platform/upstream/tensorflow.git
6 years agoMerge pull request #18907 from yongtang/18363-mpi
ekelsen [Thu, 26 Apr 2018 22:30:35 +0000 (15:30 -0700)]
Merge pull request #18907 from yongtang/18363-mpi

Fix build error with MPI support

6 years agoMerge pull request #18896 from KikaTech/fix_lite_topk
ekelsen [Thu, 26 Apr 2018 22:27:08 +0000 (15:27 -0700)]
Merge pull request #18896 from KikaTech/fix_lite_topk

Fix some issues in official tf.nn.topk() in lite

6 years agoFix cmake build issues with GPU on Linux (#18775)
Yong Tang [Thu, 26 Apr 2018 20:47:51 +0000 (13:47 -0700)]
Fix cmake build issues with GPU on Linux (#18775)

* Update include file for cmake

While attempting to build tensorflow with cmake for gpu the following
build error happens:
```
/home/ubuntu/tensorflow/tensorflow/core/platform/default/gpu/cupti_wrapper.h:26:45: fatal error: cuda/extras/CUPTI/include/cupti.h: No such file or directory
compilation terminated.
CMakeFiles/tf_core_cpu.dir/build.make:3302: recipe for target 'CMakeFiles/tf_core_cpu.dir/home/ubuntu/tensorflow/tensorflow/core/platform/default/gpu/cupti_wrapper.cc.o' failed
make[2]: *** [CMakeFiles/tf_core_cpu.dir/home/ubuntu/tensorflow/tensorflow/core/platform/default/gpu/cupti_wrapper.cc.o] Error 1
CMakeFiles/Makefile2:2402: recipe for target 'CMakeFiles/tf_core_cpu.dir/all' failed
make[1]: *** [CMakeFiles/tf_core_cpu.dir/all] Error 2
Makefile:127: recipe for target 'all' failed
make: *** [all] Error 2
```
This fix tries to fix the build error above for cmake.

Signed-off-by: Yong Tang <yong.tang.github@outlook.com>
* Add xla dependency to stream_executor

Signed-off-by: Yong Tang <yong.tang.github@outlook.com>
* Remove nccl files temporarily

Signed-off-by: Yong Tang <yong.tang.github@outlook.com>
* Add gomp library to libcusolver.so

Signed-off-by: Yong Tang <yong.tang.github@outlook.com>
* Fix cmake issue

Signed-off-by: Yong Tang <yong.tang.github@outlook.com>
6 years agoMerge pull request #17602 from joeyearsley/patch-1
Martin Wicke [Thu, 26 Apr 2018 20:45:56 +0000 (13:45 -0700)]
Merge pull request #17602 from joeyearsley/patch-1

Update fold_old_batch_norms.cc to accommodate 'NCHW' format.

6 years agoFix build error with MPI support
Yong Tang [Thu, 26 Apr 2018 20:21:18 +0000 (20:21 +0000)]
Fix build error with MPI support

This fix tries to fix the issue raised in 18363 where
the bazel build with MPI support fails as a header is missing
in the include.

This fix fixes the issue. The fix is verified locally
with MPI+CUDA on Ubuntu 16.04.

This fix fixes 18363.

Signed-off-by: Yong Tang <yong.tang.github@outlook.com>
6 years agoMerge pull request #18885 from drpngx/branch_194337205
Yifei Feng [Thu, 26 Apr 2018 18:27:25 +0000 (11:27 -0700)]
Merge pull request #18885 from drpngx/branch_194337205

Branch 194337205

6 years agoFix conflict.
Patrick Nguyen [Thu, 26 Apr 2018 17:00:06 +0000 (10:00 -0700)]
Fix conflict.

Use the exposed Scaffold.default_init_op instead of _default_init_op.

6 years agoMerge pull request #18883 from av8ramit/fix_master_keras_bin_acc
Amit Patankar [Thu, 26 Apr 2018 15:35:48 +0000 (08:35 -0700)]
Merge pull request #18883 from av8ramit/fix_master_keras_bin_acc

Fix critical metrics computation bug with Model in Eager mode.

