A. Unique TensorFlower [Fri, 26 Jan 2018 11:50:58 +0000 (03:50 -0800)]
Fix override annotations
PiperOrigin-RevId:
183367326
A. Unique TensorFlower [Fri, 26 Jan 2018 09:17:13 +0000 (01:17 -0800)]
Fix checkpoint_utils.init_from_checkpoint() to be deterministic.
PiperOrigin-RevId:
183354193
Guangda Lai [Fri, 26 Jan 2018 07:59:19 +0000 (23:59 -0800)]
Set up TensorRT configurations for external use, and add a test.
PiperOrigin-RevId:
183347199
Tayo Oguntebi [Fri, 26 Jan 2018 07:37:20 +0000 (23:37 -0800)]
Adds R1 test for ReduceWindow.
PiperOrigin-RevId:
183345779
A. Unique TensorFlower [Fri, 26 Jan 2018 06:45:59 +0000 (22:45 -0800)]
Go: Update generated wrapper functions for TensorFlow ops.
PiperOrigin-RevId:
183342483
Gunhan Gulsoy [Fri, 26 Jan 2018 06:32:15 +0000 (22:32 -0800)]
For windows cmake build turn on CMAKE_SUPPRESS_REGENERATION to avoid flaky
build failures.
PiperOrigin-RevId:
183341561
A. Unique TensorFlower [Fri, 26 Jan 2018 05:57:50 +0000 (21:57 -0800)]
Clarified documentation on resize_images.align_corners parameter.
PiperOrigin-RevId:
183339087
Max Galkin [Fri, 26 Jan 2018 04:27:11 +0000 (20:27 -0800)]
Log more info about the ill-formed node in ComputeTransitiveFanin.
PiperOrigin-RevId:
183333452
A. Unique TensorFlower [Fri, 26 Jan 2018 04:26:35 +0000 (20:26 -0800)]
internal change
PiperOrigin-RevId:
183333411
Chris Leary [Fri, 26 Jan 2018 03:50:17 +0000 (19:50 -0800)]
[XLA] Add source mapping utility translation unit, use it in the local client.
PiperOrigin-RevId:
183331075
A. Unique TensorFlower [Fri, 26 Jan 2018 03:10:13 +0000 (19:10 -0800)]
[tpu:profiler] Add infeed enqueue operation data to tf_op_stats.proto.
PiperOrigin-RevId:
183328456
Francois Chollet [Fri, 26 Jan 2018 03:05:28 +0000 (19:05 -0800)]
Update tf.keras to the Keras 2.1.3 API.
PiperOrigin-RevId:
183328052
A. Unique TensorFlower [Fri, 26 Jan 2018 02:46:51 +0000 (18:46 -0800)]
Set size of test //third_party/tensorflow/python/data/kernel_tests:dataset_from_generator_op_test to medium
It sometimes takes longer than a minute, and thus gets flaky timeouts.
PiperOrigin-RevId:
183326334
A. Unique TensorFlower [Fri, 26 Jan 2018 02:42:44 +0000 (18:42 -0800)]
Allow passing dummy/custom minmax information on a per-array basis,
unlike the existing --default_ranges_{min,max} flags which only allowed
to set a single global value for all arrays.
This takes the form of a new embedded message in ModelFlags, which is
its own message so that it can be serialized separately. The command-line
interface is --arrays_extra_info_file=some_proto.pbtxt, i.e. we don't
try to make a command-line-flags-only interface, we mandate putting the info
in a file. The rationale is that users may want to specify custom minmax
for hundreds of arrays, so it would be cumbersome to have that all in a
command line.
This should be considered an experimental feature, in the sense that
in properly quantized models, minmax information is already embedded
in the graph (e.g. in FakeQuant nodes). This is an extension of the
existing --default_ranges_{min,max} feature which had turned out to be
too restrictive for many users.
PiperOrigin-RevId:
183326000
Adam Roberts [Fri, 26 Jan 2018 01:58:48 +0000 (17:58 -0800)]
Add input_shape to seq2seq helpers.
