A. Unique TensorFlower [Tue, 13 Mar 2018 18:36:28 +0000 (11:36 -0700)]
Change back TFE_Execute logic to set '*num_retvals' on return.
PiperOrigin-RevId:
188903892
A. Unique TensorFlower [Tue, 13 Mar 2018 18:34:23 +0000 (11:34 -0700)]
Add TransformShardedTileShape helper method to HloSharding
It transforms an existing sharding to be compatible with a new shape
with an optional transform method to adjust the tile size for the
sharded dimensions.
PiperOrigin-RevId:
188903257
A. Unique TensorFlower [Tue, 13 Mar 2018 18:30:20 +0000 (11:30 -0700)]
Automated g4 rollback of changelist
188353354
PiperOrigin-RevId:
188902517
Yuefeng Zhou [Tue, 13 Mar 2018 18:27:46 +0000 (11:27 -0700)]
Replace the unknown dimension of signature input when building grappler items.
Fix the bug where same feed nodes or fetch nodes would be added more than once.
PiperOrigin-RevId:
188902101
A. Unique TensorFlower [Tue, 13 Mar 2018 17:39:33 +0000 (10:39 -0700)]
Enable arithmetic optimizations for Fill nodes that are all zeros or ones.
PiperOrigin-RevId:
188893722
A. Unique TensorFlower [Tue, 13 Mar 2018 17:22:53 +0000 (10:22 -0700)]
Add TFLite Interpreter bindings.
PiperOrigin-RevId:
188890943
A. Unique TensorFlower [Tue, 13 Mar 2018 17:13:45 +0000 (10:13 -0700)]
Performing the finalization of the LayerCollection outside of FisherEstimator's constructor. This allows layers and losses to be registered after the FisherEstimator (or KFACOptimizer) has been constructed.
PiperOrigin-RevId:
188889252
Mark Daoust [Tue, 13 Mar 2018 17:09:03 +0000 (10:09 -0700)]
include the graph.pb->tensorboard script in pip package
This will allow you to run it with:
`python -m tensorflow.python.tools.import_pb_to_tensorboard ...`
Without any additional downloads or setup.
see #8854
PiperOrigin-RevId:
188888501
A. Unique TensorFlower [Tue, 13 Mar 2018 16:11:14 +0000 (09:11 -0700)]
Internal change
PiperOrigin-RevId:
188880419
Eugene Brevdo [Tue, 13 Mar 2018 16:09:57 +0000 (09:09 -0700)]
[TFLite] Don't always treat inputs and outputs as encoded floats.
PiperOrigin-RevId:
188880237
A. Unique TensorFlower [Tue, 13 Mar 2018 15:49:17 +0000 (08:49 -0700)]
Introduce output fusion for (A dot B) * alpha
We allow now to fuse (A dot B) times a constant, which is available in cublas.
PiperOrigin-RevId:
188878210
A. Unique TensorFlower [Tue, 13 Mar 2018 12:32:55 +0000 (05:32 -0700)]
PiperOrigin-RevId:
188860028
Shanqing Cai [Tue, 13 Mar 2018 05:38:48 +0000 (22:38 -0700)]
tfdbg: add reference to the TensorBoard Debugger Plugin in README.md
PiperOrigin-RevId:
188828898
A. Unique TensorFlower [Tue, 13 Mar 2018 12:32:55 +0000 (05:32 -0700)]
PiperOrigin-RevId:
188860028
Shanqing Cai [Tue, 13 Mar 2018 05:38:48 +0000 (22:38 -0700)]
tfdbg: add reference to the TensorBoard Debugger Plugin in README.md
PiperOrigin-RevId:
188828898
A. Unique TensorFlower [Tue, 13 Mar 2018 12:32:55 +0000 (05:32 -0700)]
PiperOrigin-RevId:
188860028
Shanqing Cai [Tue, 13 Mar 2018 05:38:48 +0000 (22:38 -0700)]
tfdbg: add reference to the TensorBoard Debugger Plugin in README.md
PiperOrigin-RevId:
188828898
A. Unique TensorFlower [Tue, 13 Mar 2018 04:53:54 +0000 (21:53 -0700)]
Merges (embedding|indicator)_column with their sequence counterparts, and exposes sequence methods under tf.contrib.feature_column.
