Amit Patankar [Fri, 1 Jun 2018 22:27:16 +0000 (15:27 -0700)]
Merge branch 'r1.9' into relnotes_19b
Amit Patankar [Fri, 1 Jun 2018 22:22:54 +0000 (15:22 -0700)]
Merge pull request #19703 from av8ramit/update_version_190_rc0
Updating version for 1.9.0-rc0.
Amit Patankar [Fri, 1 Jun 2018 21:26:07 +0000 (14:26 -0700)]
Updating version for 1.9.0-rc0.
Amit Patankar [Fri, 1 Jun 2018 21:19:03 +0000 (14:19 -0700)]
Updating release notes for r1.9.
Michael Case [Fri, 1 Jun 2018 19:58:16 +0000 (12:58 -0700)]
Merge pull request #19680 from case540/branch_198811639
Branch
198811639
Clayne Robison [Fri, 1 Jun 2018 19:29:39 +0000 (12:29 -0700)]
Finished incomplete support for bad usernames in the CI build scripts. ci_build.sh now passes the environment variable to the container, and the with_the_same_user script adds the --force-badname param to addgroup as well. (#19699)
Michael Case [Fri, 1 Jun 2018 05:55:46 +0000 (22:55 -0700)]
Fix import depth issue.
Michael Case [Fri, 1 Jun 2018 04:55:11 +0000 (21:55 -0700)]
Potential fix to layout_optimizer_test.py
Jason Zaman [Fri, 1 Jun 2018 03:52:17 +0000 (11:52 +0800)]
BUILD: dont force stripping (#19599)
* BUILD: dont force stripping
Build systems must not strip binaries, it makes it impossible for
distros to ship debugging symbols for packages.
bazel build has a --strip option to allow the user to generate stripped
binaries in a configurable way, that should be used instead.
https://fedoraproject.org/wiki/Packaging:Debuginfo
https://wiki.gentoo.org/wiki/Project:Quality_Assurance/Backtraces#Stripping
Signed-off-by: Jason Zaman <jason@perfinion.com>
* configure: add --strip=always to bazelrc
Michael Case [Fri, 1 Jun 2018 03:44:41 +0000 (20:44 -0700)]
Add new line to make buildifier happy
Michael Case [Fri, 1 Jun 2018 03:27:17 +0000 (20:27 -0700)]
Merge branch 'master' into branch_198811639
Michael Case [Fri, 1 Jun 2018 02:46:48 +0000 (19:46 -0700)]
Fix sanity issues.
Felix Abecassis [Fri, 1 Jun 2018 02:45:15 +0000 (04:45 +0200)]
docker: update cuDNN to 7.1.4.18 (#19636)
Signed-off-by: Felix Abecassis <fabecassis@nvidia.com>
Pete Warden [Fri, 1 Jun 2018 02:44:05 +0000 (19:44 -0700)]
Manual roll back of PR #19443, because it causes the Raspberry Pi build to fail (#19678)
Michael Case [Fri, 1 Jun 2018 01:31:23 +0000 (18:31 -0700)]
Fix lite.py Python TypeError.
Michael Case [Fri, 1 Jun 2018 01:29:32 +0000 (18:29 -0700)]
Merge commit for internal changes
Allen Lavoie [Fri, 1 Jun 2018 01:17:48 +0000 (18:17 -0700)]
Make the TFOptimizer wrapper checkpointable.
TensorFlow Optimizers compiled with a Model will now have their state saved and restored with save_weights/load_weights.
PiperOrigin-RevId:
198811639
Peter Hawkins [Fri, 1 Jun 2018 01:09:50 +0000 (18:09 -0700)]
Automated g4 rollback of changelist
198803131
PiperOrigin-RevId:
198810875
Saurabh Saxena [Fri, 1 Jun 2018 00:20:31 +0000 (17:20 -0700)]
Throw a more informative error message when checkpointing an input pipeline
containing a ShuffleDataset with reshuffle_each_iteration=True.
This is a temporary fix till we figure out how to handle this use-case.
PiperOrigin-RevId:
198805344
Justin Lebar [Fri, 1 Jun 2018 00:19:25 +0000 (17:19 -0700)]
[XLA] Fix handling of CustomCall's window and dnums.