PiperOrigin-RevId:
183321394
Akshay Agrawal [Fri, 26 Jan 2018 01:50:40 +0000 (17:50 -0800)]
Replace instances of `control_dependencies(None)` with `init_scope` when
initializing variables.
Today, when variables are constructed (Resource and otherwise), we lift certain
operations, including the VarHandleOp and initialization ops, out of all
control flow contexts; the mechanism for doing so is entering the context
manager returned by `control_dependencies(None)`. This change replaces various
instances of this mechanism with `init_scope`, which clears control dependencies, lifts ops out of function-building graphs, and pauses the gradient tape. As a result,
variables that are created inside graph functions will be automatically hoisted
into an outer context.
PiperOrigin-RevId:
183320576
A. Unique TensorFlower [Fri, 26 Jan 2018 01:32:12 +0000 (17:32 -0800)]
Move flatbuffer verifier to a separate lib
PiperOrigin-RevId:
183318384
Yao Zhang [Fri, 26 Jan 2018 01:27:42 +0000 (17:27 -0800)]
Add an option to input a GraphDef.
PiperOrigin-RevId:
183317862
Skye Wanderman-Milne [Fri, 26 Jan 2018 01:20:07 +0000 (17:20 -0800)]
Make kernel_tests/scalar_test.py work with the C API enabled.
This also moves the set_producer_version function from a specific test
file to test_util.py, since it's needed in two test files now.
PiperOrigin-RevId:
183316990
A. Unique TensorFlower [Fri, 26 Jan 2018 01:08:50 +0000 (17:08 -0800)]
[XLA] Disable half_test_cpu as it is flaky
PiperOrigin-RevId:
183315762
Yuefeng Zhou [Fri, 26 Jan 2018 00:50:24 +0000 (16:50 -0800)]
Record requested cpu cores in OpPerformance.
PiperOrigin-RevId:
183313321
Justin Lebar [Fri, 26 Jan 2018 00:45:52 +0000 (16:45 -0800)]
Automated g4 rollback of changelist
183296506
PiperOrigin-RevId:
183312680
Benoit Steiner [Fri, 26 Jan 2018 00:45:04 +0000 (16:45 -0800)]
Deleted unused data fields
PiperOrigin-RevId:
183312596
Skye Wanderman-Milne [Fri, 26 Jan 2018 00:15:50 +0000 (16:15 -0800)]
Make batch_sequences_with_states_test.py work with C API enabled, take 2.
This fixes the original rollback by using placeholders for the
SparseTensor shapes. The flakiness was caused by the nondeterministic
ordering of the sequences dict.
PiperOrigin-RevId:
183308774
Max Galkin [Fri, 26 Jan 2018 00:08:13 +0000 (16:08 -0800)]
Show friendlier error message on failure in tf_optimizer.i
Without it we trigger a segmentation fault, but later in a different stack, which is not so helpful.
PiperOrigin-RevId:
183307729
Yuefeng Zhou [Thu, 25 Jan 2018 23:59:12 +0000 (15:59 -0800)]
In the TF cost model, ensure when id grows, records for each op's outputs match the number of outputs.
PiperOrigin-RevId:
183306267
A. Unique TensorFlower [Thu, 25 Jan 2018 23:25:58 +0000 (15:25 -0800)]
Adds loss_fn argument in remaining heads.
PiperOrigin-RevId:
183301479
A. Unique TensorFlower [Thu, 25 Jan 2018 23:01:30 +0000 (15:01 -0800)]
[XLA:GPU] Implement conditional as a sequence of thunks in the GPU backend.
This also includes the following fixes:
(1) Update buffer assignment for conditionals so that the buffers corresponding to the true operand and the true computation parameter are colocated, and similarly, the buffers corresponding to the false operand and the false computation parameter are colocated.
(2) Update GPU copy insertion pass to insert copies when constants appear as operands of conditional instructions.
