platform/upstream/tensorflow.git
6 years ago[TF/XLA] Add validation to find ops incompatible with the given device type at the...
A. Unique TensorFlower [Tue, 29 May 2018 17:54:51 +0000 (10:54 -0700)]
[TF/XLA] Add validation to find ops incompatible with the given device type at the beginning of graph compilation.

PiperOrigin-RevId: 198421828

6 years ago Support NHWC_VECT_W in MakeShapeFromFormat.
A. Unique TensorFlower [Tue, 29 May 2018 17:53:40 +0000 (10:53 -0700)]
  Support NHWC_VECT_W in MakeShapeFromFormat.

PiperOrigin-RevId: 198421617

6 years agoAdd AnonymousIteratorHandleOp for non-shared Iterator resources
Allen Lavoie [Tue, 29 May 2018 17:34:20 +0000 (10:34 -0700)]
Add AnonymousIteratorHandleOp for non-shared Iterator resources

Fixes Iterator cleanup when executing eagerly. DestroyResourceOp will now remove the last reference from the Iterator resource when it runs (after the last Python reference to an EagerIterator is removed).

Previously EagerIterator used IteratorHandleOp to create resource handles, which used one kernel per (unique) shared name since the shared name was an attribute. These kernels each held a reference to their resource, which kept it alive indefinitely.

Fixes #19499.

PiperOrigin-RevId: 198417997

6 years agoClarify argument types and relationships in docstrings of statistical_testing.py.
A. Unique TensorFlower [Tue, 29 May 2018 17:17:39 +0000 (10:17 -0700)]
Clarify argument types and relationships in docstrings of statistical_testing.py.

PiperOrigin-RevId: 198414898

6 years agoEliminate self.test_session in favor of self.evaluate in statistical_testing_test.py.
A. Unique TensorFlower [Tue, 29 May 2018 16:49:42 +0000 (09:49 -0700)]
Eliminate self.test_session in favor of self.evaluate in statistical_testing_test.py.

PiperOrigin-RevId: 198410306

6 years agoClean up: handle the hidden additional clause on for loops in a way that's clearer...
Dan Moldovan [Tue, 29 May 2018 16:46:21 +0000 (09:46 -0700)]
Clean up: handle the hidden additional clause on for loops in a way that's clearer about what it does.

PiperOrigin-RevId: 198409797

6 years agoClean up: remove useless super delegation.
Dan Moldovan [Tue, 29 May 2018 16:15:44 +0000 (09:15 -0700)]
Clean up: remove useless super delegation.

PiperOrigin-RevId: 198405670

6 years agoAllow assignment to subscripts in static analysis.
Dan Moldovan [Tue, 29 May 2018 15:45:28 +0000 (08:45 -0700)]
Allow assignment to subscripts in static analysis.

Move the handling of syntactic unpackings to a generic helper function since the pattern is used in multiple places. Update the type info analyzer to correctly process function arguments.

PiperOrigin-RevId: 198401368

6 years agoGeneralize assert_true_mean_equal and assert_true_mean_equal_two_sample to assert_tru...
A. Unique TensorFlower [Tue, 29 May 2018 15:36:14 +0000 (08:36 -0700)]
Generalize assert_true_mean_equal and assert_true_mean_equal_two_sample to assert_true_mean_in_interval.

PiperOrigin-RevId: 198400265

6 years agoRe-apply #18192.
Asim Shankar [Tue, 29 May 2018 07:50:34 +0000 (00:50 -0700)]
Re-apply #18192.

PiperOrigin-RevId: 198358055

6 years agoFix an incorrect precondition check in IndexedArrayAnalysis
Sanjoy Das [Tue, 29 May 2018 06:55:19 +0000 (23:55 -0700)]
Fix an incorrect precondition check in IndexedArrayAnalysis

PiperOrigin-RevId: 198354001

6 years agoPass HloOpcode instead of HloInstruction; NFC
Sanjoy Das [Tue, 29 May 2018 06:03:39 +0000 (23:03 -0700)]
Pass HloOpcode instead of HloInstruction; NFC

Minor code cleanup change.

PiperOrigin-RevId: 198351045

6 years agoMake IndexedArrayAnalysis behave well around StatusOr
Sanjoy Das [Tue, 29 May 2018 05:16:46 +0000 (22:16 -0700)]
Make IndexedArrayAnalysis behave well around StatusOr

PiperOrigin-RevId: 198348355

6 years agoAdding tf.data optimization for rewriting `map(...).batch(...)` to `map_and_batch...
Jiri Simsa [Mon, 28 May 2018 16:33:49 +0000 (09:33 -0700)]
Adding tf.data optimization for rewriting `map(...).batch(...)` to `map_and_batch(...)`.