CustomCall can have a window and convolution-dimension-numbers, so
HloInstruction needs to handle this in Clone() and Identical().
PiperOrigin-RevId:
198805211
Derek Murray [Fri, 1 Jun 2018 00:18:54 +0000 (17:18 -0700)]
[tf.data] Update `DatasetBase::DebugString()` to be const in the docs.
PiperOrigin-RevId:
198805143
Peter Hawkins [Fri, 1 Jun 2018 00:17:13 +0000 (17:17 -0700)]
Mark tensorflow/contrib/learn:estimator_test as optonly because it is flaky due to timeouts without optimization.
PiperOrigin-RevId:
198804880
Peter Hawkins [Fri, 1 Jun 2018 00:03:07 +0000 (17:03 -0700)]
[TF:XLA] Preliminary support for tpu.replicate() inside of TF control flow (such as tf.while_loop()).
Register the remaining control-flow operators on XLA devices.
PiperOrigin-RevId:
198803131
Blake Hechtman [Thu, 31 May 2018 23:58:05 +0000 (16:58 -0700)]
[XLA] Fix batchnorm rewriter to not use implicit broadcasts. Algebraic
simplifier reshape change is now covered by ReshapeMover.
PiperOrigin-RevId:
198802494
Bixia Zheng [Thu, 31 May 2018 23:25:24 +0000 (16:25 -0700)]
Handle FilterLayout::kOutputYXInput in FilterDescriptor::ToShortString.
This fixes an error when running resnet50_batch128_fp16 with --v=2.
PiperOrigin-RevId:
198798196
A. Unique TensorFlower [Thu, 31 May 2018 23:21:24 +0000 (16:21 -0700)]
Edited the landing page for the Performance section. Reorganized content and removed references to content that is being deleted.
PiperOrigin-RevId:
198797662
Jiri Simsa [Thu, 31 May 2018 23:18:15 +0000 (16:18 -0700)]
[tf.data] Reflect `MakeIterator` signature change in documentation.
PiperOrigin-RevId:
198797254
Brennan Saeta [Thu, 31 May 2018 23:15:45 +0000 (16:15 -0700)]
[tf.data] Mark DebugString() as const.
By marking DebugString() as const we can make some error messages more descriptive. Because DatasetIterator marks the return value of the dataset() function const, DebugString() cannot be called.
PiperOrigin-RevId:
198796894
Nupur Garg [Thu, 31 May 2018 23:07:02 +0000 (16:07 -0700)]
Fixed Python API.
PiperOrigin-RevId:
198795738
Saurabh Saxena [Thu, 31 May 2018 23:06:15 +0000 (16:06 -0700)]
Add utility for converting FunctionDef to GraphDef and _FuncGraph.
PiperOrigin-RevId:
198795625
Justin Lebar [Thu, 31 May 2018 23:02:26 +0000 (16:02 -0700)]
[XLA:GPU] Specify cudnn conv algorithm via backend_config.
Gets rid of the tricky algorithm/use-tensor-cores operands to cudnn
convolution customcalls, using instead a backend_config.
PiperOrigin-RevId:
198794988
Ruoxin Sang [Thu, 31 May 2018 23:01:50 +0000 (16:01 -0700)]
Fix one comment in prefetch_autotuner_test.cc.
PiperOrigin-RevId:
198794897
A. Unique TensorFlower [Thu, 31 May 2018 23:01:23 +0000 (16:01 -0700)]
Putting stubs for function shape inference interface
PiperOrigin-RevId:
198794845
Frank Chen [Thu, 31 May 2018 22:52:15 +0000 (15:52 -0700)]
Add warning for gcs_config_ops
PiperOrigin-RevId:
198793502
Yu-Cheng Ling [Thu, 31 May 2018 22:50:55 +0000 (15:50 -0700)]
Refactoring: Extract CombineHashes function into a shared module
PiperOrigin-RevId:
198793295
Justin Lebar [Thu, 31 May 2018 22:25:10 +0000 (15:25 -0700)]
[XLA] Check for identical backend configs in HloInstruction::Identical.
