Sanjoy Das [Tue, 20 Mar 2018 19:51:50 +0000 (12:51 -0700)]
[TF:XLA] Bump open source llvm revision to r327958
PiperOrigin-RevId:
189792132
Max Galkin [Tue, 20 Mar 2018 18:45:23 +0000 (11:45 -0700)]
Improved accuracy of op_level_cost_estimator (QuantizeV2, Dequantize, Gather).
PiperOrigin-RevId:
189779691
A. Unique TensorFlower [Tue, 20 Mar 2018 18:36:07 +0000 (11:36 -0700)]
Don't run tensorflow/python:function_test under asan.
It gets flaky timeouts.
PiperOrigin-RevId:
189777986
A. Unique TensorFlower [Tue, 20 Mar 2018 18:27:54 +0000 (11:27 -0700)]
The Quantized BroadcastSub portion of #17123
PiperOrigin-RevId:
189776376
A. Unique TensorFlower [Tue, 20 Mar 2018 18:21:40 +0000 (11:21 -0700)]
Update version of nsync used by TensorFlow.
The primary change is that on Linux, the C++11 build of nsync will now use
underlying system primitives to implement a semaphore instead of the C++11
primitives, which are currently surprisingly slow on Linux.
PiperOrigin-RevId:
189775201
Sanjoy Das [Tue, 20 Mar 2018 18:13:48 +0000 (11:13 -0700)]
Use 32 bit induction variable in gather expander
Right now this is unconditional (and we fail with Unimplemented() if a 32 bit
induction variable is not large enough), but eventually we may want to be
smarter about this.
PiperOrigin-RevId:
189773581
A. Unique TensorFlower [Tue, 20 Mar 2018 18:12:47 +0000 (11:12 -0700)]
Replace std::clock with random::New64() for setting random seed
PiperOrigin-RevId:
189773399
A. Unique TensorFlower [Tue, 20 Mar 2018 18:04:44 +0000 (11:04 -0700)]
Add support for automatically wrapping NumPy functions based on a whitelist.
PiperOrigin-RevId:
189771575
A. Unique TensorFlower [Tue, 20 Mar 2018 16:27:24 +0000 (09:27 -0700)]
Revise the main API for more consistent notation and add a supplemental shortcut to mark functions as "run in py_func". This is an intermediate step to simplifying the execution of plotting code.
PiperOrigin-RevId:
189753509
Jianwei Xie [Tue, 20 Mar 2018 15:32:30 +0000 (08:32 -0700)]
Fixed the bug that the export code triggers the TPU validation.
PiperOrigin-RevId:
189745966
Shanqing Cai [Tue, 20 Mar 2018 15:08:48 +0000 (08:08 -0700)]
tfdbg: Curses CLI: Fill line-end whitespace with default color pair
to prevent spurious color pairs from appearing in certain text
terminal environments.
RELNOTES: Bug fix: tfdbg curses CLI: fix spurious background colors in some text terminals.
PiperOrigin-RevId:
189742433
A. Unique TensorFlower [Tue, 20 Mar 2018 14:27:16 +0000 (07:27 -0700)]
Drop name_scope from operation names during quantization to avoid doubling it up.
PiperOrigin-RevId:
189737746
A. Unique TensorFlower [Tue, 20 Mar 2018 13:28:09 +0000 (06:28 -0700)]
Make TensorSpec and BoundedTensorSpec serializable.
PiperOrigin-RevId:
189731737
Brian Patton [Tue, 20 Mar 2018 13:25:19 +0000 (06:25 -0700)]
Adds float64 support for avg pool and its gradient.
Eigen NumTraits is modified to directly use std::numeric_limits, which resolves a broken test caused by inconsistency between the host and devices values of Eigen::NumTraits<double>::highest(). This returns +inf on device, due to third_party/eigen3/Eigen/src/Core/util/Meta.h, and __DBL_MAX__ (1.
7976931348623157e+308) on host, making the behavior for doubles (on device) inconsistent with both the behavior of floats Eigen::NumTraits<float>::highest() and the behavior of std::numeric_limits<double>::max()
PiperOrigin-RevId:
189731521
Justin Lebar [Tue, 20 Mar 2018 10:48:38 +0000 (03:48 -0700)]
Don't spin in a loop when we're not waiting on any GPU events.