PiperOrigin-RevId:
183297282
Justin Lebar [Thu, 25 Jan 2018 22:56:38 +0000 (14:56 -0800)]
[XLA] Add HLO matcher for CustomCall that accepts a call target.
PiperOrigin-RevId:
183296506
Peter Hawkins [Thu, 25 Jan 2018 22:53:37 +0000 (14:53 -0800)]
Whitelist "bool" as a valid TPU infeed type.
PiperOrigin-RevId:
183296017
A. Unique TensorFlower [Thu, 25 Jan 2018 22:42:03 +0000 (14:42 -0800)]
Drop the manually_create field from RnnState.
Initially, I thought that the shape of RNN state arrays could always be
determined by shape propagation. Then I came across some graphs where this
wasn't so easy to infer, so I introduced manually_create thinking of it
as a hack. Today I took another look at dropping that hack, and had a
"D'oh" moment when I realized that the cyclic nature of RNN graphs makes
it impossible to infer the shapes of all arrays by usual propagation.
For example, in a LSTM cell, the input array is concatenated with
a state array, so if we don't already know the shape of that state array,
shape propagation stops there.
Thus, this change removes manually_create by making toco always behave as
if manually_create=true, i.e. early-creating all RNN state arrays with
the shape explicitly specified by the user. The next TODO item here
(see model_flags.proto) is to introduce a generic 'shape' field, so far
the current 'size' field only allows specifying 1-D shapes.
PiperOrigin-RevId:
183294102
Jonathan Hseu [Thu, 25 Jan 2018 22:39:44 +0000 (14:39 -0800)]
Use LookupOrCreateResource when creating summary writers. The resource may
already exist because it's associated with the device.
PiperOrigin-RevId:
183293761
A. Unique TensorFlower [Thu, 25 Jan 2018 22:38:57 +0000 (14:38 -0800)]
Add OPENSOURCE extension to schema_generated.h
PiperOrigin-RevId:
183293637
Yunxing Dai [Thu, 25 Jan 2018 22:30:15 +0000 (14:30 -0800)]
Make select_and_scatter_test optonly
- Make select_and_scatter_test optonly.
- Reduce number of shards.
PiperOrigin-RevId:
183292230
Eugene Brevdo [Thu, 25 Jan 2018 22:28:41 +0000 (14:28 -0800)]
Automated g4 rollback of changelist
183273334
PiperOrigin-RevId:
183291956
Alexandre Passos [Thu, 25 Jan 2018 22:18:45 +0000 (14:18 -0800)]
Deprecation warning on Variables's += methods
PiperOrigin-RevId:
183290246
A. Unique TensorFlower [Thu, 25 Jan 2018 21:46:54 +0000 (13:46 -0800)]
Remove no longer used param fields from TfLiteResizeBilinerParams
PiperOrigin-RevId:
183284937
Yao Zhang [Thu, 25 Jan 2018 21:21:24 +0000 (13:21 -0800)]
VLOG shape inference and annotation return status.
PiperOrigin-RevId:
183280222
A. Unique TensorFlower [Thu, 25 Jan 2018 20:24:58 +0000 (12:24 -0800)]
Automated g4 rollback of changelist
183251689
PiperOrigin-RevId:
183273334
Yifei Feng [Thu, 25 Jan 2018 20:02:36 +0000 (12:02 -0800)]
Add C0330 bad-continuation check to pylint.
PiperOrigin-RevId:
183270896
Sanjoy Das [Thu, 25 Jan 2018 19:17:08 +0000 (11:17 -0800)]
[XLA] Make xla_hlo_profile_test less flaky
Instead of relying on some oeprations always taking longer than others (and this
appearing in a specific order in the rendered HLO profile), pick them out by
opcode.
PiperOrigin-RevId:
183268593
A. Unique TensorFlower [Thu, 25 Jan 2018 18:36:25 +0000 (10:36 -0800)]
Simplify the template mechanism by specifying templates using multi-line strings instead of functions. This loses the syntax verification on templates, but it avoids the clutter of lint overrides and the duplication of parameter names, so things are more readable.