PiperOrigin-RevId:
198789495
A. Unique TensorFlower [Thu, 31 May 2018 22:17:52 +0000 (15:17 -0700)]
Move reorder-cast-and-transpose optimization to optimization stage.
PiperOrigin-RevId:
198788352
A. Unique TensorFlower [Thu, 31 May 2018 22:11:26 +0000 (15:11 -0700)]
Internal change.
PiperOrigin-RevId:
198787391
A. Unique TensorFlower [Thu, 31 May 2018 22:05:23 +0000 (15:05 -0700)]
Similar to cr/
188652533, specify the `maximum_iterations` to tf.while_loop in tf.map_fn to be compatible with XLA.
PiperOrigin-RevId:
198786266
A. Unique TensorFlower [Thu, 31 May 2018 21:48:51 +0000 (14:48 -0700)]
Write checkpoint path of evaluated checkpoint to the event file.
PiperOrigin-RevId:
198783364
A. Unique TensorFlower [Thu, 31 May 2018 21:42:07 +0000 (14:42 -0700)]
Expose the ExponentialMovingAverage name as a public property.
PiperOrigin-RevId:
198782348
Jacques Pienaar [Thu, 31 May 2018 21:24:13 +0000 (14:24 -0700)]
Mark XLAShapeForArgument as const.
PiperOrigin-RevId:
198778945
A. Unique TensorFlower [Thu, 31 May 2018 21:01:45 +0000 (14:01 -0700)]
Move RemodeRedundantReshape optimization to a separate stage.
PiperOrigin-RevId:
198775276
Nupur Garg [Thu, 31 May 2018 20:58:32 +0000 (13:58 -0700)]
Add attributes to TFLite Python API.
PiperOrigin-RevId:
198774775
Jiri Simsa [Thu, 31 May 2018 20:43:43 +0000 (13:43 -0700)]
[tf.data] Changing signature of `MakeIterator` to enable propagating error status.
PiperOrigin-RevId:
198772254
Alexandre Passos [Thu, 31 May 2018 20:22:10 +0000 (13:22 -0700)]
More eager notebooks.
PiperOrigin-RevId:
198768912
Anna R [Thu, 31 May 2018 20:11:43 +0000 (13:11 -0700)]
Moving generated API to tensorflow/.
PiperOrigin-RevId:
198767512
Shashi Shekhar [Thu, 31 May 2018 20:10:07 +0000 (13:10 -0700)]
Add profiling statistics to benchmark.
PiperOrigin-RevId:
198767297
Frank Chen [Thu, 31 May 2018 19:38:35 +0000 (12:38 -0700)]
Fix links in the TensorFlow Security Advisories
PiperOrigin-RevId:
198762795
A. Unique TensorFlower [Thu, 31 May 2018 19:16:54 +0000 (12:16 -0700)]
Making the tf.name_scope blocks related to the factor and weight vars configurable. By default they will not be scoped.
PiperOrigin-RevId:
198759754
A. Unique TensorFlower [Thu, 31 May 2018 18:52:43 +0000 (11:52 -0700)]
RuntimeShapes class: minor tweak to fix builds.
PiperOrigin-RevId:
198755870
Martin Wicke [Thu, 31 May 2018 18:50:35 +0000 (11:50 -0700)]
Merge pull request #19419 from perfinion/pip_package
pip_package: modularize build script to allow distros to install more flexibly
Martin Wicke [Thu, 31 May 2018 18:49:11 +0000 (11:49 -0700)]
Merge pull request #19597 from perfinion/pkgconfig
generate-pc.sh: add option to set libdir
Justin Lebar [Thu, 31 May 2018 18:43:37 +0000 (11:43 -0700)]
[XLA] Make HloInstruction::backend_config() a JSON-encoded protobuf.
PiperOrigin-RevId:
198754463
A. Unique TensorFlower [Thu, 31 May 2018 17:39:33 +0000 (10:39 -0700)]
[XLA] Redesign: delete computation_tracker and user_computation.
PiperOrigin-RevId:
198743117
Dan Moldovan [Thu, 31 May 2018 17:33:53 +0000 (10:33 -0700)]
Initial implementation of a few of the list-specific operators.