PiperOrigin-RevId: 198310806

6 years agoSort variables in C++ instead of Python.
Tom Hennigan [Mon, 28 May 2018 13:32:04 +0000 (06:32 -0700)]
Sort variables in C++ instead of Python.

PiperOrigin-RevId: 198298103

6 years agoRelax compatibility checks for Protocol Buffer classes, to not rely on the exact...
A. Unique TensorFlower [Mon, 28 May 2018 12:21:13 +0000 (05:21 -0700)]
Relax compatibility checks for Protocol Buffer classes, to not rely on the exact layout of the generated classes. The definition of the message is checked instead.

PiperOrigin-RevId: 198292780

6 years ago[XLA] Don't display metadata inline in HLO graph dump.
Justin Lebar [Sun, 27 May 2018 20:27:28 +0000 (13:27 -0700)]
[XLA] Don't display metadata inline in HLO graph dump.

We only want to display it in the tooltip.

PiperOrigin-RevId: 198235268

6 years agoTPUEstimator.export_savedmodel() saves a SavedModel with both TPU and CPU graphs.
A. Unique TensorFlower [Sun, 27 May 2018 17:49:12 +0000 (10:49 -0700)]
TPUEstimator.export_savedmodel() saves a SavedModel with both TPU and CPU graphs.

PiperOrigin-RevId: 198229550

6 years agoExtracts the 'remove random shuffle node' optimization into its own method.
A. Unique TensorFlower [Sat, 26 May 2018 15:25:12 +0000 (08:25 -0700)]
Extracts the 'remove random shuffle node' optimization into its own method.

PiperOrigin-RevId: 198169790

6 years agoFix the issue where returned Status is not used.
Guangda Lai [Sat, 26 May 2018 06:44:11 +0000 (23:44 -0700)]
Fix the issue where returned Status is not used.

PiperOrigin-RevId: 198146500

6 years agoDepthwiseConv Optimizations
A. Unique TensorFlower [Sat, 26 May 2018 05:46:46 +0000 (22:46 -0700)]
DepthwiseConv Optimizations

PiperOrigin-RevId: 198144118

6 years agoUse dict(locals()) instead of distribution_util.parent_frame_arguments. This will...
A. Unique TensorFlower [Sat, 26 May 2018 04:38:56 +0000 (21:38 -0700)]
Use dict(locals()) instead of distribution_util.parent_frame_arguments. This will be much
faster at object construction time (going forward we'll figure out a way to make this a function to call).

PiperOrigin-RevId: 198141184

6 years ago[XLA] Don't compute relative error when the expected value is 0.
Justin Lebar [Sat, 26 May 2018 03:23:31 +0000 (20:23 -0700)]
[XLA] Don't compute relative error when the expected value is 0.

In literal_comparison, don't try to compute a relative error when the
expected value is 0, because doing so would mean that the only
acceptable value *is* zero, which probably isn't what you mean.

PiperOrigin-RevId: 198137414

6 years agoAdd support for unary and binary ops to indexed tensor analysis
Sanjoy Das [Sat, 26 May 2018 02:21:57 +0000 (19:21 -0700)]
Add support for unary and binary ops to indexed tensor analysis

I've added a TODO to clean up the use of ValueOrDie which I will address in an
immediately following CL.

PiperOrigin-RevId: 198134579

6 years ago[Hlo Graphviz] Always show metadata as tooltips
Yunxing Dai [Sat, 26 May 2018 02:18:30 +0000 (19:18 -0700)]
[Hlo Graphviz] Always show metadata as tooltips

Always show metadata as tooltips.

PiperOrigin-RevId: 198134430

6 years agoAdd a type-erased broadcast implementation to xla::Literal
Sanjoy Das [Sat, 26 May 2018 00:46:19 +0000 (17:46 -0700)]
Add a type-erased broadcast implementation to xla::Literal

And use this in HLO evaluator.  Since broadcast only moves bytes around we don't
need a type specialized implementation.

I'll use this in a later change.

PiperOrigin-RevId: 198128524

6 years agoMinor eager performance improvements
Akshay Modi [Sat, 26 May 2018 00:37:01 +0000 (17:37 -0700)]
Minor eager performance improvements

- remove linear regression dependence on global step.
  This speeds things up a lot for the benchmark (since it removes a bunch of
  unnecessary code), but is obviously not a fair comparison.
  I think its worth doing, since I don't see any reason to have a global step
  in eager.

- nn_ops dropout had an unnecessary convert_to_tensor, convert back to numpy
  (with a GPU this would copy out, copy back).
- cudnn_recurrent reshape would always fallback to the slow path - so I just
  converted it to be in the fastpath - this will be low impact.