Addresses #16318
PiperOrigin-RevId:
183260854
Akshay Modi [Thu, 25 Jan 2018 17:53:10 +0000 (09:53 -0800)]
Surface error if input is not flat list of eager tensors
PiperOrigin-RevId:
183253112
Eugene Brevdo [Thu, 25 Jan 2018 17:44:04 +0000 (09:44 -0800)]
[TF RNN] Ensure dynamic_rnn runs at least one step to propagate dynamic shape info.
PiperOrigin-RevId:
183251689
Nupur Garg [Thu, 25 Jan 2018 17:34:11 +0000 (09:34 -0800)]
Add functions to encapsulate the logic for checking and setting tensor type.
PiperOrigin-RevId:
183250334
Yunxing Dai [Thu, 25 Jan 2018 17:32:22 +0000 (09:32 -0800)]
Add shard count to select and scatter tests
RELNOTES: n/a
PiperOrigin-RevId:
183250076
A. Unique TensorFlower [Thu, 25 Jan 2018 16:20:51 +0000 (08:20 -0800)]
Automated g4 rollback of changelist
183239252
PiperOrigin-RevId:
183241034
A. Unique TensorFlower [Thu, 25 Jan 2018 16:05:30 +0000 (08:05 -0800)]
Drop the manually_create field from RnnState.
Initially, I thought that the shape of RNN state arrays could always be
determined by shape propagation. Then I came across some graphs where this
wasn't so easy to infer, so I introduced manually_create thinking of it
as a hack. Today I took another look at dropping that hack, and had a
"D'oh" moment when I realized that the cyclic nature of RNN graphs makes
it impossible to infer the shapes of all arrays by usual propagation.
For example, in a LSTM cell, the input array is concatenated with
a state array, so if we don't already know the shape of that state array,
shape propagation stops there.
Thus, this change removes manually_create by making toco always behave as
if manually_create=true, i.e. early-creating all RNN state arrays with
the shape explicitly specified by the user. The next TODO item here
(see model_flags.proto) is to introduce a generic 'shape' field, so far
the current 'size' field only allows specifying 1-D shapes.
PiperOrigin-RevId:
183239252
A. Unique TensorFlower [Thu, 25 Jan 2018 12:48:59 +0000 (04:48 -0800)]
Fixes FunctionLibraryRuntime to be destroyed before ExecutorImpl when NewLocalExecutor returns an error status.
PiperOrigin-RevId:
183220585
A. Unique TensorFlower [Thu, 25 Jan 2018 08:05:16 +0000 (00:05 -0800)]
Add WarmStartSettings configuration for all Estimators.
PiperOrigin-RevId:
183197945
Francois Chollet [Thu, 25 Jan 2018 07:36:20 +0000 (23:36 -0800)]
Fix eager Pooling1D unit test for data_format='channels_first'
PiperOrigin-RevId:
183196050
A. Unique TensorFlower [Thu, 25 Jan 2018 06:33:14 +0000 (22:33 -0800)]
Automated g4 rollback of changelist
183171572
PiperOrigin-RevId:
183192221
Chris Leary [Thu, 25 Jan 2018 04:51:33 +0000 (20:51 -0800)]
[XLA] Add more tests for ConvertElementType.
PiperOrigin-RevId:
183185601
A. Unique TensorFlower [Thu, 25 Jan 2018 04:13:14 +0000 (20:13 -0800)]
Make the graph generation for TFBT deterministic.
PiperOrigin-RevId:
183183086
A. Unique TensorFlower [Thu, 25 Jan 2018 03:47:58 +0000 (19:47 -0800)]
[XLA] Fix HloModule clone.
Currently the cloning of an instruction is usually "shallow": the called
computations of the instruction are reused in the clone. This mechanism is
useful when the hlo graph need to be modified in place (e.g. inliner, and some
hlo passes). One exception is the fusion instruction: it's always "deep"
copied, which means the fused computation is copied as well. So when we deep
cloning an HLO module, don't re-copy the fused computation, and do let the
instruction's clone function know where to put the copied fused computation.