This introduces an abstraction for a dispatch context, which allows passing local information through to the specialized operators.
PiperOrigin-RevId:
198742074
Jesse Benson [Thu, 31 May 2018 17:35:15 +0000 (10:35 -0700)]
Build libtensorflow.so and libtensorflow_framework.so for Raspberry Pi. (#18892)
A. Unique TensorFlower [Thu, 31 May 2018 17:20:00 +0000 (10:20 -0700)]
Suppress generation of the proto API's descriptor() method, it conflicts with the field name.
PiperOrigin-RevId:
198739727
A. Unique TensorFlower [Thu, 31 May 2018 17:15:59 +0000 (10:15 -0700)]
Introduce runtime shape class.
PiperOrigin-RevId:
198739017
gracehoney [Thu, 31 May 2018 16:48:54 +0000 (09:48 -0700)]
Merge pull request #19652 from samikama/WiringIssues
Fix TRT Node rewiring
Skye Wanderman-Milne [Thu, 31 May 2018 16:43:30 +0000 (09:43 -0700)]
Another handle_data fix for graph-mode functions.
PiperOrigin-RevId:
198734229
Dan Moldovan [Thu, 31 May 2018 15:53:36 +0000 (08:53 -0700)]
Cleanup: update continue_statements.py to use the base transformer facilities for tracking local state and reindenting node blocks. Rearrange the error handling in base transformer to avoid chained exceptions.
PiperOrigin-RevId:
198727946
A. Unique TensorFlower [Thu, 31 May 2018 15:33:36 +0000 (08:33 -0700)]
Standardize shifts in multiplication util functions.
PiperOrigin-RevId:
198725578
Skye Wanderman-Milne [Thu, 31 May 2018 14:56:56 +0000 (07:56 -0700)]
Make GraphConstructor create nodes in the same order as the GraphDef.
While technically the order of the created nodes doesn't matter, this
makes viewing and debugging graphs more sensible.
Fixes #19594.
PiperOrigin-RevId:
198721173
Jiri Simsa [Thu, 31 May 2018 13:57:55 +0000 (06:57 -0700)]
[tf.data] Scaling down the `batch_dataset_op_test`.
PiperOrigin-RevId:
198715407
A. Unique TensorFlower [Thu, 31 May 2018 13:05:04 +0000 (06:05 -0700)]
implementation of sparse_to_dense
PiperOrigin-RevId:
198710452
A. Unique TensorFlower [Thu, 31 May 2018 05:03:16 +0000 (22:03 -0700)]
[XLA] Redesign: delete the old service interface.
- Computation
- ComputeConstant
- Execute
- ExecuteAsync
- ExecuteParallel
- GetComputationStats
- GetComputationShape
- GetLocalShape
- IsConstant
- LoadComputationSnapshot
- Op
- SetReturnValue
- SnapshotComputation
PiperOrigin-RevId:
198669035
Jingyue Wu [Thu, 31 May 2018 05:00:32 +0000 (22:00 -0700)]
Improve ReshapeIsIdentity to work with symbolic shapes.
For example, with this CL, ArithmeticOptimizer can optimize the Reshape below
into a no-op.
s = Shape(t)
Reshape(t, Concat(s[0], s[1], s[2], s[3]))
PiperOrigin-RevId:
198668726
Nick Felt [Thu, 31 May 2018 02:27:26 +0000 (19:27 -0700)]
Add GCS_READ_CACHE_DISABLED explicit env var to GcsFileSystem
PiperOrigin-RevId:
198658074
Amit Patankar [Thu, 31 May 2018 02:25:44 +0000 (19:25 -0700)]
Merge pull request #19591 from candyzone/master
Fix typo
Amit Patankar [Thu, 31 May 2018 02:25:09 +0000 (19:25 -0700)]
Merge pull request #19655 from gunan/setuptools
Fix setuptools version to avoid a bad release.
Allen Lavoie [Thu, 31 May 2018 02:01:58 +0000 (19:01 -0700)]
Add a subclassed Model's attribute-assigned variables to Model.weights et al
Makes the Variable.trainable property public, which is sensible if we're discouraging use of the global collection (currently eager execution is using ResourceVariable._trainable in a bunch of places anyway). I'm leaving it read-only for now, since we should toggle in and out of the global collection when it changes.
