrigin-RevId: 193495506

6 years agoSet the random seed in on-demand mode.
A. Unique TensorFlower [Thu, 19 Apr 2018 09:58:31 +0000 (02:58 -0700)]
Set the random seed in on-demand mode.

PiperOrigin-RevId: 193488103

6 years agoUpdate feature_util's GetFeatures to show compile-time error for unsupported types...
A. Unique TensorFlower [Thu, 19 Apr 2018 08:56:31 +0000 (01:56 -0700)]
Update feature_util's GetFeatures to show compile-time error for unsupported types instead of a link-time error.

PiperOrigin-RevId: 193480683

6 years agoRemove proto import in header files for core/kernels/boosted_trees.
Yifei Feng [Thu, 19 Apr 2018 08:26:07 +0000 (01:26 -0700)]
Remove proto import in header files for core/kernels/boosted_trees.
Move implementations that requires declaration of TreeEnsemble to .cc files.

The goal is to make kernels mostly independent of proto headers, which will let us lock down our .so import

PiperOrigin-RevId: 193478404

6 years agodocs: Add a note on building the C and/or Java API binaries from source.
Asim Shankar [Thu, 19 Apr 2018 07:28:33 +0000 (00:28 -0700)]
docs: Add a note on building the C and/or Java API binaries from source.

See #15290

PiperOrigin-RevId: 193473886

6 years agoGo: Update generated wrapper functions for TensorFlow ops.
A. Unique TensorFlower [Thu, 19 Apr 2018 06:46:34 +0000 (23:46 -0700)]
Go: Update generated wrapper functions for TensorFlow ops.
PiperOrigin-RevId: 193471104

6 years agoUpdate ops-related pbtxt files.
A. Unique TensorFlower [Thu, 19 Apr 2018 06:18:50 +0000 (23:18 -0700)]
Update ops-related pbtxt files.

PiperOrigin-RevId: 193469437

6 years ago[tf.data] Add native implementation for `tf.contrib.data.unbatch()`.
Derek Murray [Thu, 19 Apr 2018 05:59:01 +0000 (22:59 -0700)]
[tf.data] Add native implementation for `tf.contrib.data.unbatch()`.

The implementation has two main improvements:
1. Avoid relatively expensive (~15us) function invocation for each incoming batch.
2. Use std::move() where possible to avoid copying strings/variants into the unbatched
   elements.

PiperOrigin-RevId: 193467856

6 years ago[XLA] Fix BF16 propagation bug for while condition.
Yuanzhong Xu [Thu, 19 Apr 2018 05:07:12 +0000 (22:07 -0700)]
[XLA] Fix BF16 propagation bug for while condition.

PiperOrigin-RevId: 193465140

6 years agoInitial addition of CLZ HLO
Jacques Pienaar [Thu, 19 Apr 2018 04:15:41 +0000 (21:15 -0700)]
Initial addition of CLZ HLO

* Adds the HLO op and lowering on CPU/GPU/evaluator;
* This does not update the operation semantics;

PiperOrigin-RevId: 193461989

6 years agoExpose tf.contrib.training.{prepend_from_queue_and_padded_batch_dataset}
Eugene Brevdo [Thu, 19 Apr 2018 03:15:27 +0000 (20:15 -0700)]
Expose tf.contrib.training.{prepend_from_queue_and_padded_batch_dataset}

Also its helper method "enqueue_in_queue_dataset".

PiperOrigin-RevId: 193458095

6 years agoUse the host implementation of vec permute op if the input on the host. Note
Yao Zhang [Thu, 19 Apr 2018 03:04:56 +0000 (20:04 -0700)]
Use the host implementation of vec permute op if the input on the host. Note
that the op still needs to be placed on the GPU so that it stays within the
same partiion with the neighboring ops, and as a result, no unnecessary send
and recv are created.

PiperOrigin-RevId: 193457328

6 years agoTeach the reshape mover pass about sharding.
A. Unique TensorFlower [Thu, 19 Apr 2018 03:02:52 +0000 (20:02 -0700)]
Teach the reshape mover pass about sharding.

PiperOrigin-RevId: 193457083

6 years agoGo: Update generated wrapper functions for TensorFlow ops.
A. Unique TensorFlower [Thu, 19 Apr 2018 02:49:12 +0000 (19:49 -0700)]
Go: Update generated wrapper functions for TensorFlow ops.
PiperOrigin-RevId: 193456151

6 years agoUpdate ops-related pbtxt files.
A. Unique TensorFlower [Thu, 19 Apr 2018 02:18:33 +0000 (19:18 -0700)]
Update ops-related pbtxt files.

PiperOrigin-RevId: 193454093