A. Unique TensorFlower [Fri, 19 Jan 2018 19:01:45 +0000 (11:01 -0800)]
Add option for loop unrolling in MaskedAutoregressiveFlow bijector.
PiperOrigin-RevId:
182559030
Alexandre Passos [Fri, 19 Jan 2018 18:36:05 +0000 (10:36 -0800)]
Fixes issue with using lookup ops in datasets from eager.
Addresses issue #16160
PiperOrigin-RevId:
182554969
Anna R [Fri, 19 Jan 2018 18:32:15 +0000 (10:32 -0800)]
Fixes #12466. Fix getfile in cases when .py files aren't available (only .pyc files are
used).
PiperOrigin-RevId:
182554325
A. Unique TensorFlower [Fri, 19 Jan 2018 17:53:10 +0000 (09:53 -0800)]
Fix documentation typo.
PiperOrigin-RevId:
182548006
A. Unique TensorFlower [Fri, 19 Jan 2018 17:46:59 +0000 (09:46 -0800)]
Change deprecated dim parameter to axis.
PiperOrigin-RevId:
182547126
Benoit Steiner [Fri, 19 Jan 2018 17:12:34 +0000 (09:12 -0800)]
Improved detection of swappable nodes
PiperOrigin-RevId:
182542749
Peter Hawkins [Fri, 19 Jan 2018 17:06:40 +0000 (09:06 -0800)]
[TF:XLA] Include a summary of the TF operator when TF -> XLA compilation fails.
PiperOrigin-RevId:
182542114
Ilya Biryukov [Fri, 19 Jan 2018 11:11:14 +0000 (03:11 -0800)]
Updated the version of downloadable chromium's clang toolchain
PiperOrigin-RevId:
182511847
A. Unique TensorFlower [Fri, 19 Jan 2018 07:21:48 +0000 (23:21 -0800)]
Fix typo in MklSubAllocator::ClearStats()
PiperOrigin-RevId:
182492508
A. Unique TensorFlower [Fri, 19 Jan 2018 02:54:45 +0000 (18:54 -0800)]
Support writing np.ndarray as tensors in Summary proto.
PiperOrigin-RevId:
182474037
Anna R [Fri, 19 Jan 2018 01:16:06 +0000 (17:16 -0800)]
Script to create ApiDef files automatically based on OpDef.
PiperOrigin-RevId:
182463327
A. Unique TensorFlower [Fri, 19 Jan 2018 01:02:02 +0000 (17:02 -0800)]
Reshaping vars correctly in _compute_fraction_of_zero.
PiperOrigin-RevId:
182461459
Alexandre Passos [Fri, 19 Jan 2018 00:51:40 +0000 (16:51 -0800)]
Avoid use-after-free in eager tape code.
PiperOrigin-RevId:
182460292
A. Unique TensorFlower [Fri, 19 Jan 2018 00:20:14 +0000 (16:20 -0800)]
[TF:XLA] Make ResizeBilinear and ResizeBilinearGrad faster for images of size 2^x+1.
PiperOrigin-RevId:
182456096
A. Unique TensorFlower [Fri, 19 Jan 2018 00:03:32 +0000 (16:03 -0800)]
Fix some (re)shaping issues.
PiperOrigin-RevId:
182453707
A. Unique TensorFlower [Thu, 18 Jan 2018 23:50:21 +0000 (15:50 -0800)]
Remove extra repeat in tf.data.Dataset documentation.
END_PUBLIC
Unless I am missing something:
BEGIN_PUBLIC
If num_epochs is 0, this isn't needed anyway; and if num_epochs > 0 then this will be buggy, because it will cause the output dataset to repeat for more than num_epochs.
PiperOrigin-RevId:
182451811
Nick Felt [Thu, 18 Jan 2018 23:50:16 +0000 (15:50 -0800)]
Update tensorboard dependency to minimum of 0.4.0
This should address https://github.com/tensorflow/tensorboard/issues/877.