PiperOrigin-RevId:
183181206
Rui Zhao [Thu, 25 Jan 2018 03:32:10 +0000 (19:32 -0800)]
Allow passing candidate sampling seed in to sampled_softmax_loss for testing.
PiperOrigin-RevId:
183180196
Justin Lebar [Thu, 25 Jan 2018 03:01:16 +0000 (19:01 -0800)]
[XLA] Allow buffers for CustomCalls to be reused.
PiperOrigin-RevId:
183177944
A. Unique TensorFlower [Thu, 25 Jan 2018 02:23:18 +0000 (18:23 -0800)]
Refactoring pass.
Add a container context class to shorten argument lists and expose context information in a more organized manner.
Clean up names, docs and tests.
Rename submodule to avoid clashing with the new @covert decorator.
PiperOrigin-RevId:
183174971
Skye Wanderman-Milne [Thu, 25 Jan 2018 01:56:57 +0000 (17:56 -0800)]
Don't close session in debug_gradients_test.py.
PiperOrigin-RevId:
183171960
Ian Langmore [Thu, 25 Jan 2018 01:56:55 +0000 (17:56 -0800)]
VISIBILITY_FIX: Add 'auto_correlation' to _allowed_symbols.
PiperOrigin-RevId:
183171955
Skye Wanderman-Milne [Thu, 25 Jan 2018 01:52:19 +0000 (17:52 -0800)]
Make batch_sequences_with_states_test.py work with C API enabled.
PiperOrigin-RevId:
183171572
Skye Wanderman-Milne [Thu, 25 Jan 2018 01:49:28 +0000 (17:49 -0800)]
Make resource_variable_ops_test.py work with the C API enabled.
PiperOrigin-RevId:
183171341
Yunxing Dai [Thu, 25 Jan 2018 01:24:50 +0000 (17:24 -0800)]
Add R3 and R5 tests to select and scatter.
- Added evaluator support so that we can test S&S in arbitrary dimensions.
RELNOTES: n/a
PiperOrigin-RevId:
183168473
A. Unique TensorFlower [Thu, 25 Jan 2018 01:18:47 +0000 (17:18 -0800)]
Add copy method to EagerVariableStore.
EagerVariableStore.copy creates a new EagerVariableStore instance containing new variables so that they can be modified without affecting the variables in the old store.
PiperOrigin-RevId:
183167776
A. Unique TensorFlower [Thu, 25 Jan 2018 01:16:58 +0000 (17:16 -0800)]
Fix comparison bug in remove_trivial_binary
PiperOrigin-RevId:
183167508
A. Unique TensorFlower [Thu, 25 Jan 2018 01:11:42 +0000 (17:11 -0800)]
Dedup control dependencies where the node already depends on the source as a non-control input.
PiperOrigin-RevId:
183166819
Russell Power [Thu, 25 Jan 2018 00:41:50 +0000 (16:41 -0800)]
Defer logging infeed error messages for a short time to see if the main session returns.
PiperOrigin-RevId:
183162866
A. Unique TensorFlower [Thu, 25 Jan 2018 00:28:20 +0000 (16:28 -0800)]
Prepare internal kernels for BroadcastMul and BroadcastAdd
PiperOrigin-RevId:
183160907
A. Unique TensorFlower [Thu, 25 Jan 2018 00:23:09 +0000 (16:23 -0800)]
Basic AddN support in toco
PiperOrigin-RevId:
183160197
Sanjoy Das [Thu, 25 Jan 2018 00:10:37 +0000 (16:10 -0800)]
[TF:XLA:CPU] Fix an issue with explicitly vectorized reductions
The explicitly vectorized reduction kernel assumes that the input and the output
layouts are "same" (not identical which is not possible, but have the same
ordering of the un-reduced dimensions). This change makes IrEmitter check this
precondition.