Same change for checkpointable data structures with respect to gathering extra variables. They'll behave like subclassed Models.
I think this makes more sense than trying to have a distinction between "variables" and "weights". It's also more sensible than collecting everything that would get checkpointed, since that will include Optimizer slot variables and metrics. Collecting those is generally pointless, and accidentally adding them to gradient tapes would be horribly confusing.
PiperOrigin-RevId:
198656079
HyoukJoong Lee [Thu, 31 May 2018 01:43:40 +0000 (18:43 -0700)]
Automated g4 rollback of changelist
195379693
PiperOrigin-RevId:
198654780
Sanjoy Das [Thu, 31 May 2018 01:35:42 +0000 (18:35 -0700)]
Expose xla_disable_hlo_passes via ExecutableBuildOptions.
PiperOrigin-RevId:
198654099
Ruoxin Sang [Thu, 31 May 2018 01:11:10 +0000 (18:11 -0700)]
Remove code returning bad status when the input pointer is nullptr in internal
functions. That should be a programmatic error and we have full control of
internal functions, so it is OK to crash if error happens.
PiperOrigin-RevId:
198651749
Justin Lebar [Thu, 31 May 2018 00:59:50 +0000 (17:59 -0700)]
[XLA] Add parsers for Window and ConvolutionDimensionNumbers.
Also modify relevant ToString functions so we can have the property
Parse(ToString(x)) == x.
PiperOrigin-RevId:
198650340
Nupur Garg [Thu, 31 May 2018 00:54:02 +0000 (17:54 -0700)]
Enable TOCO pip command line binding.
PiperOrigin-RevId:
198649827
Chris Ying [Thu, 31 May 2018 00:38:13 +0000 (17:38 -0700)]
Fix bug with renorm + virtual_batch_size.
PiperOrigin-RevId:
198648273
Justin Lebar [Thu, 31 May 2018 00:00:50 +0000 (17:00 -0700)]
[XLA] Switch replay_computation to use LocalClient.
This lets replay_computation build an executable once and run it
multiple times. This is particularly important because in XLA:GPU, the
first run of an executable does some autotuning and therefore is
unrepresentative.
This change removes --xla_hlo_profile_last_run, because I don't see how
to support it in LocalClient -- LocalClient wants the do-profile bit to
be set when we *compile*. (There may not be an easy fix for this; it
worked with regular Client because we were recompiling every time we
ran.)
PiperOrigin-RevId:
198643577
Sanjoy Das [Wed, 30 May 2018 23:54:00 +0000 (16:54 -0700)]
[TF:XLA] Bump open source llvm revision to r333547
PiperOrigin-RevId:
198642698
Ruoxin Sang [Wed, 30 May 2018 23:38:59 +0000 (16:38 -0700)]
Regard a path as a directory if it ends with "/" in GCS. This implies the assumption that if a real GCS object has file name ending with "/", it is always a directory mark rather than an object carrying actual contents.
PiperOrigin-RevId:
198640604
Michael Case [Wed, 30 May 2018 23:34:18 +0000 (16:34 -0700)]
Merge pull request #19654 from case540/branch_198629366
Branch
198629366
Sanjoy Das [Wed, 30 May 2018 23:29:25 +0000 (16:29 -0700)]
Always delete old while loop after LICM
Right now the old while loop can stick around if it had side effects, which is
incorrect.
PiperOrigin-RevId:
198639203
A. Unique TensorFlower [Wed, 30 May 2018 23:27:26 +0000 (16:27 -0700)]
Add SerialDeviceBatchScheduler which offers similar performance as the AdaptiveSharedBatchScheduler, but increased reliablility and stability.
ASBS assumes request latency can be minimized at a specific number of batch processing threads. Under reasonable load, this is true and ASBS performs well, but under low load latency is basically unaffected by the number of threads, and ASBS can learn a wide variety of 'optimal' values. If load resumes suddenly, these values can give very poor latencies. In most cases, ASBS will recover, eventually rediscovering the correct value, but we have observed other cases where the latency is so large and noisy that ASBS can't get a good signal to guide its learning and the number of threads remains stuck at the bad value.