PiperOrigin-RevId:
182451796
Saurabh Saxena [Thu, 18 Jan 2018 23:27:12 +0000 (15:27 -0800)]
Handle case when buffer_size is larger than total dataset size in shuffle dataset op.
Currently ShuffleDatasetOpBase::GetNext throws a "Attempted to repeat an empty dataset infinitely." if the number of outputs produced from the dataset are < buffer_size.
Usually the buffer size is set to <= total dataset size so we never run into this.
PiperOrigin-RevId:
182448427
Benoit Steiner [Thu, 18 Jan 2018 23:08:48 +0000 (15:08 -0800)]
Entice TensorFlow to swap data out of GPU memory sooner
PiperOrigin-RevId:
182445531
A. Unique TensorFlower [Thu, 18 Jan 2018 22:28:45 +0000 (14:28 -0800)]
Add a convenient function that makes an HloProto from an HloModule.
PiperOrigin-RevId:
182438633
Sung Jin Hwang [Thu, 18 Jan 2018 22:24:50 +0000 (14:24 -0800)]
Fixed Eigen-version Conv2DBackpropInput kernel, marked with "eigen_tensor".
LaunchConv2DBackpropInputOp, a template functor class used in
Conv2DFastBackpropInputOp class was specialized in a wrong place, in
conv_grad_filter_ops.cc (ODR violation). This caused wrong functor to be called
inside the op kernel class, because the correct functor definition was overriden
by a wrong specialization. The functor class that should have been specialized
in conv_grad_filter_ops.cc was LaunchConv2DBackpropFilterOp.
PiperOrigin-RevId:
182438025
Anna R [Thu, 18 Jan 2018 22:20:56 +0000 (14:20 -0800)]
Use snake-case name when generating @tf_export decorator in python_op_gen.cc.
PiperOrigin-RevId:
182437341
A. Unique TensorFlower [Thu, 18 Jan 2018 22:20:14 +0000 (14:20 -0800)]
Make bfloat16 work correctly with matmul
PiperOrigin-RevId:
182437226
Tayo Oguntebi [Thu, 18 Jan 2018 22:01:07 +0000 (14:01 -0800)]
Adds ReduceWindow R1 test case for windows of length 128.
PiperOrigin-RevId:
182433727
Derek Murray [Thu, 18 Jan 2018 21:36:52 +0000 (13:36 -0800)]
[tf.data] Update API guide to include latest methods.
PiperOrigin-RevId:
182430038
A. Unique TensorFlower [Thu, 18 Jan 2018 21:31:04 +0000 (13:31 -0800)]
Supports Squeeze in Tf Lite.
PiperOrigin-RevId:
182429180
Yu-Cheng Ling [Thu, 18 Jan 2018 21:16:01 +0000 (13:16 -0800)]
Bazel rule to generate LICENSE for TFLite C library.
PiperOrigin-RevId:
182426880
Benoit Steiner [Thu, 18 Jan 2018 21:07:51 +0000 (13:07 -0800)]
Don't try to swap out persistent data
PiperOrigin-RevId:
182425781
Sanjoy Das [Thu, 18 Jan 2018 20:53:42 +0000 (12:53 -0800)]
[TF:XLA] Bump open source llvm revision to r322850
PiperOrigin-RevId:
182423651
A. Unique TensorFlower [Thu, 18 Jan 2018 20:50:29 +0000 (12:50 -0800)]
Add basic class support:
* frontend API extended to allow converting an entire class
* function conversion now supports object methods
* responsibility for renaming the top level object is now moved up from
the call tree transformer - this is because functions may or may not be renamed based on context (e.g. object members are not renamed)
Not yet implemented:
* static analysis and side effect protection for object fields
PiperOrigin-RevId:
182423203
A. Unique TensorFlower [Thu, 18 Jan 2018 20:26:38 +0000 (12:26 -0800)]
Add tests for R2 reduce window with multiply reduction.