PiperOrigin-RevId:
188826187
A. Unique TensorFlower [Tue, 13 Mar 2018 04:18:25 +0000 (21:18 -0700)]
Update ops-related pbtxt files.
PiperOrigin-RevId:
188824362
A. Unique TensorFlower [Tue, 13 Mar 2018 04:12:52 +0000 (21:12 -0700)]
Support NHWC <--> HWNC permutation.
PiperOrigin-RevId:
188824058
A. Unique TensorFlower [Tue, 13 Mar 2018 02:46:08 +0000 (19:46 -0700)]
Go: Update generated wrapper functions for TensorFlow ops.
PiperOrigin-RevId:
188817976
Shanqing Cai [Tue, 13 Mar 2018 02:33:52 +0000 (19:33 -0700)]
Merge changes from github.
PiperOrigin-RevId:
188817194
A. Unique TensorFlower [Tue, 13 Mar 2018 02:09:28 +0000 (19:09 -0700)]
Remove integration_tests directory and associated files.
PiperOrigin-RevId:
188815493
Max Galkin [Tue, 13 Mar 2018 01:35:15 +0000 (18:35 -0700)]
Demystify MaterializeShapes a bit.
PiperOrigin-RevId:
188812445
A. Unique TensorFlower [Tue, 13 Mar 2018 00:48:57 +0000 (17:48 -0700)]
Make tf.add_to_collections visible.
PiperOrigin-RevId:
188807786
A. Unique TensorFlower [Tue, 13 Mar 2018 00:13:09 +0000 (17:13 -0700)]
[XLA:Tools] Make hlo_runner understand --xla_hlo_profile.
PiperOrigin-RevId:
188803724
Skye Wanderman-Milne [Tue, 13 Mar 2018 00:12:32 +0000 (17:12 -0700)]
Don't evaluate control flow in EvaluateConstantTensor.
ExtractConstantSubgraph doesn't copy control edges, which are sometimes necessary
to correctly evaluate conds (at the very least). Avoid evaluating conds at all to
address this.
PiperOrigin-RevId:
188803649
A. Unique TensorFlower [Mon, 12 Mar 2018 23:54:10 +0000 (16:54 -0700)]
BREAKING_CHANGE: Split out event_ndims=0 bijectors from Affine and CholeskyOuterProduct.
- Deprecate event_ndims argument
- Create a Square bijector for the scalar case of CholeskyOuterProduct (which now only operates on matrices).
- Create a AffineScalar bijector for the scalar case of Affine (which now only operates on vectors)
PiperOrigin-RevId:
188801116
Derek Murray [Mon, 12 Mar 2018 23:46:48 +0000 (16:46 -0700)]
Switch BuildGraphOptions to wrap CallableOptions.
This change harmonizes the graph construction codepaths for DirectSession and
MasterSession, which will make it easier to add new subgraph creation features.
PiperOrigin-RevId:
188799932
Benoit Steiner [Mon, 12 Mar 2018 23:25:46 +0000 (16:25 -0700)]
Don't use shapes unless it's safe to do so.
PiperOrigin-RevId:
188796626
Derek Murray [Mon, 12 Mar 2018 22:40:47 +0000 (15:40 -0700)]
[tf.data] Enable Dataset.make_one_shot_iterator() and Dataset.__iter__() in eager mode.
This change partially replicates the code in `tf.contrib.eager.Iterator`.
However, since that class depends on contrib-level functionality (viz.
cross-device prefetching support), we cannot move it wholesale to core.
PiperOrigin-RevId:
188790349
Yu-Cheng Ling [Mon, 12 Mar 2018 22:05:11 +0000 (15:05 -0700)]
TFLite Delegate: Expose input / output tensor indicies in `Init`
PiperOrigin-RevId:
188784614
Max Galkin [Mon, 12 Mar 2018 21:52:24 +0000 (14:52 -0700)]
Fix typo in the doc: tf.multiplytiply
https://www.tensorflow.org/api_docs/python/tf/multiply
PiperOrigin-RevId:
188782466
A. Unique TensorFlower [Mon, 12 Mar 2018 21:23:49 +0000 (14:23 -0700)]
K-FAC: FisherBlocks for tf.nn.{depthwise_conv2d, separable_conv2d, convolution}.