- tensor_shape should not generate a new object every time
- remove unnecessary list creation and searching in some dtypes functions

PiperOrigin-RevId: 198127757

6 years ago[TF:XLA] Add direct implementation of AssignVariableOp for XLA devices.
Peter Hawkins [Sat, 26 May 2018 00:29:37 +0000 (17:29 -0700)]
[TF:XLA] Add direct implementation of AssignVariableOp for XLA devices.

This allows us to avoid an XLA compilation and tensor copies when assigning to a variable placed on an XLA device.

PiperOrigin-RevId: 198127062

6 years agoTurn on heuristic (mostly-NHWC) convolution layout assignment for (V100, fp16) by...
A. Unique TensorFlower [Sat, 26 May 2018 00:22:17 +0000 (17:22 -0700)]
Turn on heuristic (mostly-NHWC) convolution layout assignment for (V100, fp16) by default.

Also increase the column reduction tile size to reduce atomic operations.

PiperOrigin-RevId: 198126505

6 years agoEnable while loop constant sinking for GPU
Sanjoy Das [Sat, 26 May 2018 00:22:11 +0000 (17:22 -0700)]
Enable while loop constant sinking for GPU

To avoid keeping constants in while loop bodies after optimization (where they
may cause extra copies) we run a late pass of LICM that has been asked to hoist
constants when it can.

PiperOrigin-RevId: 198126497

6 years agoProvide a step container to OpKernelContexts when running eagerly.
Akshay Agrawal [Sat, 26 May 2018 00:12:49 +0000 (17:12 -0700)]
Provide a step container to OpKernelContexts when running eagerly.

This lets us run some ops that require step containers (e.g.
TensorArray-related ops).

 Before change:
 Benchmark                    Time(ns)        CPU(ns)     Iterations
-------------------------------------------------------------------
 BM_CreateGraph                  61292          80915           8581
 BM_RunGraph                      7899          13398          51251
 BM_CreateAndDestroySession       2588           2594         269838
 BM_KernelAndDeviceInit           2971           2976         235908
 BM_KernelAndDeviceRun             505            506        1000000

 After change:
 Benchmark                    Time(ns)        CPU(ns)     Iterations
 -------------------------------------------------------------------
 BM_CreateGraph                  78295         105539           8698
 BM_RunGraph                      9907          16988          47908
 BM_CreateAndDestroySession       2773           2778         247635
 BM_KernelAndDeviceInit           2678           2682         270054
 BM_KernelAndDeviceRun             553            554        1000000

PiperOrigin-RevId: 198125630

6 years ago[tf.data] Fixing concurrency issue in `map_and_batch`.
Jiri Simsa [Sat, 26 May 2018 00:05:33 +0000 (17:05 -0700)]
[tf.data] Fixing concurrency issue in `map_and_batch`.

PiperOrigin-RevId: 198124860

6 years agoIgnore while loops instead of mangling them in the automatic control dependencies.
Alexandre Passos [Fri, 25 May 2018 23:43:40 +0000 (16:43 -0700)]
Ignore while loops instead of mangling them in the automatic control dependencies.

PiperOrigin-RevId: 198122188

6 years agoExtracts the 'remove reverse node' optimization into its own method.
A. Unique TensorFlower [Fri, 25 May 2018 23:43:29 +0000 (16:43 -0700)]
Extracts the 'remove reverse node' optimization into its own method.

PiperOrigin-RevId: 198122165

6 years agoAutomated g4 rollback of changelist 198087342
A. Unique TensorFlower [Fri, 25 May 2018 23:07:25 +0000 (16:07 -0700)]
Automated g4 rollback of changelist 198087342

PiperOrigin-RevId: 198117552

6 years agoAdd warning to LookupOrCreate about reentrancy issue
Nick Felt [Fri, 25 May 2018 22:11:46 +0000 (15:11 -0700)]
Add warning to LookupOrCreate about reentrancy issue

PiperOrigin-RevId: 198110382

6 years agoAdd EagerTensor profiler and device shape utilities
Igor Ganichev [Fri, 25 May 2018 20:58:51 +0000 (13:58 -0700)]
Add EagerTensor profiler and device shape utilities

This change includes the following steps to make
EagerTensor profiler work:
 - Add a PaddedShapeFn to XlaDevice::Metadata. We need a
   backend-independent way to get a fully-padded shape and
   its layout on the device. This function is set during
   device construction. CPU and GPU devices effectively get
   an identity function since they neither change the layout
   nor pad. TPU gets the appropriate function.
 - Add TFE_TensorDebugInfo struct and C API methods for it.
   These methods are necessary to fetch the shape and layout
   from under the C API to the Python level. This can be a home
   for more debug information later.
 - Make EagerTensor weak referencable. This involves adding a
   pointer to the list of current weak references. This addition
   should have negligible overhead when profiler is not used.
   The only operations on this field are setting it to null on
   construction and checking if it is null on destruction.
 - Adding C++ functions callable from Python to register an instance
   of EagerTensorProfiler and retrieve debug information for a given
   EagerTensor. These functions are used in the new "inspect" module.
 - Finally, writing the actual profiler.