PiperOrigin-RevId:
183158249
Sanjoy Das [Thu, 25 Jan 2018 00:10:12 +0000 (16:10 -0800)]
[TF:XLA] Replace most of HloProfilePrinter by a protocol buffer
This change replaces the meat of HloProfilePrinter with a protobuf
HloProfilePrinterData. The original plan was to serialize HloProfilePrinter
into C++ source code and put that in a .cc file along with the string for the
xla::ProgramShape. However, since we now directly serialize xla::ProgramShape
into a .o file, for consistency I think we should do the same thing for
HloProfilePrinter (instead of adding yet another output file to tfcompile).
The change itself is fairly simple, it is large mostly due to the mass renaming
I had to do.
PiperOrigin-RevId:
183158192
Yutaka Leon [Wed, 24 Jan 2018 23:45:55 +0000 (15:45 -0800)]
Set export_outs in KMeans' EstimatorSpec.
PiperOrigin-RevId:
183154542
Skye Wanderman-Milne [Wed, 24 Jan 2018 23:23:12 +0000 (15:23 -0800)]
Make tensor_util_test.py work with the C API enabled.
Some of the shape inference error messages have changed. We also need
to wrap an InvalidArgumentError as a ValueError for backwards compat.
PiperOrigin-RevId:
183150857
A. Unique TensorFlower [Wed, 24 Jan 2018 23:10:52 +0000 (15:10 -0800)]
internal change
PiperOrigin-RevId:
183148922
Eugene Brevdo [Wed, 24 Jan 2018 23:10:28 +0000 (15:10 -0800)]
Modify dynamic_rnn to use max_seq_length for its while_loop.
This is especially important with XLA compilation, which will usually have
a maximum time dimension on the input that is (sometimes much larger than)
max(sequence_length). In these cases, a lot of compute can be saved.
PiperOrigin-RevId:
183148831
Amit Patankar [Wed, 24 Jan 2018 23:10:19 +0000 (15:10 -0800)]
Parameterized docker build now supports a local pip whl file path.
PiperOrigin-RevId:
183148798
A. Unique TensorFlower [Wed, 24 Jan 2018 23:07:21 +0000 (15:07 -0800)]
When a device list API such as TF_DeviceListName() succeeds, set the output status to OK.
PiperOrigin-RevId:
183148319
A. Unique TensorFlower [Wed, 24 Jan 2018 22:56:16 +0000 (14:56 -0800)]
Go: Update generated wrapper functions for TensorFlow ops.
PiperOrigin-RevId:
183146357
A. Unique TensorFlower [Wed, 24 Jan 2018 22:41:35 +0000 (14:41 -0800)]
[XLA] Enable F16 for the CPU and GPU backends.
Enhance the IR emitter to handle mathematics operations with F16 types.
Add tests for F16 mathematics operations.
PiperOrigin-RevId:
183144087
Sanjoy Das [Wed, 24 Jan 2018 22:23:02 +0000 (14:23 -0800)]
Remove THIRD_PARTY_ from #include guards
They don't make sense in the open source repository.
PiperOrigin-RevId:
183140889
A. Unique TensorFlower [Wed, 24 Jan 2018 22:18:30 +0000 (14:18 -0800)]
Update ops-related pbtxt files.
PiperOrigin-RevId:
183140179
A. Unique TensorFlower [Wed, 24 Jan 2018 22:17:07 +0000 (14:17 -0800)]
Expose shared_embedding_columns in tf.feature_column.
PiperOrigin-RevId:
183139975
A. Unique TensorFlower [Wed, 24 Jan 2018 22:17:05 +0000 (14:17 -0800)]
Adds loss_reduction argument to canned estimators.
PiperOrigin-RevId:
183139970
Derek Murray [Wed, 24 Jan 2018 22:03:09 +0000 (14:03 -0800)]
[Eager mode] Add an Eager-optimized version of the IteratorGetNext op.