PiperOrigin-RevId:
182419763
Derek Murray [Thu, 18 Jan 2018 20:18:49 +0000 (12:18 -0800)]
[tf.data] Further simplify the `CapturedFunction::Run*()` interfaces.
All users of these interfaces had to use the same boilerplate to create a
`FunctionLibraryRuntime::Options`. This change moves that boilerplate inside
the `CapturedFunction` implementation.
PiperOrigin-RevId:
182418501
A. Unique TensorFlower [Thu, 18 Jan 2018 19:42:09 +0000 (11:42 -0800)]
Enable a no longer flaky test.
PiperOrigin-RevId:
182412902
Yu-Cheng Ling [Thu, 18 Jan 2018 19:31:21 +0000 (11:31 -0800)]
TFLite iOS Makefile: Enable SSE4.1 for x86_64 build.
This makes the simulator slightly faster.
PiperOrigin-RevId:
182411030
Kay Zhu [Thu, 18 Jan 2018 19:30:02 +0000 (11:30 -0800)]
[TF:JIT] Make :xla_device visible to public, which is necessary for 3rd party
to use XLA. Also add a build target to tf/compiler/plugin/BUILD that depends on
:xla_device to help ensure it stays visible to public.
PiperOrigin-RevId:
182410844
Allen Lavoie [Thu, 18 Jan 2018 19:09:26 +0000 (11:09 -0800)]
Disable flaky test
Due to a known input function issue.
PiperOrigin-RevId:
182407471
Peter Hawkins [Thu, 18 Jan 2018 18:51:41 +0000 (10:51 -0800)]
[TF:XLA] Add support for atrous (RHS dilated) convolutions via the "dilations" attribute on convolution operators.
PiperOrigin-RevId:
182404352
Andrew Selle [Thu, 18 Jan 2018 18:45:25 +0000 (10:45 -0800)]
Warn against using the name RELU1 again as an op.
PiperOrigin-RevId:
182403185
Sanjoy Das [Thu, 18 Jan 2018 18:29:39 +0000 (10:29 -0800)]
[XLA:AOT] Emit a .o file to hold the ProgramShape protobuf
This changes tfcompile to emit two .o files, one containing the TF model, and
one containing "helpers" which for now only contains the relevant
xla::ProgramShape protobuf in serialized form.
PiperOrigin-RevId:
182400539
A. Unique TensorFlower [Thu, 18 Jan 2018 18:24:13 +0000 (10:24 -0800)]
Adds SUM_OVER_BATCH_SIZE in losses.Reduction.
PiperOrigin-RevId:
182399706
A. Unique TensorFlower [Thu, 18 Jan 2018 18:10:07 +0000 (10:10 -0800)]
Allow numpy_input_fn to take an ndarray representing a single input feature.
PiperOrigin-RevId:
182397583
A. Unique TensorFlower [Thu, 18 Jan 2018 16:47:09 +0000 (08:47 -0800)]
Go: Update generated wrapper functions for TensorFlow ops.
PiperOrigin-RevId:
182384458
A. Unique TensorFlower [Thu, 18 Jan 2018 16:18:22 +0000 (08:18 -0800)]
Update ops-related pbtxt files.
PiperOrigin-RevId:
182378232
A. Unique TensorFlower [Thu, 18 Jan 2018 16:10:50 +0000 (08:10 -0800)]
Added matrix logarithm op to linear algebra ops.
PiperOrigin-RevId:
182376730
A. Unique TensorFlower [Thu, 18 Jan 2018 15:27:48 +0000 (07:27 -0800)]
Fix Markdown for examples in WarmStartSettings docstring.
PiperOrigin-RevId:
182368459
A. Unique TensorFlower [Thu, 18 Jan 2018 15:03:48 +0000 (07:03 -0800)]
Expose Sequential as tfe.Sequential to go with the already exposed tfe.Network.
