Joshua V. Dillon [Tue, 6 Feb 2018 01:55:26 +0000 (17:55 -0800)]
Cleanup markdown errors in `Bijector`.
PiperOrigin-RevId:
184616392
Justin Lebar [Tue, 6 Feb 2018 01:45:26 +0000 (17:45 -0800)]
[XLA] Add HloBindings::ToString().
PiperOrigin-RevId:
184615306
Igor Saprykin [Tue, 6 Feb 2018 01:24:31 +0000 (17:24 -0800)]
Correctly treat "devices=/gpu:0" argument of replicate_model_fn.
At the moment if "devices=/GPU:0" are specified by the user, then variables are going to be placed on the GPU. However, if "devices=/gpu:0" are given, then they are going to be placed on the CPU. Instead, the latter case should be equivalent to the former case.
PiperOrigin-RevId:
184612823
Mingsheng Hong [Tue, 6 Feb 2018 01:16:27 +0000 (17:16 -0800)]
Misc cleanups and tweaks:
1. Removed obsolete constructors in ProcessLunctionLibraryRuntime
2. Add const annotations to Tensor::PrintOneDim(), and removed unnecessary vector copy.
PiperOrigin-RevId:
184611531
Justin Lebar [Tue, 6 Feb 2018 01:11:39 +0000 (17:11 -0800)]
[XLA:GPU] Split IrEmitter{Unn,N}ested out of ir_emitter.h.
Also add a bunch of clarifying comments.
PiperOrigin-RevId:
184610674
A. Unique TensorFlower [Tue, 6 Feb 2018 00:32:38 +0000 (16:32 -0800)]
Support negative axis in concatenation
PiperOrigin-RevId:
184605786
A. Unique TensorFlower [Tue, 6 Feb 2018 00:12:19 +0000 (16:12 -0800)]
Shard linear operator tests.
PiperOrigin-RevId:
184602704
Amit Patankar [Tue, 6 Feb 2018 00:08:16 +0000 (16:08 -0800)]
Add filepaths to test_local support.
PiperOrigin-RevId:
184602010
Yifei Feng [Tue, 6 Feb 2018 00:04:06 +0000 (16:04 -0800)]
Remove makefile build dependency on all_opensource_files, as part of the effort to remove all_opensource_files #15758.
PiperOrigin-RevId:
184601439
Skye Wanderman-Milne [Mon, 5 Feb 2018 23:49:57 +0000 (15:49 -0800)]
"frame_name" attr must be altered when importing/exporting MetaGraphDefs.
The frame_name attr of Enter operations must be the name of the
associated WhileLoopContext, otherwise taking the gradient of the loop
will result in frame errors.
I'm not happy that the export logic is in meta_graph.py (all other
control flow de/serialization is in control_flow_ops.py). However, I
can't think of how else to do it, since only export_scoped_meta_graph
has access to the NodeDefs being exported.
PiperOrigin-RevId:
184599323
A. Unique TensorFlower [Mon, 5 Feb 2018 23:46:50 +0000 (15:46 -0800)]
[XLA] Sink layout sensitivity from CSE into HloInstruction::Identical, and make it the default.
PiperOrigin-RevId:
184598903
Suharsh Sivakumar [Mon, 5 Feb 2018 23:32:32 +0000 (15:32 -0800)]
Make fold batch norm code use OneofPattern and rearrange functions to (maybe) be more readable.
PiperOrigin-RevId:
184597111
A. Unique TensorFlower [Mon, 5 Feb 2018 23:30:54 +0000 (15:30 -0800)]
Assign total_loss in order not to crash if training loop exists early.
PiperOrigin-RevId:
184596877
Sergio Guadarrama [Mon, 5 Feb 2018 23:16:29 +0000 (15:16 -0800)]
Adding TensorSpec to represent the specification of Tensors.
PiperOrigin-RevId:
184594856
A. Unique TensorFlower [Mon, 5 Feb 2018 23:14:20 +0000 (15:14 -0800)]
Verify tflite model in TFLite Java API
PiperOrigin-RevId:
184594561
Max Galkin [Mon, 5 Feb 2018 23:10:44 +0000 (15:10 -0800)]
Add logging to diagnose device properties parsing problem in Grappler.