PiperOrigin-RevId:
189719711
A. Unique TensorFlower [Tue, 20 Mar 2018 10:11:32 +0000 (03:11 -0700)]
- Added support for data to be specified in RNN classes as large tensors with time folded into the batch dimension instead of lists of tensors
- Significant refactoring of RNN classes
- Fixed a bunch of issues in the LayerCollection docstrings, especially around the 'reuse' argument.
PiperOrigin-RevId:
189716331
A. Unique TensorFlower [Tue, 20 Mar 2018 09:26:31 +0000 (02:26 -0700)]
Fix bug
PiperOrigin-RevId:
189712233
Sanjoy Das [Tue, 20 Mar 2018 06:15:42 +0000 (23:15 -0700)]
Fix some edge cases around scalar indices in the gather expander
I discovered these when changing the tf2xla bridge to directly emit gather
operations.
- DeScalarizeGatherIndices was assuming that gather_indices must be of at least
rank 1. Fix this to be more general.
- We were passing in the wrong version of gather indices to
ExpandFirstDimIntoNDims. We don't strictly need to pass in
transposed_gather_indices (since if transposed_gather_indices is rank 1 then
the transpose has to be an identity transpose), passing in
descalarized_gather_indices would also have been fine, but
transposed_gather_indices seems more uniform.
- ExpandGatherDimsInAccumulator was assuming that gather_indices must be of at
least rank 1 (by calling CollapseFirstNDims). Fix this to be more general.
- We were trying to go through with emitting zero sized gather operations. I
don't think it is worth dealing with all of the edge cases this would expose
so now we just punt to ZeroSizedHloElimination.
PiperOrigin-RevId:
189696444
A. Unique TensorFlower [Tue, 20 Mar 2018 05:56:01 +0000 (22:56 -0700)]
Predictions have to be updated for exported output signatures
PiperOrigin-RevId:
189694707
Mingsheng Hong [Tue, 20 Mar 2018 05:29:33 +0000 (22:29 -0700)]
Added infeed support for experimental C APIs associated with TPU graph rewrite.
This initial design of the C API is different from (and mostly higher level
than) the python API counterparts for infeed, in that the python API has
explicit graph construction APIs for generating infeed enqueue/dequeue
ops (e.g. split_inputs_and_generate_enqueue_ops() and generate_dequeue_op()),
while the C API takes an input graph and redirects all input nodes to feed the
infeed enqueue.
One requirement/restriction is that the input nodes in the TF
graph (e.g. Placeholder) must specify their tensor shapes, for infeed enqueue
and dequeue nodes to properly compile with XLA. The API for more general shape
support will be designed and implemented later.
PiperOrigin-RevId:
189693028
A. Unique TensorFlower [Tue, 20 Mar 2018 04:45:45 +0000 (21:45 -0700)]
Go: Update generated wrapper functions for TensorFlow ops.
PiperOrigin-RevId:
189690096
A. Unique TensorFlower [Tue, 20 Mar 2018 04:18:06 +0000 (21:18 -0700)]
Update ops-related pbtxt files.
PiperOrigin-RevId:
189688675
A. Unique TensorFlower [Tue, 20 Mar 2018 03:55:51 +0000 (20:55 -0700)]
Add `ostream<<` to `tensorflow::TensorShapeBase`.
Reason: Allow `LOG(ERROR) << shape` (currently disallowed).
PiperOrigin-RevId:
189687162
Suharsh Sivakumar [Tue, 20 Mar 2018 03:42:00 +0000 (20:42 -0700)]
Quantize bypasses after activations.
PiperOrigin-RevId:
189686219
Alexandre Passos [Tue, 20 Mar 2018 03:21:45 +0000 (20:21 -0700)]
Always imports the contrib summary ops when importing tensorflow.
Fixes #17802
PiperOrigin-RevId:
189684619
Jianwei Xie [Tue, 20 Mar 2018 03:06:26 +0000 (20:06 -0700)]
Adds final partial batch support for TPUEstimator.predict.