PiperOrigin-RevId:
182365046
A. Unique TensorFlower [Thu, 18 Jan 2018 09:35:01 +0000 (01:35 -0800)]
Basic feature selection for boosted trees.
The idea is that we grow the trees normally until we reach the requested number of unique features, and once we reach that limit, we avoid using any new features.
PiperOrigin-RevId:
182336278
A. Unique TensorFlower [Thu, 18 Jan 2018 06:28:11 +0000 (22:28 -0800)]
[XLA] Add Pad to the local Python XLA client.
PiperOrigin-RevId:
182320104
Chris Leary [Thu, 18 Jan 2018 05:25:40 +0000 (21:25 -0800)]
[XLA:python] Simplify frame data extraction, fix Py3 compatibility bugs.
PiperOrigin-RevId:
182315965
A. Unique TensorFlower [Thu, 18 Jan 2018 05:14:42 +0000 (21:14 -0800)]
Add WarmStartSettings configuration for canned Estimators.
PiperOrigin-RevId:
182315322
Justine Tunney [Thu, 18 Jan 2018 04:55:46 +0000 (20:55 -0800)]
Say SQLite BLOB protos are ideally temporary
We ideally want a schema that's pure SQL and doesn't embed
protos. However some protos we don't have time to translate
quite yet. So we put a BLOB as the last field so it can be
migrated later.
Please note there might end up being some things highly
specific to TensorFlow and/or Google that end up here. In
those cases we might consider keeping the proto blobs
permanently, if they're difficult to express in SQL.
PiperOrigin-RevId:
182313847
Jianwei Xie [Thu, 18 Jan 2018 02:56:27 +0000 (18:56 -0800)]
Slightly tune the usage of ctx to be consistent with other places.
PiperOrigin-RevId:
182305106
A. Unique TensorFlower [Thu, 18 Jan 2018 01:49:23 +0000 (17:49 -0800)]
Dynamically print helper message based on user queries.
PiperOrigin-RevId:
182298241
A. Unique TensorFlower [Thu, 18 Jan 2018 01:13:03 +0000 (17:13 -0800)]
internal change
PiperOrigin-RevId:
182294233
Brennan Saeta [Thu, 18 Jan 2018 01:07:11 +0000 (17:07 -0800)]
Nicer error messages when missing filesystem.
PiperOrigin-RevId:
182293612
Sung Jin Hwang [Thu, 18 Jan 2018 01:05:21 +0000 (17:05 -0800)]
Add range encode and decode to new tensorflow contrib.
PiperOrigin-RevId:
182293433
Rohan Jain [Thu, 18 Jan 2018 00:56:14 +0000 (16:56 -0800)]
Changes the CapturedFunction implementation to not create a new FunctionLibraryRuntime etc. and just re-use the existing one provided. This allows for Dataset ops to run cross device and cross process.
PiperOrigin-RevId:
182292325
Chris Leary [Thu, 18 Jan 2018 00:54:30 +0000 (16:54 -0800)]
[XLA] Add source mapping support to SWIG API.
PiperOrigin-RevId:
182292142
Blake Hechtman [Thu, 18 Jan 2018 00:49:54 +0000 (16:49 -0800)]
Use explicit broadcasts in ResizeBilinear Kernel creation to make fusion easier.
PiperOrigin-RevId:
182291519
A. Unique TensorFlower [Thu, 18 Jan 2018 00:30:21 +0000 (16:30 -0800)]
Update contrib/all_reduce/python/all_reduce.py in consequence of
earlier change to nccl.broadcast function signature.
PiperOrigin-RevId:
182288943
Yuefeng Zhou [Wed, 17 Jan 2018 23:58:10 +0000 (15:58 -0800)]
Combine host and device memory proto fields.
PiperOrigin-RevId:
182284426
Kiril Gorovoy [Wed, 17 Jan 2018 23:28:25 +0000 (15:28 -0800)]
Make _check_bazel_version_at_least in workspace.bzl public.