PiperOrigin-RevId:
184594084
Yu-Cheng Ling [Mon, 5 Feb 2018 23:05:43 +0000 (15:05 -0800)]
Fix CBLAS Conv reference implementation in TFLite.
PiperOrigin-RevId:
184592951
Alexandre Passos [Mon, 5 Feb 2018 22:47:23 +0000 (14:47 -0800)]
Automated g4 rollback of changelist
184573795
PiperOrigin-RevId:
184590080
A. Unique TensorFlower [Mon, 5 Feb 2018 22:43:35 +0000 (14:43 -0800)]
Backward pass implementation for fusion optimizer.
PiperOrigin-RevId:
184589487
A. Unique TensorFlower [Mon, 5 Feb 2018 22:38:07 +0000 (14:38 -0800)]
[XLA] Fix documentation for Clamp.
PiperOrigin-RevId:
184588630
Peter Hawkins [Mon, 5 Feb 2018 22:12:02 +0000 (14:12 -0800)]
[TF:XLA] Implement GatherNd.
PiperOrigin-RevId:
184584104
Derek Murray [Mon, 5 Feb 2018 21:43:52 +0000 (13:43 -0800)]
[tf.data] Fix use-after-free bug when closing down an input pipeline.
This fix affects the distributed runtime; DirectSession use is unaffected.
Before this change, an iterator that used a background prefetching
thread might attempt to use a captured FunctionLibraryRuntime from a
subgraph that had been deregistered (and hence its
FunctionLibraryRuntime would have been deleted). This change
introduces a mechanism for "cloning" the necessary parts of the
FunctionLibraryRuntime so that it can be owned by the
IteratorResource.
PiperOrigin-RevId:
184579490
A. Unique TensorFlower [Mon, 5 Feb 2018 21:43:08 +0000 (13:43 -0800)]
[XLA] Add tests for Clamp of S32 and U32 vectors with broadcasted scalars.
PiperOrigin-RevId:
184579375
Alexandre Passos [Mon, 5 Feb 2018 21:03:53 +0000 (13:03 -0800)]
Fixes issue where external control dependencies in while loops are dropped.
Fixes #15891
PiperOrigin-RevId:
184573795
Asim Shankar [Mon, 5 Feb 2018 20:40:51 +0000 (12:40 -0800)]
contrib/rnn: Fix #16703
(Bug introduced in
https://github.com/tensorflow/tensorflow/commit/
3f579020bab8f00e4621e9c7c740cbf13136a809)
Kudos to @akhti for pointing this out.
PiperOrigin-RevId:
184570448
A. Unique TensorFlower [Mon, 5 Feb 2018 20:16:04 +0000 (12:16 -0800)]
Bug fix: Don't dereference nullptr in OpKernelContext::input_alloc_attr().
PiperOrigin-RevId:
184566770
Yunxing Dai [Mon, 5 Feb 2018 20:15:10 +0000 (12:15 -0800)]
[TF:XLA] Making constant folding deterministic.
Making constant folding deterministic by doing DFS deterministically and inserting a serialization point based on nodes' names.
This is the last source of non-determinism remaining in the TF:XLA stack.
RELNOTES: Constant folding pass is now deterministic.
PiperOrigin-RevId:
184566644
Benoit Steiner [Mon, 5 Feb 2018 20:08:36 +0000 (12:08 -0800)]
Serialize the evaluation of the AssignAdd nodes to make the result more
deterministic
Improved testing
PiperOrigin-RevId:
184565483
Yao Zhang [Mon, 5 Feb 2018 19:23:40 +0000 (11:23 -0800)]
Support parsing from text and fused op in contrib.
PiperOrigin-RevId:
184558131
Joshua V. Dillon [Mon, 5 Feb 2018 18:45:33 +0000 (10:45 -0800)]
Automated g4 rollback of changelist
184323369
PiperOrigin-RevId:
184551259
A. Unique TensorFlower [Mon, 5 Feb 2018 18:39:18 +0000 (10:39 -0800)]
Proper reallocation of dynamic tensors.