PiperOrigin-RevId:
189683528
A. Unique TensorFlower [Tue, 20 Mar 2018 02:58:03 +0000 (19:58 -0700)]
Apply output_min/output_max to the result in the NEON implementation of Add operator.
Both non-NEON and reference implementation have this, but it's missing from NEON version.
PiperOrigin-RevId:
189682984
Igor Ganichev [Tue, 20 Mar 2018 02:52:06 +0000 (19:52 -0700)]
Handle non-broadcastables shapes in eager assert_equal
Before this change assert_equal would fail when producing an error
message for non-equal shapes because array_ops.boolean_mask only
works for equal shapes.
This part of the error message is fairly confusing in presence
of non-equal shapes. This change removes it.
PiperOrigin-RevId:
189682518
A. Unique TensorFlower [Tue, 20 Mar 2018 02:30:23 +0000 (19:30 -0700)]
Avoid attaching fqn annotations to live values that don't have a `__name__`.
PiperOrigin-RevId:
189680937
Suharsh Sivakumar [Tue, 20 Mar 2018 02:24:26 +0000 (19:24 -0700)]
Disable freeze_bn_delay by default.
PiperOrigin-RevId:
189680481
A. Unique TensorFlower [Tue, 20 Mar 2018 02:24:24 +0000 (19:24 -0700)]
Update GraphProperties comments
PiperOrigin-RevId:
189680477
Surya Bhupatiraju [Tue, 20 Mar 2018 01:51:06 +0000 (18:51 -0700)]
Make L2 norm computation more stable.
Avoids the potentially numerically instable square root in the linalg_ops.norm() function because we 'undo' that operation with a math_ops.square() operation anyway.
PiperOrigin-RevId:
189677716
Asim Shankar [Tue, 20 Mar 2018 01:32:13 +0000 (18:32 -0700)]
Export tf.GradientTape
tf.GradientTape can be used both for eager execution and graph construction
to compute gradients (unlike tf.gradients, which works only for graph
construction).
PiperOrigin-RevId:
189676004
A. Unique TensorFlower [Tue, 20 Mar 2018 01:20:12 +0000 (18:20 -0700)]
Support general permutation.
PiperOrigin-RevId:
189675019
Suharsh Sivakumar [Tue, 20 Mar 2018 00:56:02 +0000 (17:56 -0700)]
Add docstring pointing to tf.contrib.quantize.
PiperOrigin-RevId:
189672549
Martin Wicke [Tue, 20 Mar 2018 00:50:05 +0000 (17:50 -0700)]
Register gradient for argmin (cf. #15278).
PiperOrigin-RevId:
189671974
A. Unique TensorFlower [Tue, 20 Mar 2018 00:45:10 +0000 (17:45 -0700)]
add option to save trace table to model directory's profile plugin subdirectory.
PiperOrigin-RevId:
189671290
Dustin Tran [Tue, 20 Mar 2018 00:41:25 +0000 (17:41 -0700)]
Standardize bib references and Examples subsection in docstrings.
Recipe:
+ Write a #### Examples subsection below Args/Returns/Raises to illustrate examples. If the docstring's last line is a ``` closing a code snippet, add an empty line before closing the docstring with """. This properly displays the code snippet.
+ Write a #### References subsection at the bottom of any docstring with citations. Enumerate all references in alphabetical order. Individual bibentries use ICLR?s bibliography style, which borrows from icml2010.bst and which itself borrows from plainnl.bst. Add a link to the paper if the publication is open source (ideally, arXiv).
PiperOrigin-RevId:
189670932
Skye Wanderman-Milne [Tue, 20 Mar 2018 00:34:47 +0000 (17:34 -0700)]
Make _USE_C_API = True and_USE_C_SHAPES = False work with import_graph_def.
Without this change, shapes wouldn't be correctly computed for
operations created via import_graph_def.
PiperOrigin-RevId:
189670312
A. Unique TensorFlower [Tue, 20 Mar 2018 00:29:19 +0000 (17:29 -0700)]
Add a clif build rule for saved_model.