PiperOrigin-RevId:
182280342
A. Unique TensorFlower [Wed, 17 Jan 2018 23:07:39 +0000 (15:07 -0800)]
Prune away Identity nodes without regular outputs. (Addresses TODO from cl/
175780178).
PiperOrigin-RevId:
182277131
Amit Patankar [Wed, 17 Jan 2018 22:44:17 +0000 (14:44 -0800)]
Disabling flaky replicate_model_fn_test.py for cmake.
PiperOrigin-RevId:
182273024
Frank Chen [Wed, 17 Jan 2018 22:29:30 +0000 (14:29 -0800)]
Add functionality to auto-discover project and zone when they are not supplied to the TPUClusterResolver
PiperOrigin-RevId:
182270565
A. Unique TensorFlower [Wed, 17 Jan 2018 22:23:28 +0000 (14:23 -0800)]
Update ops-related pbtxt files.
PiperOrigin-RevId:
182269654
A. Unique TensorFlower [Wed, 17 Jan 2018 21:58:05 +0000 (13:58 -0800)]
K-FAC: Fixes problem where initial eigendecomposition used by RNN classes would be computed during the call to instantiate_factors (via the "registration" functions) instead of the call to make_inverse_update_ops. This messed up the device placement of these ops and interacted badly with other parts of the code.
Also changed how covariance op creation is done in ConvDiagonalFactor in anticipation of similar problems in the future.
As of this CL, none of the op creation methods will modify the state of the class, and no ops will be created outside of the op creation methods. We should try to follow this convention going forward.
PiperOrigin-RevId:
182265266
Igor Ganichev [Wed, 17 Jan 2018 21:46:17 +0000 (13:46 -0800)]
Add TF_GraphNumFunctions and TF_GraphGetFunctions
PiperOrigin-RevId:
182263576
Sanjoy Das [Wed, 17 Jan 2018 21:41:05 +0000 (13:41 -0800)]
[TF:XLA] Bump open source llvm revision to r322618
PiperOrigin-RevId:
182262817
A. Unique TensorFlower [Wed, 17 Jan 2018 21:29:50 +0000 (13:29 -0800)]
Adding support for to resolve constant FloorDiv, FloorMod, StridedSlice, Stack, Rank and Range.
PiperOrigin-RevId:
182261052
Sourabh Bajaj [Wed, 17 Jan 2018 21:14:48 +0000 (13:14 -0800)]
Merge changes from github.
PiperOrigin-RevId:
182258809
Tayo Oguntebi [Wed, 17 Jan 2018 20:55:00 +0000 (12:55 -0800)]
* Adds rank-3 ReduceWindow computation to the reference util test suite.
* Adds rank-3 test harness for ReduceWindow, along with some test cases.
* Minor style correction.
PiperOrigin-RevId:
182256143
Yu-Cheng Ling [Wed, 17 Jan 2018 20:28:46 +0000 (12:28 -0800)]
Make int64 inputs workable, and testable in generate_examples_test.
PiperOrigin-RevId:
182252829
A. Unique TensorFlower [Wed, 17 Jan 2018 20:20:20 +0000 (12:20 -0800)]
[XLA:GPU] Fix a problem in calculating the address of the memory used to
implement small data type atomic operations.
To calculate the address of the enclosing 4 byte memory,
output_address&(-2) was used while it should be output_address&(-4).
Add a test case.
PiperOrigin-RevId:
182251760
Shanqing Cai [Wed, 17 Jan 2018 19:51:07 +0000 (11:51 -0800)]
TFE: Improve spinn/README.md
PiperOrigin-RevId:
182247687
Igor Saprykin [Wed, 17 Jan 2018 19:37:48 +0000 (11:37 -0800)]
Make `replicate_model_fn` friendlier to distributed training.
I verified that async distributed training works as is. One quirk is that when replicating over a single GPU, variables end up being placed on /gpu:0 on PSs, which works correctly only thanks to allow_soft_placement=True.