PiperOrigin-RevId: 198098380

6 years agoDisable //tensorflow/contrib/lite/python:lite_test on Windows
A. Unique TensorFlower [Fri, 25 May 2018 20:44:36 +0000 (13:44 -0700)]
Disable //tensorflow/contrib/lite/python:lite_test on Windows

PiperOrigin-RevId: 198096344

6 years ago [tpu:profiler] Capture the data for generating a memory viewer of the profiling...
A. Unique TensorFlower [Fri, 25 May 2018 20:39:25 +0000 (13:39 -0700)]
   [tpu:profiler] Capture the data for generating a memory viewer of the profiling results.

PiperOrigin-RevId: 198095564

6 years ago[TF:XLA] Bump open source llvm revision to r333273
Sanjoy Das [Fri, 25 May 2018 20:38:24 +0000 (13:38 -0700)]
[TF:XLA] Bump open source llvm revision to r333273

PiperOrigin-RevId: 198095416

6 years agoPublic API to switch between eager execution and graph building.
Alexandre Passos [Fri, 25 May 2018 20:20:13 +0000 (13:20 -0700)]
Public API to switch between eager execution and graph building.

Now, after tf.enable_eager_execution() has been executed, entering the context
manager of a tf.Graph will enable graph mode. So, for example

```
tf.enable_eager_execution()
with tf.Graph().as_default():
  c = tf.constant(1.0)  # this is a graph tensor
c2 = tf.constant(1.0)  # this is an eager tensor
```

The main use-case of this is allowing documentation writers to make a single
notebook which starts with eager execution and seamlessly transitions to
building graphs.

This also makes many explicit enablings of graph mode in the code redundant
(a cleanup cl will follow).

PiperOrigin-RevId: 198092991

6 years agoUse functions to build dense splits. Tensorflow Function invocations share the same...
A. Unique TensorFlower [Fri, 25 May 2018 19:58:55 +0000 (12:58 -0700)]
Use functions to build dense splits. Tensorflow Function invocations share the same graph so using them reduces the graph construction overhead.

PiperOrigin-RevId: 198090110

6 years ago [tpu:profiler] Minor change in the description of tool name proto.
A. Unique TensorFlower [Fri, 25 May 2018 19:56:40 +0000 (12:56 -0700)]
   [tpu:profiler] Minor change in the description of tool name proto.

PiperOrigin-RevId: 198089875

6 years agoAdd ScopedAllocatorOptimizer in support of CollectiveReduce.
A. Unique TensorFlower [Fri, 25 May 2018 19:54:49 +0000 (12:54 -0700)]
Add ScopedAllocatorOptimizer in support of CollectiveReduce.

The efficiency of CollectiveReduce is greatly improved by merging
multiple parallel reductions over smaller tensors into a single
reduction over a larger tensor that is the concatentation of the
smaller tensors.  Because CollectiveReduce is essentially an
element-wise array operation which operates on a 1-D reshape of
the input tensor it is eligible for a ScopedAllocation optimization.

The optimization works by looking for serially independent instances
of CollectiveReduce that lie within the same name-scope tier and
have the same control-flow (e.g. loop) embedding structure.  Where
two or more such nodes are found the upstream nodes that generate
their inputs are modified to write their outputs into consecutive
regions of a single tensor buffer maintained by a ScopedAllocator.
The multiple CollectiveReduce nodes are then replaced by a single
CollectiveReduce that operates in-place on the backing buffer.

The effectiveness of the optimization depends on there being candidate
CollectiveReduce nodes with these characteristics that become eligible
for execution at close to the same time.  If the name scope is too
large, and includes nodes that become execution eligible at very different
times, this graph rewrite could result in a slowdown.

Note that this optimization is experimental: it is not guaranteed to
work, especially for ops other than CollectiveReduce.