PiperOrigin-RevId:
189669509
A. Unique TensorFlower [Tue, 20 Mar 2018 00:23:20 +0000 (17:23 -0700)]
Improve flatbuffer verification.
PiperOrigin-RevId:
189668634
Derek Murray [Tue, 20 Mar 2018 00:13:11 +0000 (17:13 -0700)]
[tf.data] Combine implementations of FlatMapDataset, InterleaveDataset and ParallelInterleaveDataset.
PiperOrigin-RevId:
189667086
Yuefeng Zhou [Tue, 20 Mar 2018 00:04:28 +0000 (17:04 -0700)]
Fix test failure
PiperOrigin-RevId:
189666053
Lukasz Kaiser [Mon, 19 Mar 2018 23:56:44 +0000 (16:56 -0700)]
Automated g4 rollback of changelist
188440916
PiperOrigin-RevId:
189664854
Zhixian Yan [Mon, 19 Mar 2018 22:50:02 +0000 (15:50 -0700)]
Disable lstm test in generated_example due to state non-definitive init.
PiperOrigin-RevId:
189654943
Xiaoqiang Zheng [Mon, 19 Mar 2018 22:40:37 +0000 (15:40 -0700)]
A few changes to improve the real data performance:
* Turn off the force_gpu_compatible by default.
* Move the cast operator within the processing operator.
* Have the map_and_batch operator produce gpu_compatible output.
* Add an option to produce fp16 tensors for network transfer by default.
On DGX-1 V100, with resnet50, I got 5050 images/sec on real data, 5395 images/sec on synthetic data.
With trivial model, I got 13000+ images/sec on real data.
PiperOrigin-RevId:
189653575
A. Unique TensorFlower [Mon, 19 Mar 2018 22:13:53 +0000 (15:13 -0700)]
Run flatbuffer verifier before reading a TFLITE file into toco.
PiperOrigin-RevId:
189649236
A. Unique TensorFlower [Mon, 19 Mar 2018 22:09:23 +0000 (15:09 -0700)]
Use fully-qualified function names and avoid the need to replace attributes.
PiperOrigin-RevId:
189648496
Rohan Jain [Mon, 19 Mar 2018 22:06:40 +0000 (15:06 -0700)]
Allowing the FunctionBufferingResource to be passed in thread_pool_size=0 in which case we wouldn't pass in a runner to the FLR::Run call and rely on the underlying device threadpool instead.
PiperOrigin-RevId:
189648051
Akshay Modi [Mon, 19 Mar 2018 21:57:09 +0000 (14:57 -0700)]
Maintain an updateable map of devices in the eager context.
PiperOrigin-RevId:
189646358
Ilya Biryukov [Mon, 19 Mar 2018 21:48:23 +0000 (14:48 -0700)]
Fix build breakage with downloadable clang and -fopenmp.
By disabling openmp when building with clang.
If we want to enable openmp with clang, we'll probably have to have libomp as
an explicit dependency.
This fixes a breakage found by OS CI:
https://ci.tensorflow.org/view/Nightly/job/nightly-matrix-linux-gpu-clang/215/
PiperOrigin-RevId:
189644968
Yu-Cheng Ling [Mon, 19 Mar 2018 21:27:52 +0000 (14:27 -0700)]
TFLite Delegate: Add an `allow_dynamic_tensors` parameter.
PiperOrigin-RevId:
189641833
A. Unique TensorFlower [Mon, 19 Mar 2018 21:27:06 +0000 (14:27 -0700)]
Enable stack push removal optimization by default.
PiperOrigin-RevId:
189641729
Benoit Steiner [Mon, 19 Mar 2018 21:24:00 +0000 (14:24 -0700)]
Turned on gradient optimization by default
PiperOrigin-RevId:
189641300
A. Unique TensorFlower [Mon, 19 Mar 2018 21:09:52 +0000 (14:09 -0700)]
Add a helper that allows constructing simple expression ASTs from string. Useful to simplify the representation of composite symbols, e.g. 'py2tf.foo'.