In addition, the incremental learning nature of this algorithm means that ASBS is always exploring to some extent, which can give rise to periods of non-optimal latency. This is most significant at high utilization where the wrong number of threads can potentially overload the system.
ASBS uses latency as a proxy for keeping the tensorflow processing pipeline optimally loaded. SDBS, on the other hand, uses a direct measurement of the pipeline fullness, and adjusts its number of batch processing threads accordingly. This solves the exploration problem. SDBS solves the low load problem by not adjusting its thread count when the threads pass some idleness threshold.
PiperOrigin-RevId:
198638918
Karmel Allison [Wed, 30 May 2018 23:25:00 +0000 (16:25 -0700)]
Add a convenience function, build_supervised_input_receiver_fn_from_input_fn,
that takes an Estimator input_fn and returns an input receiver function.
PiperOrigin-RevId:
198638593
A. Unique TensorFlower [Wed, 30 May 2018 23:17:45 +0000 (16:17 -0700)]
Automated g4 rollback of changelist
198444757
PiperOrigin-RevId:
198637528
A. Unique TensorFlower [Wed, 30 May 2018 23:01:05 +0000 (16:01 -0700)]
Makes empty() support uint8 on cpu.
PiperOrigin-RevId:
198634986
Gunhan Gulsoy [Wed, 30 May 2018 23:05:33 +0000 (16:05 -0700)]
Fix setuptools version to avoid a bad release.
Skye Wanderman-Milne [Wed, 30 May 2018 23:00:26 +0000 (16:00 -0700)]
Remove environment variable to disable C API.
This is staging for removing the _USE_C_API toggle altogether.
PiperOrigin-RevId:
198634886
Alexandre Passos [Wed, 30 May 2018 22:50:43 +0000 (15:50 -0700)]
Makes most variable writes depend on the cached value.
This disallows some undefined behavior with unordered reads and writes.
PiperOrigin-RevId:
198633444
A. Unique TensorFlower [Wed, 30 May 2018 22:39:03 +0000 (15:39 -0700)]
Add HloProto support to replay_computation
PiperOrigin-RevId:
198631733
Michael Case [Wed, 30 May 2018 22:34:30 +0000 (15:34 -0700)]
Merge commit for internal changes
A. Unique TensorFlower [Wed, 30 May 2018 22:24:17 +0000 (15:24 -0700)]
Avoid recursion in ExpandDomain() as stack is not happy.
PiperOrigin-RevId:
198629366
A. Unique TensorFlower [Wed, 30 May 2018 22:03:25 +0000 (15:03 -0700)]
Add kwargs support for tpu.outside_compilation
PiperOrigin-RevId:
198625799
A. Unique TensorFlower [Wed, 30 May 2018 21:55:54 +0000 (14:55 -0700)]
Move RemoveInvolution optimization to optimizer stage.
PiperOrigin-RevId:
198624394
Brennan Saeta [Wed, 30 May 2018 22:25:46 +0000 (15:25 -0700)]
Add GCS configure ops.
PiperOrigin-RevId:
198624285
Sami Kama [Wed, 30 May 2018 22:08:34 +0000 (15:08 -0700)]
Review changes
Yifei Feng [Wed, 30 May 2018 21:58:10 +0000 (14:58 -0700)]
Merge pull request #19633 from yifeif/branch_198434814
Branch
198434814
Joshua V. Dillon [Wed, 30 May 2018 21:52:57 +0000 (14:52 -0700)]
Add `fill_triangular_inverse`, which flattens a triangular matrix in a way such that:
# Lower triangular matrix
x = tf.matrix_band_part(x, -1, 0)
x == fill_triangular(fill_triangular_inverse(x))
Code by srvasude@ which I'm submitting on his behalf.
PiperOrigin-RevId:
198623887
Benoit Steiner [Wed, 30 May 2018 21:45:56 +0000 (14:45 -0700)]
Add control dependencies to the correct graph when simplifying packing ops.
PiperOrigin-RevId:
198622727