PiperOrigin-RevId: 198089642

6 years agoenhance Tensorflow GBDT and GBRT model by exposing a new two dimensional output in...
A. Unique TensorFlower [Fri, 25 May 2018 19:35:50 +0000 (12:35 -0700)]
enhance Tensorflow GBDT and GBRT model by exposing a new two dimensional output in prediction ops (example id, tree leaf node index id) for input as other model features

PiperOrigin-RevId: 198087342

6 years agoExtracts the 'simplify slice' optimization into its own method.
A. Unique TensorFlower [Fri, 25 May 2018 19:22:45 +0000 (12:22 -0700)]
Extracts the 'simplify slice' optimization into its own method.

PiperOrigin-RevId: 198085532

6 years ago[TF:XLA] Register Switch and Merge ops on XLA devices.
Peter Hawkins [Fri, 25 May 2018 19:04:49 +0000 (12:04 -0700)]
[TF:XLA] Register Switch and Merge ops on XLA devices.

PiperOrigin-RevId: 198083156

6 years agoAutomated g4 rollback of changelist 192848921
Derek Murray [Fri, 25 May 2018 18:42:33 +0000 (11:42 -0700)]
Automated g4 rollback of changelist 192848921

PiperOrigin-RevId: 198079927

6 years agoExtracts the 'simplify strided slice' optimization into its own method.
A. Unique TensorFlower [Fri, 25 May 2018 18:34:30 +0000 (11:34 -0700)]
Extracts the 'simplify strided slice' optimization into its own method.

PiperOrigin-RevId: 198078724

6 years agoBump TPU batch size and wrap apply_grads in defun
Igor Ganichev [Fri, 25 May 2018 18:27:39 +0000 (11:27 -0700)]
Bump TPU batch size and wrap apply_grads in defun

PiperOrigin-RevId: 198077643

6 years agoRelease C++ lock before calling back into python
Akshay Modi [Fri, 25 May 2018 18:02:42 +0000 (11:02 -0700)]
Release C++ lock before calling back into python

PiperOrigin-RevId: 198073059

6 years agoDepthwiseConv optimizations.
A. Unique TensorFlower [Fri, 25 May 2018 17:54:38 +0000 (10:54 -0700)]
DepthwiseConv optimizations.

PiperOrigin-RevId: 198071709

6 years agoLink to tf.estimator docs for premade estimators.
Mark Daoust [Fri, 25 May 2018 17:45:27 +0000 (10:45 -0700)]
Link to tf.estimator docs for premade estimators.

PiperOrigin-RevId: 198070157

6 years agoCode simplification in dump_graphviz.cc:
A. Unique TensorFlower [Fri, 25 May 2018 15:55:24 +0000 (08:55 -0700)]
Code simplification in dump_graphviz.cc:
Just output all arrays, before writing edges, so we don't
need to keep track of which arrays we've already output.

PiperOrigin-RevId: 198055327

6 years agoMinor clarification to model_to_estimator() doc string
Shanqing Cai [Fri, 25 May 2018 13:56:38 +0000 (06:56 -0700)]
Minor clarification to model_to_estimator() doc string

PiperOrigin-RevId: 198044106

6 years agoeager: Update introduction notebooks.
Asim Shankar [Fri, 25 May 2018 09:23:06 +0000 (02:23 -0700)]
eager: Update introduction notebooks.

PiperOrigin-RevId: 198022387

6 years agoFix typo, fix build.
Asim Shankar [Fri, 25 May 2018 08:36:23 +0000 (01:36 -0700)]
Fix typo, fix build.

PiperOrigin-RevId: 198017870

6 years agoExtracts the 'simplify tile node' optimization into its own method.
A. Unique TensorFlower [Fri, 25 May 2018 03:36:45 +0000 (20:36 -0700)]
Extracts the 'simplify tile node' optimization into its own method.

PiperOrigin-RevId: 197996636

6 years agoGo: Update generated wrapper functions for TensorFlow ops.
A. Unique TensorFlower [Fri, 25 May 2018 02:49:05 +0000 (19:49 -0700)]
Go: Update generated wrapper functions for TensorFlow ops.
PiperOrigin-RevId: 197993384

6 years agoInitialize the score threshold to -inf to avoid filtering out negative logits
A. Unique TensorFlower [Fri, 25 May 2018 02:45:27 +0000 (19:45 -0700)]
Initialize the score threshold to -inf to avoid filtering out negative logits

PiperOrigin-RevId: 197993147

6 years agoUpdate ops-related pbtxt files.
A. Unique TensorFlower [Fri, 25 May 2018 02:20:31 +0000 (19:20 -0700)]
Update ops-related pbtxt files.

PiperOrigin-RevId: 197991672

6 years agoMerge changes from github.
Yifei Feng [Fri, 25 May 2018 02:12:26 +0000 (19:12 -0700)]
Merge changes from github.
Revert #18413. Too many internal test failures due to the name scope change caused by this change.
Revert #18192. Cannot use re2::StringPiece internally. Need alternative for set call. Will pull and clean this up in a separate change.