For sync distributed training using SyncReplicasOptimizer the only quirk is that SyncReplicasOptimizerHook insists on SyncReplicasOptimizer.apply_gradients to be called. That happens only in the last tower, yet any tower could create the hook. To accommodate that requirement hooks from the last tower are taken as part of this CL. Before this, hooks from the first tower were taken.
SyncReplicasOptimizer doesn't behave perfectly in tests. The queue keeps hanging waiting for new token to arrive until `stop_grace_period_seconds` which is set for 120 seconds. The latter isn't exposed through the Estimator interface, which means the test is slower.
PiperOrigin-RevId:
182245657
Dan Ringwalt [Wed, 17 Jan 2018 19:33:12 +0000 (11:33 -0800)]
Add tf.contrib.image.connected_components.
Comparable to scipy.ndimage.measurements.label.
PiperOrigin-RevId:
182244926
Yuanzhong Xu [Wed, 17 Jan 2018 19:32:05 +0000 (11:32 -0800)]
Allow specifying the device in CreateParameterAndTransferLiteral.
PiperOrigin-RevId:
182244708
A. Unique TensorFlower [Wed, 17 Jan 2018 19:05:24 +0000 (11:05 -0800)]
Internal change.
PiperOrigin-RevId:
182240175
A. Unique TensorFlower [Wed, 17 Jan 2018 17:55:59 +0000 (09:55 -0800)]
Make logging more accurate
PiperOrigin-RevId:
182229043
A. Unique TensorFlower [Wed, 17 Jan 2018 17:49:03 +0000 (09:49 -0800)]
Adds regularization_losses in canned heads.
PiperOrigin-RevId:
182228149
Mark Heffernan [Wed, 17 Jan 2018 17:41:43 +0000 (09:41 -0800)]
Add instruction count method to HloModule.
PiperOrigin-RevId:
182227249
Andrew Selle [Wed, 17 Jan 2018 17:36:55 +0000 (09:36 -0800)]
Create support for div and sub in tensorflow lite.
PiperOrigin-RevId:
182226574
A. Unique TensorFlower [Wed, 17 Jan 2018 16:22:06 +0000 (08:22 -0800)]
export_strategy: Add strip_default_attrs to ExportStrategy and it's export_fn signature.
This allows for a single place (ExportStrategy) where the config for strip_default_attrs is stored. If the export_fn callback accepts the `strip_default_attrs` argument, it can
then propagate the `strip_default_attrs` arg to estimator.export_savedmodel().
Also changed the default for `strip_default_attrs` in the ExportStrategy factory methods (saved_model_export_utils.py) to None. This allows clients that process a given ExportStrategy object to determine if a value has been provided for ExportStrategy.strip_default_attrs field (i.e., `is not None`).
If no value has been provided, those clients can then use whatever default they wish to use.
PiperOrigin-RevId:
182217376
A. Unique TensorFlower [Wed, 17 Jan 2018 16:11:09 +0000 (08:11 -0800)]
Rename the newly introduced TF_SetAttrFunc function to make it clear that it is
setting the attribute to a function name. This permits future patches to add
things like TF_SetAttrFuncAttrList if necessary later.
PiperOrigin-RevId:
182216098
A. Unique TensorFlower [Wed, 17 Jan 2018 13:03:05 +0000 (05:03 -0800)]
K-FAC: Expose protected functions from fisher_blocks and fisher_factors and constant strings from layer_collection in their respective library modules. This allows consistent development of blocks and factors outside tensorflow.contrib.kfac.
PiperOrigin-RevId:
182197356
A. Unique TensorFlower [Wed, 17 Jan 2018 06:55:02 +0000 (22:55 -0800)]
[XLA] Make the hlo_evaluator handle reduce window in the same ways as the cpu backend. Also,
- Added a test that would fail without this change.
- Enabled reduce_window_test_interpreter.