PiperOrigin-RevId:
184550199
Dan Ringwalt [Mon, 5 Feb 2018 18:36:24 +0000 (10:36 -0800)]
Make flat_transforms_to_matrices and matrices_to_flat_transforms public (#781).
PiperOrigin-RevId:
184549704
Amit Patankar [Mon, 5 Feb 2018 18:16:02 +0000 (10:16 -0800)]
Changing the link to point to new android job.
PiperOrigin-RevId:
184546160
A. Unique TensorFlower [Mon, 5 Feb 2018 18:06:10 +0000 (10:06 -0800)]
[XLA] add Conditional to the local Python XLA client.
PiperOrigin-RevId:
184544483
Dan Ringwalt [Mon, 5 Feb 2018 17:49:01 +0000 (09:49 -0800)]
Clarify that tf.contrib.image.rotate angles are counterclockwise.
PiperOrigin-RevId:
184541776
A. Unique TensorFlower [Mon, 5 Feb 2018 17:33:42 +0000 (09:33 -0800)]
Enable aggressive identity node pruning in dependency optimizer.
PiperOrigin-RevId:
184539756
Skye Wanderman-Milne [Mon, 5 Feb 2018 16:41:54 +0000 (08:41 -0800)]
Get control_flow_ops.py ready to support de/serializing nested control flow.
With this change, ControlFlowContexts keep track of their nested
contexts (the reverse lookup as ControlFlowContext.outer_context).
This is to enable de/serializing the nested contexts of each "root"
context, and only adding the root contexts to collections. This allows
for simple deserialization of each root context by recursively
deserializing its nested contexts.
The de/serialization logic is disabled and the corresponding
control_flow.proto changes are omitted for now for forwards
compatability (i.e. three-week-old binaries must be ready to accept
the new proto format once its commited). After this is committed for
three weeks, I'll commit a follow-up change enabling the new behavior.
Design note: I chose to serialize the nested contexts, rather than the
outer contexts, because it makes it easy to deserialize the contexts
in topological order and to assign the right outer context. If we
serialized the outer contexts, there'd need to be some mechanism for
either sorting all the serialized contexts first, or deserializing all
of them and then doing another pass to assign the outer contexts.
PiperOrigin-RevId:
184533406
A. Unique TensorFlower [Mon, 5 Feb 2018 16:32:18 +0000 (08:32 -0800)]
Internal Change
PiperOrigin-RevId:
184532417
A. Unique TensorFlower [Mon, 5 Feb 2018 15:56:27 +0000 (07:56 -0800)]
Expand the activity analysis to composite names.
Fix a bug in the cond template that caused bad syntax when there it no symbol that needs aliasing.
More refactoring in the process, including:
* introduce the QN (qualified name) class to hold symbol information; it has value semantics and can generate the original symbol, a corresponding AST tree or a single-symbol form (e.g. "a.b" -> a_b)
* allow the template mechanism to use QNs for substitutions
* annotate *all* symbol nodes with their corresponding QN object; this is done as first step during static analysis, and automatically performed on all template expansions
* start using typed annotation keys (Enum values) instead of plain strings
* rename access.py to activity.py
* sanitize nodes in template expansion by deep copying the AST without annotations, to avoid common references
PiperOrigin-RevId:
184528586
A. Unique TensorFlower [Mon, 5 Feb 2018 09:42:16 +0000 (01:42 -0800)]
mini documentation fix
PiperOrigin-RevId:
184496843
A. Unique TensorFlower [Mon, 5 Feb 2018 09:03:32 +0000 (01:03 -0800)]
Fixed sequence_mask behavior on unknown shape.
`sequence_mask` crashed when fed with a tensor of unknown rank.
Added a test for that, and expanded a bit existing tests.
Also fixed pre-existing lint errors.
PiperOrigin-RevId:
184493239
A. Unique TensorFlower [Mon, 5 Feb 2018 03:58:18 +0000 (19:58 -0800)]
Support for quantized LSTM models.
PiperOrigin-RevId:
184476753
Mingsheng Hong [Sun, 4 Feb 2018 19:54:35 +0000 (11:54 -0800)]
Minor fixes to the get started doc.