PiperOrigin-RevId: 197991247

6 years agoExtracts the 'simplify pad node' optimization into its own method.
A. Unique TensorFlower [Fri, 25 May 2018 01:55:30 +0000 (18:55 -0700)]
Extracts the 'simplify pad node' optimization into its own method.

PiperOrigin-RevId: 197989813

6 years agoRename TileLoader to MemoryTile; NFC
Sanjoy Das [Fri, 25 May 2018 01:23:48 +0000 (18:23 -0700)]
Rename TileLoader to MemoryTile; NFC

In a later change I will expand MemoryTile to store tiles and load "3d" tiles
(where we broadcast along one dimension as we load).

PiperOrigin-RevId: 197987185

6 years agoAdd heuristic on picking NHWC layout for (V100, fp16) convolutions.
A. Unique TensorFlower [Fri, 25 May 2018 00:48:21 +0000 (17:48 -0700)]
Add heuristic on picking NHWC layout for (V100, fp16) convolutions.

Also move AlgorithmPicker after layout assignment, as now
cudnn_convolution_runner will return failures on invalid input layouts.

Also add a backend debug option to switch the layout heuristic. By default
it has the old behavior (all NCHW).

PiperOrigin-RevId: 197983747

6 years agoEnabling some potential optimization using the restrict qualifier.
A. Unique TensorFlower [Fri, 25 May 2018 00:06:34 +0000 (17:06 -0700)]
Enabling some potential optimization using the restrict qualifier.

PiperOrigin-RevId: 197979118

6 years agoWhen converting a numpy float64 to an EagerTensor, always ensure that it
Akshay Modi [Thu, 24 May 2018 23:53:33 +0000 (16:53 -0700)]
When converting a numpy float64 to an EagerTensor, always ensure that it
becomes a float64 tensor.

Earlier py_seq_tensor would fall back to a float32 if not explicitly requesting
a float64 (which would not happen if we had no other information).

PiperOrigin-RevId: 197977260

6 years agoDon't XLA-compile naked variable reads
Igor Ganichev [Thu, 24 May 2018 23:44:17 +0000 (16:44 -0700)]
Don't XLA-compile naked variable reads

Before this change, when we executed a naked variable read (i.e. outside of
a defun, directly running <xla_device>->Compute()), tf2xla kernel would
copy the variable's tensor leading to many unnecessary copies.

This change uses the regular non-tf2xla kernel for naked variable reads
and marks the tf2xla one for CompilationOnly().

PiperOrigin-RevId: 197976146

6 years agomove wide string manipulations out of windows_file_system
A. Unique TensorFlower [Thu, 24 May 2018 23:31:48 +0000 (16:31 -0700)]
move wide string manipulations out of windows_file_system

PiperOrigin-RevId: 197974385

6 years agoRemove _get_backward_fn and depend on _gradient_function directly.
Akshay Modi [Thu, 24 May 2018 23:20:31 +0000 (16:20 -0700)]
Remove _get_backward_fn and depend on _gradient_function directly.

(_magic_gradient_function was renamed to _gradient_function)

Before:
entry {
  name: "MicroBenchmarks.benchmark_tf_gradient_forward_identity"
  iters: 30000
  wall_time: 5.88456789653
  extras {
    key: "examples_per_sec"
    value {
      double_value: 169936.011885
    }
  }
}

After:
entry {
  name: "MicroBenchmarks.benchmark_tf_gradient_forward_identity"
  iters: 30000
  wall_time: 5.04853725433
  extras {
    key: "examples_per_sec"
    value {
      double_value: 198077.175551
    }
  }
}
PiperOrigin-RevId: 197972668

6 years agoFix the generated builtin_ops enum header.
Yu-Cheng Ling [Thu, 24 May 2018 23:02:11 +0000 (16:02 -0700)]
Fix the generated builtin_ops enum header.

PiperOrigin-RevId: 197969642

6 years agoExtracts the 'simplify squeeze node' optimization into its own method.
A. Unique TensorFlower [Thu, 24 May 2018 22:53:44 +0000 (15:53 -0700)]
Extracts the 'simplify squeeze node' optimization into its own method.

PiperOrigin-RevId: 197968452

6 years ago[XLA] Remove maps with a single instruction
David Majnemer [Thu, 24 May 2018 22:45:25 +0000 (15:45 -0700)]
[XLA] Remove maps with a single instruction

These maps aren't really pulling their weight, fold them to the instruction
that they compute.