PiperOrigin-RevId:
188778072
A. Unique TensorFlower [Mon, 12 Mar 2018 20:50:35 +0000 (13:50 -0700)]
Move "hoist common factor out of aggregation" optimization
to a separate stage.
1) Use a new naming scheme for optimized ops,
share it with AddOpsRewrite
2) Make sure that tests actually test that optimized
nodes exists in a graph
PiperOrigin-RevId:
188772892
Sanjoy Das [Mon, 12 Mar 2018 20:44:58 +0000 (13:44 -0700)]
[TF:XLA] Bump open source llvm revision to r327201
PiperOrigin-RevId:
188771994
Eugene Brevdo [Mon, 12 Mar 2018 20:35:03 +0000 (13:35 -0700)]
[TFLite] Don't require a std::vector for Interpreter::SetTensorParameters*.
PiperOrigin-RevId:
188770522
Igor Saprykin [Mon, 12 Mar 2018 20:07:12 +0000 (13:07 -0700)]
Move `loss_reduction` argument from `replicate_model_fn` to `TowerOptimizer.
PiperOrigin-RevId:
188766477
A. Unique TensorFlower [Mon, 12 Mar 2018 20:05:26 +0000 (13:05 -0700)]
Add is_discrete, is_continuous, is_bounded methods to TensorSpecs.
PiperOrigin-RevId:
188766232
Allen Lavoie [Mon, 12 Mar 2018 20:00:24 +0000 (13:00 -0700)]
Fix another eager PyObject leak
Shockingly this one was also due to PySequence_GetItem.
PiperOrigin-RevId:
188765548
Joshua V. Dillon [Mon, 12 Mar 2018 19:58:49 +0000 (12:58 -0700)]
Add custom_gradient function.
PiperOrigin-RevId:
188765271
A. Unique TensorFlower [Mon, 12 Mar 2018 19:44:29 +0000 (12:44 -0700)]
Extend TF Eager C API to allow asynchronous execution.
PiperOrigin-RevId:
188763442
Yunxing Dai [Mon, 12 Mar 2018 18:49:36 +0000 (11:49 -0700)]
[XLA] [Copy insertion] Deterministically iterate through instructions to copy
- Use HloInstructionMap to get deterministic iteration order.
PiperOrigin-RevId:
188755375
Skye Wanderman-Milne [Mon, 12 Mar 2018 18:43:57 +0000 (11:43 -0700)]
Switch op_hint.py to use _set_attr.
This is in preparation for enabling the C API. Modifying an op's NodeDef
directly has no effect with the C API enabled.
PiperOrigin-RevId:
188754464
Mark Daoust [Mon, 12 Mar 2018 18:39:08 +0000 (11:39 -0700)]
Standardize "op" capitalization, see "adding_an_op".
PiperOrigin-RevId:
188753529
Joshua V. Dillon [Mon, 12 Mar 2018 18:29:24 +0000 (11:29 -0700)]
Improve usability of `tf.contrib.bayesflow.custom_gradient` by removing need for `axis` arg and support taking lists.
PiperOrigin-RevId:
188751894
A. Unique TensorFlower [Mon, 12 Mar 2018 18:24:22 +0000 (11:24 -0700)]
Transposes are can be merged into reshapes when the ordering of non-one
dimensions remains unchanged.
PiperOrigin-RevId:
188751074
A. Unique TensorFlower [Mon, 12 Mar 2018 18:19:08 +0000 (11:19 -0700)]
boosted_trees: infer the output shapes of Quantiles Op from the input shapes.
PiperOrigin-RevId:
188750079
Benoit Steiner [Mon, 12 Mar 2018 18:04:59 +0000 (11:04 -0700)]
Avoid capturing unused variables in lambda functions
PiperOrigin-RevId:
188747641
Skye Wanderman-Milne [Mon, 12 Mar 2018 18:02:29 +0000 (11:02 -0700)]
Fix race in C API.
RecordMutation could race with ExtendSessionGraphHelper, which would
release the graph lock and only keep the session lock when extending
the session.
Also makes sure thread annotations are on declarations, not definitions
(otherwise they have no effect).
PiperOrigin-RevId:
188747158
A. Unique TensorFlower [Mon, 12 Mar 2018 17:37:20 +0000 (10:37 -0700)]
Turn trivial Pack ops with a single input into ExpandDims ops to avoid copying the tensor.