PiperOrigin-RevId:
189638901
Derek Murray [Mon, 19 Mar 2018 20:44:23 +0000 (13:44 -0700)]
Automated g4 rollback of changelist
189416074
PiperOrigin-RevId:
189634491
Alexandre Passos [Mon, 19 Mar 2018 20:43:50 +0000 (13:43 -0700)]
Moves TFE_Executor to tensorflow::EagerExecutor in tensorflow/core/common_runtime/eager
PiperOrigin-RevId:
189634404
Akshay Agrawal [Mon, 19 Mar 2018 20:38:23 +0000 (13:38 -0700)]
TFE: Fix bug encountered when using `optimizer.apply_gradients` in a defun.
Prior to this change, `Optimizer` assumed that `not
context.executing_eagerly()` implied that every variable that it was to update
was constructed in a graph. That assumption is incorrect --- TensorFlow
functions can mutate variables captured from or lifted into the eager context. As such, this change removes that assumption.
Fixes #17792
PiperOrigin-RevId:
189633630
A. Unique TensorFlower [Mon, 19 Mar 2018 20:26:19 +0000 (13:26 -0700)]
Add bfloat16 support for CPU ops.
PiperOrigin-RevId:
189631659
A. Unique TensorFlower [Mon, 19 Mar 2018 20:03:22 +0000 (13:03 -0700)]
Extract GraphOptimizer{Stage,Context}, and use it as a base class
in ArithmeticOptimizer.
PiperOrigin-RevId:
189628227
Allen Lavoie [Mon, 19 Mar 2018 20:01:58 +0000 (13:01 -0700)]
Checkpointable: Small cleanup making better use of NewCheckpointReader.
PiperOrigin-RevId:
189627956
A. Unique TensorFlower [Mon, 19 Mar 2018 19:15:59 +0000 (12:15 -0700)]
Add a map from TPU core id to name to TfOpStats.
PiperOrigin-RevId:
189620850
A. Unique TensorFlower [Mon, 19 Mar 2018 18:29:44 +0000 (11:29 -0700)]
Do not use SparseMatmul to for bfloat16 as Matmul is already supported.
PiperOrigin-RevId:
189614197
A. Unique TensorFlower [Mon, 19 Mar 2018 18:27:45 +0000 (11:27 -0700)]
Internal change
PiperOrigin-RevId:
189613870
A. Unique TensorFlower [Mon, 19 Mar 2018 18:11:28 +0000 (11:11 -0700)]
Optimizations to DepthwiseConv
PiperOrigin-RevId:
189610985
Asim Shankar [Mon, 19 Mar 2018 17:37:00 +0000 (10:37 -0700)]
Documentation tweaks and tests for GradientTape with graph execution.
PiperOrigin-RevId:
189604536
A. Unique TensorFlower [Mon, 19 Mar 2018 16:50:00 +0000 (09:50 -0700)]
Fix misc typos in tensorflow/compiler/xla.
PiperOrigin-RevId:
189596520
A. Unique TensorFlower [Mon, 19 Mar 2018 16:43:59 +0000 (09:43 -0700)]
Simple rewrite to remove negation nodes.
PiperOrigin-RevId:
189595735
A. Unique TensorFlower [Mon, 19 Mar 2018 16:42:14 +0000 (09:42 -0700)]
Refactor code to improve TensorDataSet construction speed.
PiperOrigin-RevId:
189595482
A. Unique TensorFlower [Mon, 19 Mar 2018 16:28:58 +0000 (09:28 -0700)]
Remove a few unused #includes
PiperOrigin-RevId:
189593522
A. Unique TensorFlower [Mon, 19 Mar 2018 14:29:45 +0000 (07:29 -0700)]
Adds missing protobuf dep to tf.contrib.data ops.
PiperOrigin-RevId:
189580464
A. Unique TensorFlower [Mon, 19 Mar 2018 13:27:00 +0000 (06:27 -0700)]
Adding non-linear image warping ops to tf.contrib.image
New ops are:
tf.contrib.image.sparse_image_warp, tf.contrib.image.dense_image_warp, and tf.contrib.image.interpolate_spline.
PiperOrigin-RevId:
189574951
A. Unique TensorFlower [Mon, 19 Mar 2018 12:01:05 +0000 (05:01 -0700)]
Add new helpers to HLO sharding.