PiperOrigin-RevId: 197967117

6 years agoAvoid infinite recursion when checking for indexed slices.
Priya Gupta [Thu, 24 May 2018 22:35:13 +0000 (15:35 -0700)]
Avoid infinite recursion when checking for indexed slices.

PiperOrigin-RevId: 197965508

6 years agoAllow combinations to be used on the class level. Make "mode" optional.
Igor Saprykin [Thu, 24 May 2018 22:28:03 +0000 (15:28 -0700)]
Allow combinations to be used on the class level.  Make "mode" optional.

Applying a generator to a class is the same as applying that generator to every member of that class.  It is meant to allow avoiding repetition in some cases.

The implementation relies on some internals of parameterized tests and how it works with a class level declaration:  https://github.com/abseil/abseil-py/blob/master/absl/testing/parameterized.py#L319.

The "mode" argument is required before this change.  To accommodate cases where execution mode isn't the point of the test, "mode" became optional with "graph" mode being default.  Another idea I had was to pick a random mode by default.

PiperOrigin-RevId: 197964501

6 years agoAdd local_init_run_options to SessionManager and Supervisor so that
A. Unique TensorFlower [Thu, 24 May 2018 22:27:00 +0000 (15:27 -0700)]
Add local_init_run_options to SessionManager and Supervisor so that
collective_graph_key can be passed in when collective ops are used
in variable initialization.

PiperOrigin-RevId: 197964316

6 years agoRename getInt64 to GetInt64 to follow Google style
Sanjoy Das [Thu, 24 May 2018 22:19:40 +0000 (15:19 -0700)]
Rename getInt64 to GetInt64 to follow Google style

PiperOrigin-RevId: 197963232

6 years agoWindows build script change for release job
A. Unique TensorFlower [Thu, 24 May 2018 21:59:29 +0000 (14:59 -0700)]
Windows build script change for release job

PiperOrigin-RevId: 197959602

6 years agoSmall fix so that GDN can run on TPU
A. Unique TensorFlower [Thu, 24 May 2018 21:59:05 +0000 (14:59 -0700)]
Small fix so that GDN can run on TPU

PiperOrigin-RevId: 197959536

6 years agoRaise ValueError when calling model.summary() before it is built
Francois Chollet [Thu, 24 May 2018 21:58:15 +0000 (14:58 -0700)]
Raise ValueError when calling model.summary() before it is built

PiperOrigin-RevId: 197959372

6 years ago[TF:XLA] Avoid buffer copy when copying a Tensor onto an XLA device.
Peter Hawkins [Thu, 24 May 2018 21:20:39 +0000 (14:20 -0700)]
[TF:XLA] Avoid buffer copy when copying a Tensor onto an XLA device.

PiperOrigin-RevId: 197952565

6 years ago[XLA] Convert infeed call to take a LiteralSlice.
Chris Leary [Thu, 24 May 2018 21:03:41 +0000 (14:03 -0700)]
[XLA] Convert infeed call to take a LiteralSlice.

PiperOrigin-RevId: 197949637

6 years agotfdbg: fix issue where total source file size exceeds gRPC message size limit
Shanqing Cai [Thu, 24 May 2018 21:02:30 +0000 (14:02 -0700)]
tfdbg: fix issue where total source file size exceeds gRPC message size limit

* Source file content is now sent one by one, making it less likely that individual
  messages will have sizes above the 4-MB gRPC message size limit.
* In case the message for a single source file exceeds the limit, the client handles
  it gracefully by skipping the sending and print a warning message.

Fixes: https://github.com/tensorflow/tensorboard/issues/1118
PiperOrigin-RevId: 197949416

6 years agoFix bugs with the code blocks in defun's docstring.
Akshay Agrawal [Thu, 24 May 2018 20:30:15 +0000 (13:30 -0700)]
Fix bugs with the code blocks in defun's docstring.

PiperOrigin-RevId: 197943921

6 years agoAutomated g4 rollback of changelist 197868028
A. Unique TensorFlower [Thu, 24 May 2018 20:19:47 +0000 (13:19 -0700)]
Automated g4 rollback of changelist 197868028

PiperOrigin-RevId: 197942379

6 years agoadd maxpoolgrad transposer for layout optimizer.
A. Unique TensorFlower [Thu, 24 May 2018 20:18:32 +0000 (13:18 -0700)]
add maxpoolgrad transposer for layout optimizer.

PiperOrigin-RevId: 197942180

6 years agoRemoving outdated links.
Amit Patankar [Thu, 24 May 2018 20:15:37 +0000 (13:15 -0700)]
Removing outdated links.