PiperOrigin-RevId:
184457074
Derek Murray [Sun, 4 Feb 2018 19:24:51 +0000 (11:24 -0800)]
Avoid retaining two copies of each constant in `ConstantOp`.
Presently, the kernel keeps two copies of the constant tensor value, which can be large:
1. In the `ConstantOp::tensor_` field.
2. In the `OpKernel::def_` field (as an attr of the `NodeDef`).
Since we can be sure that `ConstantOp` will never need to access the
tensor value from `OpKernel::def_`, this change introduces a mechanism
for `OpKernel` implementations to store a stripped `NodeDef` in the
base class, and uses it in `ConstantOp` to avoid storing the tensor
value attr.
PiperOrigin-RevId:
184455793
Blake Hechtman [Sun, 4 Feb 2018 07:45:00 +0000 (23:45 -0800)]
[XLA] Assign mandatory constraints in a DFS order and non-manatory constraints in a BFS order.
PiperOrigin-RevId:
184429818
RJ Ryan [Sat, 3 Feb 2018 17:46:31 +0000 (09:46 -0800)]
[tf-signal] Fix exception when input shape is unknown in mfccs_from_log_mel_spectrograms.
PiperOrigin-RevId:
184400783
A. Unique TensorFlower [Sat, 3 Feb 2018 07:02:16 +0000 (23:02 -0800)]
[XLA] Add tests for Clamp with scalars S32 and U32.
PiperOrigin-RevId:
184376425
Yu-Cheng Ling [Sat, 3 Feb 2018 06:39:45 +0000 (22:39 -0800)]
Fix breakage: Can't build TFLite with Bazel for Mac / iOS.
PiperOrigin-RevId:
184375534
A. Unique TensorFlower [Sat, 3 Feb 2018 03:40:30 +0000 (19:40 -0800)]
A more efficient implementation of the Op using batch operations.
PiperOrigin-RevId:
184367562
Mingsheng Hong [Sat, 3 Feb 2018 03:10:39 +0000 (19:10 -0800)]
Extended TFE_OpSetDevice() with the ability to set an op device from non-GPU back to GPU.
Added unit testing, and also refined unit test logic for checking the presence
of a GPU device. The latter is needed when we add XLA device support.
PiperOrigin-RevId:
184366172
Justin Lebar [Sat, 3 Feb 2018 02:57:10 +0000 (18:57 -0800)]
[XLA] Minor cleanups related to multi-output fusion.
- Add some comments about preexisting invariants, and add some CHECKs.
- In the LoopEmitter constructor, materialize the given
ArraySlice<IrArray> to a vector, so we don't rely on the given
ArraySlice having any particular lifetime.
- Add the invariant that the LoopEmitter constructor which takes a
list of IrArrays is only for multi-output fusion. Previously it said:
If you only pass one array, then treat it as regular fusion. But this
results in an LLVM type mismatch, because the given
target_element_generator should be passing a struct with one element.
PiperOrigin-RevId:
184365310
A. Unique TensorFlower [Sat, 3 Feb 2018 02:03:11 +0000 (18:03 -0800)]
Propagate outfeed sharding, if specified from TensorFlow.
PiperOrigin-RevId:
184361221
Anna R [Sat, 3 Feb 2018 01:59:16 +0000 (17:59 -0800)]
Automated g4 rollback of changelist
184347012
PiperOrigin-RevId:
184360818
Derek Murray [Sat, 3 Feb 2018 01:09:59 +0000 (17:09 -0800)]
[tf.data] Add public header "tensorflow/core/framework/dataset.h".
This adds the ability to create a custom C++ Dataset implementation
without linking it statically into the TensorFlow library. Note that
this internal API is experimental and subject to change between
versions of TensorFlow.
Fixes #16682.
PiperOrigin-RevId:
184356318
Derek Murray [Sat, 3 Feb 2018 01:09:16 +0000 (17:09 -0800)]
Check that the type of an implicitly dereferenced tensor matches the expected input type.