PiperOrigin-RevId:
188742516
A. Unique TensorFlower [Mon, 12 Mar 2018 17:34:34 +0000 (10:34 -0700)]
Make default number of threads trigger the default behavior for both eigen
and gemmlowp. In gemmlowp the default is '1', while in eigen it is 'number of
processors'.
PiperOrigin-RevId:
188742087
A. Unique TensorFlower [Mon, 12 Mar 2018 17:24:56 +0000 (10:24 -0700)]
Convert Squeeze into Reshape: Support empty output shapes.
PiperOrigin-RevId:
188740288
A. Unique TensorFlower [Mon, 12 Mar 2018 17:12:22 +0000 (10:12 -0700)]
Supporting quantization of Gather ops and removal of trivial Relu1s when quantized.
PiperOrigin-RevId:
188738133
Allen Lavoie [Mon, 12 Mar 2018 16:28:18 +0000 (09:28 -0700)]
Plug a few more PyObject leaks, test for them.
PiperOrigin-RevId:
188731961
Benoit Steiner [Mon, 12 Mar 2018 16:16:08 +0000 (09:16 -0700)]
Don't let the grappler item builder fail if the graph contains unknown custom
ops.
PiperOrigin-RevId:
188730560
A. Unique TensorFlower [Mon, 12 Mar 2018 16:12:41 +0000 (09:12 -0700)]
Don't remove identity nodes if they follow a device crossing and have consumers on a device different than themselves. They may be used to cache or route data between devices in a deliberate manner.
Simplify code in DependencyOptimizer a bit.
PiperOrigin-RevId:
188730185
Asim Shankar [Mon, 12 Mar 2018 16:05:33 +0000 (09:05 -0700)]
ResourceScatterUpdate: Gracefully handle inconsistent indices and updates in
the kernel.
With graph execution, consistency between the shapes of the arguments to
ResourceScatterUpdate is validated by the shape inference functions at graph
construction time.
With eager execution, the shape inference logic isn't executed, so inconsistent
arguments could be provided to the kernel, which would result in a segmentation
fault prior to this change. As demonstrated by the added tests.
PiperOrigin-RevId:
188729154
Benoit Steiner [Mon, 12 Mar 2018 15:03:54 +0000 (08:03 -0700)]
Turn on function optimization by default
PiperOrigin-RevId:
188722505
A. Unique TensorFlower [Mon, 12 Mar 2018 14:26:13 +0000 (07:26 -0700)]
Reuse the linear index when broadcasting a contiguous range of dimensions.
This potentially allows us to get rid of additional mod and div operations.
PiperOrigin-RevId:
188719238
A. Unique TensorFlower [Mon, 12 Mar 2018 10:13:15 +0000 (03:13 -0700)]
Lint some files.
PiperOrigin-RevId:
188698275
Olivia Nordquist [Mon, 12 Mar 2018 07:30:38 +0000 (00:30 -0700)]
this test is also timing out in cuda so disabling for now
PiperOrigin-RevId:
188685611
Mark Daoust [Mon, 12 Mar 2018 04:23:15 +0000 (21:23 -0700)]
propagate fix from tensorflow/models#3561
PiperOrigin-RevId:
188675327
Shashi Shekhar [Mon, 12 Mar 2018 03:20:45 +0000 (20:20 -0700)]
Fix assets for the TF camera example.
Mobile net model is downloaded from tf_http_archive("tf_mobilenet") rule and renaming the asset file in assets folder has no effect.
PiperOrigin-RevId:
188672531
Olivia Nordquist [Mon, 12 Mar 2018 02:35:34 +0000 (19:35 -0700)]
disable flaky asan test
PiperOrigin-RevId:
188670616
Derek Murray [Mon, 12 Mar 2018 00:49:34 +0000 (17:49 -0700)]
Fix typo in description of INTERNAL error code.
PiperOrigin-RevId:
188666142
A. Unique TensorFlower [Sun, 11 Mar 2018 23:22:47 +0000 (16:22 -0700)]
Removed duplicate statement.
PiperOrigin-RevId:
188663018
Derek Murray [Sun, 11 Mar 2018 22:38:16 +0000 (15:38 -0700)]
Fixes a race condition in function instantiation.