PiperOrigin-RevId: 197941740

6 years agoExtracts the Simplify Pack optimization into its own method.
A. Unique TensorFlower [Thu, 24 May 2018 20:13:42 +0000 (13:13 -0700)]
Extracts the Simplify Pack optimization into its own method.

PiperOrigin-RevId: 197941474

6 years agoEnsure ResourceMgr::LookupOrCreate calls create fn just once
Nick Felt [Thu, 24 May 2018 20:07:50 +0000 (13:07 -0700)]
Ensure ResourceMgr::LookupOrCreate calls create fn just once

This addresses a race condition where LookupOrCreate is called at the same time from two threads, and both Lookup()s fail, so the creator() function is run twice, even though only a single Create() will then succeed.

The motivation is that some creator() functions have side-effects, e.g. tf.contrib.summary.create_file_writer()'s init op opens an events file.  This change ensures that if two init ops for file writers with the same resource name are run in the same session.run() call, only one events file will be created.  (Current behavior will often open two files; typically the second one overwrites the first but this won't happen if the filename_suffix values are different or the timestamps happen to straddle a second boundary.)

PiperOrigin-RevId: 197940607

6 years agoUpdated documentation for tf.reduce_join.
A. Unique TensorFlower [Thu, 24 May 2018 20:03:10 +0000 (13:03 -0700)]
Updated documentation for tf.reduce_join.

PiperOrigin-RevId: 197939808

6 years agoOnly wait for one of the input tensors to be ready.
A. Unique TensorFlower [Thu, 24 May 2018 20:00:07 +0000 (13:00 -0700)]
Only wait for one of the input tensors to be ready.
The waiting was implemented to avoid reading stale models as much as possible.
However with this dependency, each input column creates a Send/Recv to PS0
which slows down training significantly.
Colocate Quantile and Stats accumulators for the same handler.

PiperOrigin-RevId: 197939327

6 years agoModify tf.image.central_crop to support batched-input.
A. Unique TensorFlower [Thu, 24 May 2018 19:34:02 +0000 (12:34 -0700)]
Modify tf.image.central_crop to support batched-input.

Currently central_crop works on singular images with dynamic dimensions. For large image classification models, it would be nice if central_crop can be modified to support batched input. This CL makes that change.

PiperOrigin-RevId: 197935606

6 years agoMark queue related ops as having side effect
Benoit Steiner [Thu, 24 May 2018 19:23:32 +0000 (12:23 -0700)]
Mark queue related ops as having side effect

PiperOrigin-RevId: 197933941

6 years agoDon't use hex floats.
Jacques Pienaar [Thu, 24 May 2018 19:22:04 +0000 (12:22 -0700)]
Don't use hex floats.

Hex float literals are in C11 and C++17, but not in C++11, so use plain float notation.

PiperOrigin-RevId: 197933744

6 years agoFix doc: "--input_arrays" instead of "--input_array".
A. Unique TensorFlower [Thu, 24 May 2018 19:11:41 +0000 (12:11 -0700)]
Fix doc: "--input_arrays" instead of "--input_array".

PiperOrigin-RevId: 197932202

6 years ago[TF:XLA] Bump open source llvm revision to r333167
Sanjoy Das [Thu, 24 May 2018 18:54:56 +0000 (11:54 -0700)]
[TF:XLA] Bump open source llvm revision to r333167

PiperOrigin-RevId: 197929434

6 years agoFix `tf_inspect.getargspec` callable objects other than functions.
Mark Daoust [Thu, 24 May 2018 18:44:15 +0000 (11:44 -0700)]
Fix `tf_inspect.getargspec` callable objects other than functions.

PiperOrigin-RevId: 197927601

6 years ago[tf.data] Add `tf.contrib.data.choose_from_datasets()`.
Derek Murray [Thu, 24 May 2018 18:37:12 +0000 (11:37 -0700)]
[tf.data] Add `tf.contrib.data.choose_from_datasets()`.

This is a deterministic counterpart to `tf.contrib.data.sample_from_datasets()`.

PiperOrigin-RevId: 197926386

6 years agoExtracts the 'Move Constants Past Enter Node' optimization into its own method.
A. Unique TensorFlower [Thu, 24 May 2018 18:29:34 +0000 (11:29 -0700)]
Extracts the 'Move Constants Past Enter Node' optimization into its own method.

PiperOrigin-RevId: 197924962

6 years agoMake the existing checkpointable data structure a CheckpointableDataStructure
Allen Lavoie [Thu, 24 May 2018 18:23:18 +0000 (11:23 -0700)]
Make the existing checkpointable data structure a CheckpointableDataStructure

Gives it better/more consistent handling of Layers.

PiperOrigin-RevId: 197923880