The dtype of a tensor reference can change between the point when it is "produced" by an
operation and consumed by the next operation. This evades checks in the executor that the
type of tensor on each edge matches the type signatures of the producing and consuming operation, which could lead to undefined behavior. Although there is no existing operation that changes the type of a tensor reference, it is possible to use the OpKernelContext API to do so, so we add a further check in the runtime to defend against operations that might be added in the future.
PiperOrigin-RevId:
184356242
Jonathan Hseu [Sat, 3 Feb 2018 00:35:10 +0000 (16:35 -0800)]
Update global_step by default if the user specifies a host_call.
PiperOrigin-RevId:
184352399
A. Unique TensorFlower [Sat, 3 Feb 2018 00:18:59 +0000 (16:18 -0800)]
Tpu cluster resolver only returns TF server addresses for 'HEALTHY' tpu nodes.
PiperOrigin-RevId:
184350480
Alexandre Passos [Sat, 3 Feb 2018 00:02:22 +0000 (16:02 -0800)]
Replacing _container_prefix with _graph_key to preserve its use in optimizer and deprecate others.
PiperOrigin-RevId:
184348303
Alexandre Passos [Fri, 2 Feb 2018 23:52:18 +0000 (15:52 -0800)]
TF_CALL_ALL_TYPES should include variant
PiperOrigin-RevId:
184347081
Alexandre Passos [Fri, 2 Feb 2018 23:51:41 +0000 (15:51 -0800)]
Rollback of recent changes to tensor.{h,cc}
PiperOrigin-RevId:
184347012
Jianwei Xie [Fri, 2 Feb 2018 23:00:17 +0000 (15:00 -0800)]
Adds batch inference support on TPU with TPUEstimator.predict.
PiperOrigin-RevId:
184339842
Eugene Brevdo [Fri, 2 Feb 2018 22:30:50 +0000 (14:30 -0800)]
[TF RNN] Small optimization to rnn: only calculate copy_cond once.
PiperOrigin-RevId:
184335231
Tatiana Shpeisman [Fri, 2 Feb 2018 22:14:13 +0000 (14:14 -0800)]
Changed minimal required version of bleach Python package from 1.5 to 2.0. bleach 1.5 causes problems with Jupyter as reported in #16424
PiperOrigin-RevId:
184332663
Mingsheng Hong [Fri, 2 Feb 2018 21:53:08 +0000 (13:53 -0800)]
A few misc tweaks to TFE APIs.
PiperOrigin-RevId:
184329345
Yu-Cheng Ling [Fri, 2 Feb 2018 21:49:35 +0000 (13:49 -0800)]
TFLite: Conv CBLAS kernel
PiperOrigin-RevId:
184328848
Derek Murray [Fri, 2 Feb 2018 21:39:38 +0000 (13:39 -0800)]
[tf.data] Move framework/dataset.h to framework/dataset_stateful_op_whitelist.h.
This will make way for the move of kernels/data/dataset.h to
framework/dataset.h, while preserving version control history, after
which we might recombine the headers.
PiperOrigin-RevId:
184327481
Joshua V. Dillon [Fri, 2 Feb 2018 21:08:31 +0000 (13:08 -0800)]
Support `float16` `dtype` in `tf.linalg.*`.
Note: not all `LinearOperator` functions will support `float16`. This change
merely enables constructing the `LinearOperator` object(s) using this `dtype`.
PiperOrigin-RevId:
184323477
Joshua V. Dillon [Fri, 2 Feb 2018 21:07:37 +0000 (13:07 -0800)]
Breaking change: Revise HMC interface to accept a list of Tensors representing a partitioning of chain state.
PiperOrigin-RevId:
184323369
Yuefeng Zhou [Fri, 2 Feb 2018 20:46:34 +0000 (12:46 -0800)]
Skip the node that has unexpected number of outputs.
PiperOrigin-RevId:
184320865
Sanjoy Das [Fri, 2 Feb 2018 20:23:47 +0000 (12:23 -0800)]
[TF:XLA] Bump open source llvm revision to r324073
PiperOrigin-RevId:
184318036
A. Unique TensorFlower [Fri, 2 Feb 2018 20:19:43 +0000 (12:19 -0800)]
Update ops-related pbtxt files.