Since in eager mode the kernel execution happens on the calling thread
(which would be blocked anyway), we can invoke the iterator
synchronously, and do not need to perform a context switch to and from
a background thread.
Improves the latency benchmarks in
//third_party/tensorflow/contrib/eager/python:datasets_test by
approximately 7us to 13us per element.
PiperOrigin-RevId:
183137488
Alexandre Passos [Wed, 24 Jan 2018 22:02:14 +0000 (14:02 -0800)]
Silent copies for int32 tensors in TFE.
Fixes #16106
PiperOrigin-RevId:
183137298
Sanjoy Das [Wed, 24 Jan 2018 21:39:02 +0000 (13:39 -0800)]
[TF:XLA] Bump open source llvm revision to r323309
PiperOrigin-RevId:
183133556
Nupur Garg [Wed, 24 Jan 2018 21:21:54 +0000 (13:21 -0800)]
Make TFLite Pad op have parity with TF Pad op.
PiperOrigin-RevId:
183131070
A. Unique TensorFlower [Wed, 24 Jan 2018 21:09:10 +0000 (13:09 -0800)]
Portability fix for parse_testdata.cc
PiperOrigin-RevId:
183129228
Alexandre Passos [Wed, 24 Jan 2018 20:47:00 +0000 (12:47 -0800)]
tf_contextlib forgotten oops
PiperOrigin-RevId:
183126126
Alexandre Passos [Wed, 24 Jan 2018 20:46:09 +0000 (12:46 -0800)]
Initializes the context before setting the device policy
PiperOrigin-RevId:
183126037
Justin Lebar [Wed, 24 Jan 2018 20:12:02 +0000 (12:12 -0800)]
[XLA] Fix crash in HloOrdering::ToString().
HloOrdering::ToString() was crashing when given a computation containing
a fusion instruction.
PiperOrigin-RevId:
183121645
Anna R [Wed, 24 Jan 2018 19:44:55 +0000 (11:44 -0800)]
Internal change.
PiperOrigin-RevId:
183117665
Yunxing Dai [Wed, 24 Jan 2018 19:42:22 +0000 (11:42 -0800)]
[BatchNorm] Remove CPU implementation
We now use batchnorm rewriter (tensorflow/compiler/xla/service/batchnorm_rewriter.h) to expand batch norm into smaller ops. A specific implementation should not be needed anymore (for CPU).
RELNOTES:n/a
PiperOrigin-RevId:
183117252
Jianwei Xie [Wed, 24 Jan 2018 19:31:06 +0000 (11:31 -0800)]
Fixed linter errors.
PiperOrigin-RevId:
183115307
A. Unique TensorFlower [Wed, 24 Jan 2018 19:29:08 +0000 (11:29 -0800)]
fix typos in schema.fbs comments.
PiperOrigin-RevId:
183114994
Suharsh Sivakumar [Wed, 24 Jan 2018 18:57:28 +0000 (10:57 -0800)]
Introduce experimental quantize training and eval apis.
These APIs will include options that may have undefined behavior, still being researched. The non experimental APIs will have better hardware support and will be more suitable for users aiming to rewrite graphs to provide to backends (e.g. TFLite).
PiperOrigin-RevId:
183109634
Anna R [Wed, 24 Jan 2018 18:55:55 +0000 (10:55 -0800)]
Fix update_api_def.sh script to output ApiDefs instead of ApiDef proto.
PiperOrigin-RevId:
183109303
A. Unique TensorFlower [Wed, 24 Jan 2018 18:35:22 +0000 (10:35 -0800)]
Internal Change
PiperOrigin-RevId:
183105591
Akshay Agrawal [Wed, 24 Jan 2018 18:34:07 +0000 (10:34 -0800)]
TFE: Add graph benchmark for linear regression, for comparison with our eager execution benchmark.
PiperOrigin-RevId:
183105403
Jianwei Xie [Wed, 24 Jan 2018 18:02:35 +0000 (10:02 -0800)]
Merge changes from github.
PiperOrigin-RevId:
183100142