PiperOrigin-RevId:
189569053
A. Unique TensorFlower [Sun, 18 Mar 2018 22:18:06 +0000 (15:18 -0700)]
Add precision and recall metrics to _BinaryLogisticHeadWithSigmoidCrossEntropyLoss.
This change makes most of the binary classifiers in the canned estimators provide precision and recall metrics during evaluation. This matches the behavior of the canned estimators defined in the deprecated tf.contrib.learn.estimator.
PiperOrigin-RevId:
189522420
A. Unique TensorFlower [Sun, 18 Mar 2018 14:48:30 +0000 (07:48 -0700)]
Fix build
PiperOrigin-RevId:
189506945
A. Unique TensorFlower [Sat, 17 Mar 2018 18:21:02 +0000 (11:21 -0700)]
Normally tf2xla (autoclustering, jit_scope and rewrite) rely on graph optimization
passes to outline subgraphs. The XLA device itself only sees Compute() calls for
_XlaLaunch ops. All other ops are registered with a dummy op factory that just
prints an error.
This patch adds an alternative, selected at registration time, that disables
default graph optimization and instead registers a non-dummy op implementation.
This op implementation compiles the op "on demand"; it generates a fake graph containing
_Arg and _Retval nodes and calls into the XlaCompiler code as usual.
This allows the device to be used as a "normal" TensorFlow device, as well as from
Eager mode, at the expense of performance.
Later additions will add the ability to create traces to amortize kernel launch overhead,
and the ability to combine op-by-op/tracing and autoclustering with jit_scope annotations.
PiperOrigin-RevId:
189463593
Peter Hawkins [Sat, 17 Mar 2018 16:57:03 +0000 (09:57 -0700)]
[XLA] Fix points-to set calculation in HLO ListScheduler.
Previously the list scheduler considered that an instruction used only the buffers defined by its operands. This is inaccurate in the presence of aliasing?an instruction may potentially use anything in the points-to set of the operand, including buffers defined by an ancestor of an operand. Change to use the full points-to set instead.
PiperOrigin-RevId:
189460681
Raghuraman Krishnamoorthi [Sat, 17 Mar 2018 02:26:56 +0000 (19:26 -0700)]
Update docs for fake quant to reflect support for bitwidths from 2 to 16 inclusive.
PiperOrigin-RevId:
189426173
Derek Murray [Sat, 17 Mar 2018 00:10:16 +0000 (17:10 -0700)]
Automated g4 rollback of changelist
189228094
PiperOrigin-RevId:
189416074
Anna R [Sat, 17 Mar 2018 00:05:51 +0000 (17:05 -0700)]
Add user_ops.my_fact to the new TensorFlow API.
PiperOrigin-RevId:
189415577
Allen Lavoie [Sat, 17 Mar 2018 00:00:17 +0000 (17:00 -0700)]
TFTS: Allow cold-starting from SavedModels
This means the model starts from its default start state and is fed a series
(filtering) to warm up its state. This warmed up state can then be used to
make predictions.
Some shape fiddling with the receiver_fn to make feeding state optional, and a
new signature for cold-starting which uses the model's default start state.
Some other shape fiddling to make feeding strings to SavedModels work more
smoothly in the cold-start part of the LSTM example. I was squeezing out the
last dimension of "scalar" exogenous features, now I'm leaving them, which
matches the placeholder generation logic.
PiperOrigin-RevId:
189414869
Skye Wanderman-Milne [Fri, 16 Mar 2018 23:27:02 +0000 (16:27 -0700)]
Relax "_output_shapes" error checking in C++ graph importer.
This is to make the behavior inline with what Python's
import_graph_def does. It'd probably be better to raise an error, but
some people are already depending on the import_graph_def behavior in
order to import modified ops.
PiperOrigin-RevId:
189411025
Alexandre Passos [Fri, 16 Mar 2018 23:10:46 +0000 (16:10 -0700)]
Always sets self._built in tfe.metrics when build() is called.