PiperOrigin-RevId:
184317521
Peter Hawkins [Fri, 2 Feb 2018 19:31:06 +0000 (11:31 -0800)]
[TF:XLA] Improve/refactor the handling of resource types/shapes.
Previously we used an xla::Shape to track the shape of a resource (Variable, TensorArray, Stack) shape. The xla::Shape described how the resource was represented to XLA, e.g., as a (buffer, size) pair for a Stack resource.
Instead, separate the TensorFlow abstract shape representation from the XLA shape representation and track it separately. This leads to simpler and more readable code.
PiperOrigin-RevId:
184310694
A. Unique TensorFlower [Fri, 2 Feb 2018 19:24:17 +0000 (11:24 -0800)]
Allow ResizeBilinear to resize the output tensor in Prepare(), if the size tensor is const.
PiperOrigin-RevId:
184309687
Alexandre Passos [Fri, 2 Feb 2018 19:24:11 +0000 (11:24 -0800)]
Register resource_scatter_update for string types.
PiperOrigin-RevId:
184309674
A. Unique TensorFlower [Fri, 2 Feb 2018 19:06:09 +0000 (11:06 -0800)]
Add shape inference for outside_compilation graph rewrite. Pull out enough of the graph to enable inference of the shape of a SendFromHost Op once the shape of corresponding RecvAtHost Ops are known.
END_PUBLIC
Fixed open source build breaks.
BEGIN_PUBLIC
Automated g4 rollback of changelist
184169668
PiperOrigin-RevId:
184306845
A. Unique TensorFlower [Fri, 2 Feb 2018 18:58:12 +0000 (10:58 -0800)]
Register GPU host kernels for Identity and RefIdentity.
PiperOrigin-RevId:
184305574
A. Unique TensorFlower [Fri, 2 Feb 2018 18:46:34 +0000 (10:46 -0800)]
Automated g4 rollback of changelist
184273245
PiperOrigin-RevId:
184303789
Anna R [Fri, 2 Feb 2018 18:28:20 +0000 (10:28 -0800)]
Remove hidden_ops.txt file. Instead, switch to use visibility attribute in ApiDef proto.
PiperOrigin-RevId:
184301076
Jacques Pienaar [Fri, 2 Feb 2018 18:04:15 +0000 (10:04 -0800)]
Consider beyond immediate neighbors to find exit node.
Most of the exit nodes are immediate neighbors of the switch, except we do have
cases where the switch feeds into an identity that feeds into a exit.
PiperOrigin-RevId:
184297180
A. Unique TensorFlower [Fri, 2 Feb 2018 18:01:40 +0000 (10:01 -0800)]
Show that we must over reallocate after resizing dynamic tensors.
PiperOrigin-RevId:
184296680
A. Unique TensorFlower [Fri, 2 Feb 2018 17:20:26 +0000 (09:20 -0800)]
Fix latent bug in dependency optimizer.
PiperOrigin-RevId:
184291701
Yao Zhang [Fri, 2 Feb 2018 17:00:41 +0000 (09:00 -0800)]
Disable graph optimizations (CSE) in test so that constant nodes are not deduped.
PiperOrigin-RevId:
184289685
A. Unique TensorFlower [Fri, 2 Feb 2018 16:09:48 +0000 (08:09 -0800)]
[comment-only change]: Fix grammar error.
PiperOrigin-RevId:
184285125
Derek Murray [Fri, 2 Feb 2018 16:09:13 +0000 (08:09 -0800)]
Fix a bug in function inlining when the argument is an implicitly dereferenced ref tensor.
Previously the inliner would add an identity node with an invalid ref-type attr when the actual parameter had ref type. The changed version removes the reference.
PiperOrigin-RevId:
184285084
A. Unique TensorFlower [Fri, 2 Feb 2018 13:38:50 +0000 (05:38 -0800)]
Enabling partitioned variables to work with TPU.
When partitioned variables are used in a TPU training loop,
concat gradient operations get generated for which XLA requires
the concat dimension argument to be a constant (or foldable to a constant).