- Added the DISALBED_ON_INTERPRETER macro and disalbed large tests on the interpreter.
The evaluator assumed the input value to be 0 when out of bound, which is likely to produce wrong output when the reduction function is not Add.
PiperOrigin-RevId:
182164740
Gunhan Gulsoy [Wed, 17 Jan 2018 05:48:26 +0000 (21:48 -0800)]
Bump the size of keras:core_test as it is flakily timing out.
PiperOrigin-RevId:
182159937
A. Unique TensorFlower [Wed, 17 Jan 2018 04:47:19 +0000 (20:47 -0800)]
Automated g4 rollback of changelist
182127737
PiperOrigin-RevId:
182155292
A. Unique TensorFlower [Wed, 17 Jan 2018 03:52:30 +0000 (19:52 -0800)]
[TF:XLA] Remove workaround for b/
37575001.
The lowering bug has been long fixed, so the workaround of constructing R3s is no longer required.
PiperOrigin-RevId:
182150664
A. Unique TensorFlower [Wed, 17 Jan 2018 03:02:15 +0000 (19:02 -0800)]
Introduce a new C API entrypoint to set a 'func' attribute on an op
description.
PiperOrigin-RevId:
182146876
Justine Tunney [Wed, 17 Jan 2018 02:25:34 +0000 (18:25 -0800)]
Disable exceptions in TensorFlow jsoncpp build
This is necessary to compile with Clang. Since the TF GCP code doesn't have any
try/catch statements, this could be a bug fix.
PiperOrigin-RevId:
182143536
Justine Tunney [Wed, 17 Jan 2018 02:24:54 +0000 (18:24 -0800)]
Make SQLite binary smaller in TensorFlow
This makes it 400kB smaller. This is worth it because even
-O6 would only make SQLite 3% faster.
https://www.sqlite.org/footprint.html
PiperOrigin-RevId:
182143488
A. Unique TensorFlower [Wed, 17 Jan 2018 02:24:39 +0000 (18:24 -0800)]
Add decay/epsilon/updates_collections parameters to Inception v3 arg_scope
PiperOrigin-RevId:
182143473
A. Unique TensorFlower [Wed, 17 Jan 2018 02:20:02 +0000 (18:20 -0800)]
Remove a duplicate bfloat16 registration.
PiperOrigin-RevId:
182143007
A. Unique TensorFlower [Wed, 17 Jan 2018 01:21:12 +0000 (17:21 -0800)]
K-FAC: Example using tf.estimator and K-FAC.
- Removes FisherEstimator.inv_updates_dict. Users should create directly from
FisherEstimator.inv_update_ops.
- Adds (cov|inv)_update_(thunks|ops) to KfacOptimizer.
PiperOrigin-RevId:
182135826
A. Unique TensorFlower [Wed, 17 Jan 2018 00:48:56 +0000 (16:48 -0800)]
Allow ReorderAxes to work on QUINT8 arrays. This will be necessary in order to support models that include tensorflow's DequantizeOp (such models were "quantized" using tensorflow's transform_graph tool).
PiperOrigin-RevId:
182131833
Saurabh Saxena [Wed, 17 Jan 2018 00:43:13 +0000 (16:43 -0800)]
Fix flakiness of dataset_constructor_op_test and flat_map_dataset_op_test due to
timeouts.
PiperOrigin-RevId:
182131168
Eugene Brevdo [Wed, 17 Jan 2018 00:17:36 +0000 (16:17 -0800)]
[TF] Fix XLA Control Flow gradient stacks max_size creation.
Stack creation uses tf.while_loop's maximum_iterations iff the while_loop
was created inside an XLA/TPU context. Added several error checks to ensure
this provides useful error messages if the limited use case is not supported.
PiperOrigin-RevId:
182128135
Michael Case [Wed, 17 Jan 2018 00:14:57 +0000 (16:14 -0800)]
Internal Change.
PiperOrigin-RevId:
182127737