PiperOrigin-RevId:
189408745
Benoit Steiner [Fri, 16 Mar 2018 23:07:06 +0000 (16:07 -0700)]
Deleted dead code
PiperOrigin-RevId:
189408200
A. Unique TensorFlower [Fri, 16 Mar 2018 23:00:11 +0000 (16:00 -0700)]
Increase kMaxEagerTensorParentSize to 64.
PiperOrigin-RevId:
189407226
A. Unique TensorFlower [Fri, 16 Mar 2018 22:55:14 +0000 (15:55 -0700)]
Remove identity transpose nodes.
PiperOrigin-RevId:
189406518
Yuefeng Zhou [Fri, 16 Mar 2018 22:36:15 +0000 (15:36 -0700)]
Consolidate all moving_average updates in batchnorm into one implementation.
PiperOrigin-RevId:
189404070
Benoit Steiner [Fri, 16 Mar 2018 22:31:10 +0000 (15:31 -0700)]
Don't fail when optimizing the gradients of noinline functions
PiperOrigin-RevId:
189403170
Chris Leary [Fri, 16 Mar 2018 22:13:25 +0000 (15:13 -0700)]
[XLA] Fix forward for HLO profiling test, explicitly set profiling preference.
PiperOrigin-RevId:
189400869
Benoit Steiner [Fri, 16 Mar 2018 22:10:39 +0000 (15:10 -0700)]
Don't inline functions in the grappler item builder since this part of the code doesn't
support custom ops. Instead we will rely on the function optimizer.
PiperOrigin-RevId:
189400462
Skye Wanderman-Milne [Fri, 16 Mar 2018 21:39:08 +0000 (14:39 -0700)]
Downgrade run-after-mutation error to a log warning.
This is to ease the transition to the C API. Some tests currently mutate the
graph after running it but currently pass. This error was meant to guard
against existing behavior, so it's not a regression to make it a warning
instead for now.
PiperOrigin-RevId:
189395472
A. Unique TensorFlower [Fri, 16 Mar 2018 20:43:12 +0000 (13:43 -0700)]
Fixed a typo.
PiperOrigin-RevId:
189386932
Jianwei Xie [Fri, 16 Mar 2018 20:11:36 +0000 (13:11 -0700)]
Estimate prediction size and error out if it is larger than protobuf limit.
PiperOrigin-RevId:
189382429
A. Unique TensorFlower [Fri, 16 Mar 2018 20:00:31 +0000 (13:00 -0700)]
Move if_op kernel to //third_party/tensorflow/compiler/tf2xla/kernels
PiperOrigin-RevId:
189381067
Yuanzhong Xu [Fri, 16 Mar 2018 19:56:26 +0000 (12:56 -0700)]
[XLA] BF16 conversion folding for CRS; remove no-op conversions in propagation.
If CRS has tuple output, it needs special handling in conversion folding.
BF16 propagation could result in BF16->BF16 conversions, which can be removed.
PiperOrigin-RevId:
189380578
A. Unique TensorFlower [Fri, 16 Mar 2018 19:55:18 +0000 (12:55 -0700)]
BREAKING_CHANGE: Remove SigmoidCentered bijector.
- SoftmaxCentered solely works on vector events, and supports broadcasting.
- Sigmoid exists for event_ndims=0 cases.
PiperOrigin-RevId:
189380445
Chris Leary [Fri, 16 Mar 2018 19:34:34 +0000 (12:34 -0700)]
[XLA:python] Plumb hlo_profile flag.
PiperOrigin-RevId:
189377860
A. Unique TensorFlower [Fri, 16 Mar 2018 19:12:52 +0000 (12:12 -0700)]
Remove empty buckets in latency_stats as it makes the report unreadable. This is also consistent with how SummaryHistoOp in summary_op.cc works.
PiperOrigin-RevId:
189375113
Derek Murray [Fri, 16 Mar 2018 19:11:18 +0000 (12:11 -0700)]
[tf.data] Fix typo in `Dataset.prefetch()` docstring.
PiperOrigin-RevId:
189374898
Sanjoy Das [Fri, 16 Mar 2018 18:53:14 +0000 (11:53 -0700)]
[TF:XLA] Bump open source llvm revision to r327616
PiperOrigin-RevId:
189372065