However since such constant is defined outside of the train while context
an Enter node is generated in order to pass it.
The fix consists in detecting such case, and to duplicate the (scalar) constant
inside the while context, so that XLA can succesfully process the resulting
graph.
PiperOrigin-RevId:
184273245
Gunhan Gulsoy [Fri, 2 Feb 2018 07:47:33 +0000 (23:47 -0800)]
Fix some tf-lite tests
PiperOrigin-RevId:
184247187
A. Unique TensorFlower [Fri, 2 Feb 2018 05:43:12 +0000 (21:43 -0800)]
Fix tolerance too tight for Wasserstein distance test.
PiperOrigin-RevId:
184240222
Gunhan Gulsoy [Fri, 2 Feb 2018 07:47:33 +0000 (23:47 -0800)]
Fix some tf-lite tests
PiperOrigin-RevId:
184247187
A. Unique TensorFlower [Fri, 2 Feb 2018 05:43:12 +0000 (21:43 -0800)]
Fix tolerance too tight for Wasserstein distance test.
PiperOrigin-RevId:
184240222
Justin Lebar [Fri, 2 Feb 2018 05:34:18 +0000 (21:34 -0800)]
Internal change
PiperOrigin-RevId:
184239740
A. Unique TensorFlower [Fri, 2 Feb 2018 04:35:29 +0000 (20:35 -0800)]
Automated g4 rollback of changelist
183874527
PiperOrigin-RevId:
184236409
A. Unique TensorFlower [Fri, 2 Feb 2018 03:50:34 +0000 (19:50 -0800)]
Supporting new saving op structure.
PiperOrigin-RevId:
184233513
Yu-Cheng Ling [Fri, 2 Feb 2018 02:27:31 +0000 (18:27 -0800)]
Add darwin_x86_64 config in TF Lite BUILD file.
PiperOrigin-RevId:
184227786
Benoit Steiner [Fri, 2 Feb 2018 02:05:15 +0000 (18:05 -0800)]
Fixed the description of the fake GPU device to avoid a division by 0
PiperOrigin-RevId:
184225409
Brennan Saeta [Fri, 2 Feb 2018 01:20:11 +0000 (17:20 -0800)]
Fix tests
PiperOrigin-RevId:
184220615
Benoit Steiner [Fri, 2 Feb 2018 01:19:32 +0000 (17:19 -0800)]
Skip unknown devices since we can't optimize for them
PiperOrigin-RevId:
184220515
Andrew Selle [Fri, 2 Feb 2018 00:51:59 +0000 (16:51 -0800)]
Allow reordering of execution order of nodes with indirect execution_plan.
Now whenever we want to operate in dependency order we use execution_plan.
It begins as identity map (0, ..., nodes_size()) but can be changed
in the future. This is the basis for more pluggable delegation.
PiperOrigin-RevId:
184216885
Jacques Pienaar [Fri, 2 Feb 2018 00:29:30 +0000 (16:29 -0800)]
Automated g4 rollback of changelist
184188816
PiperOrigin-RevId:
184213576
Brennan Saeta [Fri, 2 Feb 2018 00:22:24 +0000 (16:22 -0800)]
GCS Throttle: 1 token == 1 Kb
Previously, 1 token was approximately 256 bytes. This is slightly less
intuitive than 1 kb.
PiperOrigin-RevId:
184212503
Raghuraman Krishnamoorthi [Fri, 2 Feb 2018 00:14:51 +0000 (16:14 -0800)]
Add functionality to fold batch norm (supporting both fused and unfused batch norm) to support quantized training. The weights are always now scaled by gamma/sigma, where sigma is the moving standard deviation for stability prior to quantization. For improved performance, the moving means and variances are frozen and the training graph modified accordingly.
An additional parameter freeze_batch_norm_delay is added to foldbatchnorm function to set the delay at which training switches from regular batch norm to frozen mean and variances.
Remove placement options within FoldBatchNorm as this causes folded training to place all ops on a single GPU. The modification now significantly speeds up distributed training.
The tests for folding batch norms are also updated to reflect the additional topological changes to the graph.
PiperOrigin-RevId:
184211434