Previously, if the same function was being concurrently instantiated
and released:
1. Thread one could begin to instantiate the function, determine
that it already existed in the runtime, then be preempted.
2. Thread two could release the handle on the function, causing it to
be freed and removed from the `FunctionLibraryRuntime::items_` map.
3. Thread one could then incorrectly assume that the function still
existed, and fail to find it in the `FunctionLibraryRuntime::items_`
map, causing a segfault when it attempted to increment the refcount
on an uninitialized object.
PiperOrigin-RevId:
188661500
A. Unique TensorFlower [Sun, 11 Mar 2018 17:44:02 +0000 (10:44 -0700)]
Specify the `maximum_iterations` to tf.while_loop in tf.scan to be compatible with XLA.
PiperOrigin-RevId:
188652533
Bixia Zheng [Sun, 11 Mar 2018 17:15:18 +0000 (10:15 -0700)]
Selectively re-enable bfloat16 tests for the GPU backend.
PiperOrigin-RevId:
188651655
A. Unique TensorFlower [Sun, 11 Mar 2018 17:00:02 +0000 (10:00 -0700)]
Improve errors raised when an object does not match the RNNCell interface.
PiperOrigin-RevId:
188651070
Gunhan Gulsoy [Sun, 11 Mar 2018 06:18:23 +0000 (22:18 -0800)]
Fix windows GPU build scripts.
PiperOrigin-RevId:
188629017
Jianwei Xie [Sun, 11 Mar 2018 03:44:13 +0000 (19:44 -0800)]
Adds a warning to help user to debug the TPU program hanging for predict() call.
PiperOrigin-RevId:
188624174
A. Unique TensorFlower [Sat, 10 Mar 2018 20:03:19 +0000 (12:03 -0800)]
Turn the following ops into Identity.
* Slice when the Size input matches the size of the input tensor
* Tile when the multiples input is a tensor of '1'
* Pad/PadV2 when the paddings input is a tensor of 0
* Squeeze when the squeeze dimensions are known to be > 1
PiperOrigin-RevId:
188609800
Rui Zhao [Sat, 10 Mar 2018 08:29:37 +0000 (00:29 -0800)]
Increment node_ids when merging CostGraphDef.
PiperOrigin-RevId:
188586552
Michael Kuperstein [Sat, 10 Mar 2018 06:49:30 +0000 (22:49 -0800)]
[XLA] Speed up colocated buffer merging.
PiperOrigin-RevId:
188581202
A. Unique TensorFlower [Sat, 10 Mar 2018 06:43:51 +0000 (22:43 -0800)]
Fix DepthToSpace and SpaceToDepth to silently return instead of failing when the input tensor is empty.
PiperOrigin-RevId:
188580972
A. Unique TensorFlower [Sat, 10 Mar 2018 04:50:32 +0000 (20:50 -0800)]
Fix docstring for `embedding_lookup_sparse`.
Example with weighted mean combiner implies that single-key embeddings not normalized (the weighted sum answer). However, the code and test shows normalization regardless of number of keys.
PiperOrigin-RevId:
188575982
Mingsheng Hong [Sat, 10 Mar 2018 03:07:44 +0000 (19:07 -0800)]
Unified test util PlaceHolderFloat() into PlaceHolder(), and extended the latter
to take a TF_DataType param.
PiperOrigin-RevId:
188570493
A. Unique TensorFlower [Sat, 10 Mar 2018 02:54:41 +0000 (18:54 -0800)]
Propagate min/max for StridedSlice
PiperOrigin-RevId:
188569611
A. Unique TensorFlower [Sat, 10 Mar 2018 02:50:06 +0000 (18:50 -0800)]
Move optimizations to arithmetic optimizer stages
1) Redundant Bitcast
2) Redundant Cast
3) Remove inverse transpose
PiperOrigin-RevId:
188569367
Derek Murray [Sat, 10 Mar 2018 02:12:02 +0000 (18:12 -0800)]
Add experimental Session::MakeCallable() API and implement it for DirectSession.
The intent of this new API matches the Python `tf.Session.make_callable()`
method: it splits the two roles of the `Session::Run()` method into separate
methods:
1. `Session::MakeCallable()` takes information about a subgraph (such as the
names of nodes to feed and fetch), and prunes and optimizes that graph,
returning a simple handle.
2. `Session::RunCallable()` takes that handle, plus any values to be fed,
and executes the graph, returning whatever outputs are produced.
This split moves string processing off the critical path of running a
step. We also add a new method `Session::ReleaseCallable()` that makes
it possible to free the resources associated with a cached subgraph,
and could be useful for seldom-executed graphs such as initializers.
PiperOrigin-RevId:
188566635
Derek Murray [Sat, 10 Mar 2018 02:00:26 +0000 (18:00 -0800)]
Fix flakiness in common_runtime/function_test.cc.
The flakiness was due to nondeterministic names being chosen for
folded constants; the fix was to split out the source of the
nondetermism into a separate test.
PiperOrigin-RevId:
188565362
A. Unique TensorFlower [Sat, 10 Mar 2018 00:34:42 +0000 (16:34 -0800)]
Allowing ReorderActivationFunctions to reorder output arrays and adding support for propagating activation functions into Gather ops and as constant ops.
PiperOrigin-RevId:
188556574
A. Unique TensorFlower [Sat, 10 Mar 2018 00:17:08 +0000 (16:17 -0800)]
[XLA] Add a whole graph execution interface.
PiperOrigin-RevId:
188554206
Brennan Saeta [Fri, 9 Mar 2018 23:39:11 +0000 (15:39 -0800)]
ProcFLR: Include the remote function target in the function_key
PiperOrigin-RevId:
188548941
Akshay Modi [Fri, 9 Mar 2018 23:34:44 +0000 (15:34 -0800)]
Move a utility function into the eager C internal API
PiperOrigin-RevId:
188548393
Igor Saprykin [Fri, 9 Mar 2018 23:28:15 +0000 (15:28 -0800)]
Copy `replicate_model_fn` to core.
PiperOrigin-RevId:
188547527
Joshua V. Dillon [Fri, 9 Mar 2018 23:27:50 +0000 (15:27 -0800)]
Migrate tf.contrib.bayesflow.optimizers to tfp.optimziers.
PiperOrigin-RevId:
188547477
Allen Lavoie [Fri, 9 Mar 2018 22:42:51 +0000 (14:42 -0800)]
Eager: Fix a Dimension PyObject leak, test for it.
PiperOrigin-RevId:
188540944
Michael Case [Fri, 9 Mar 2018 22:41:06 +0000 (14:41 -0800)]
Internal Change.
PiperOrigin-RevId:
188540659
Francois Chollet [Fri, 9 Mar 2018 22:40:18 +0000 (14:40 -0800)]
Part of the update of tf.keras to the Keras 2.1.5 API.
PiperOrigin-RevId:
188540513
A. Unique TensorFlower [Fri, 9 Mar 2018 22:17:03 +0000 (14:17 -0800)]
Automated g4 rollback of changelist
188501394
PiperOrigin-RevId:
188536863
Sanjoy Das [Fri, 9 Mar 2018 22:15:23 +0000 (14:15 -0800)]
[TF:XLA] Bump open source llvm revision to r326989
PiperOrigin-RevId:
188536576
Skye Wanderman-Milne [Fri, 9 Mar 2018 22:00:23 +0000 (14:00 -0800)]
Add smart_case that calls smart_cond.
PiperOrigin-RevId:
188534066
A. Unique TensorFlower [Fri, 9 Mar 2018 21:54:17 +0000 (13:54 -0800)]
Remove the nondeterminism from a test for initializing variables from checkpoints.
PiperOrigin-RevId:
188533156
Kay Zhu [Fri, 9 Mar 2018 21:47:36 +0000 (13:47 -0800)]
[XLA:Doc] Fix an typo "Alternately" -> "Alternatively".
PiperOrigin-RevId:
188532135
A. Unique TensorFlower [Fri, 9 Mar 2018 21:26:41 +0000 (13:26 -0800)]
LSTM support: Add non-uint8 quantized operators.
PiperOrigin-RevId:
188529107
A. Unique TensorFlower [Fri, 9 Mar 2018 21:24:23 +0000 (13:24 -0800)]
Internal Change
PiperOrigin-RevId:
188528771
A. Unique TensorFlower [Fri, 9 Mar 2018 21:00:12 +0000 (13:00 -0800)]
Automated g4 rollback of changelist
188492233
PiperOrigin-RevId:
188525453