6 years agoMerge pull request #18842 from aaroey/substr_bug_fix
gracehoney [Thu, 26 Apr 2018 14:33:34 +0000 (07:33 -0700)]
Merge pull request #18842 from aaroey/substr_bug_fix

Fix a bug where string::substr is used with wrong position.

6 years agoFix some issues in official tf.nn.topk() in lite
Scott Tseng [Thu, 26 Apr 2018 09:30:08 +0000 (17:30 +0800)]
Fix some issues in official tf.nn.topk() in lite

6 years agoMerge commit for internal changes.
Patrick Nguyen [Thu, 26 Apr 2018 03:30:00 +0000 (20:30 -0700)]
Merge commit for internal changes.

6 years agoGo: Update generated wrapper functions for TensorFlow ops.
A. Unique TensorFlower [Thu, 26 Apr 2018 02:47:06 +0000 (19:47 -0700)]
Go: Update generated wrapper functions for TensorFlow ops.
PiperOrigin-RevId: 194337205

6 years agoImprove shape invariant error message for tf.while_loop.
Skye Wanderman-Milne [Thu, 26 Apr 2018 02:43:14 +0000 (19:43 -0700)]
Improve shape invariant error message for tf.while_loop.

PiperOrigin-RevId: 194336902

6 years agoReduce number of combinations that are tested
A. Unique TensorFlower [Thu, 26 Apr 2018 02:18:53 +0000 (19:18 -0700)]
Reduce number of combinations that are tested

PiperOrigin-RevId: 194335483

6 years agoUpdate ops-related pbtxt files.
A. Unique TensorFlower [Thu, 26 Apr 2018 02:18:20 +0000 (19:18 -0700)]
Update ops-related pbtxt files.

PiperOrigin-RevId: 194335460

6 years agoCudnn RNN v2 kernels with autotune capability
James Qin [Thu, 26 Apr 2018 02:00:21 +0000 (19:00 -0700)]
Cudnn RNN v2 kernels with autotune capability

CudnnRNN V2 kernels run all applicable cudnn rnn algorithms and pick the best one for following runs.
* To enable autotune, TF_CUDNN_RNN_USE_AUTOTUNE and TF_CUDNN_RNN_USE_V2 need to be set to {"1" or unset}.
* TF_CUDNN_RNN_USE_AUTOTUNE does not work with existing CudnnRNN kernels.
* V2 kernels work with existing cudnn checkpoints, since it doesn't change persistence format.

This change
* Introduces v2 kernels as templates inheriting the v1 kernels.
* Profiles fwd and bak runs in v2 kernel (forward pass)
* Exposes the chosen algorithm as fwd op output and bak op input.
* Changes rnn descriptor cache key to include AlgorithmDesc (since cudnn rnn descriptor can't be reused across different algorithms)
* Updates unittests s.t. it tests both v1 and v2 kernels. When testing v2 kernels, autotune is turned on.

PiperOrigin-RevId: 194333948

6 years agoFix critical metrics computation bug with Model in Eager mode.
Francois Chollet [Wed, 25 Apr 2018 00:00:40 +0000 (17:00 -0700)]
Fix critical metrics computation bug with Model in Eager mode.

6 years agoPreserve guarantees about const-ness when creating TensorFlow functions.
Akshay Agrawal [Thu, 26 Apr 2018 00:59:01 +0000 (17:59 -0700)]
Preserve guarantees about const-ness when creating TensorFlow functions.

PiperOrigin-RevId: 194328218

6 years agoAutomated g4 rollback of changelist 194268101
A. Unique TensorFlower [Wed, 25 Apr 2018 23:37:02 +0000 (16:37 -0700)]
Automated g4 rollback of changelist 194268101

PiperOrigin-RevId: 194318022

6 years ago[XLA] Redesign: local_client_test_base and tests that depend on it.
A. Unique TensorFlower [Wed, 25 Apr 2018 23:31:14 +0000 (16:31 -0700)]
[XLA] Redesign: local_client_test_base and tests that depend on it.

PiperOrigin-RevId: 194317245

6 years agoVarious small cleanups and simplifications
Benoit Steiner [Wed, 25 Apr 2018 23:22:29 +0000 (16:22 -0700)]
Various small cleanups and simplifications

PiperOrigin-RevId: 194316037

6 years agoCheckpointable: better handling of objects which aren't being restored
Allen Lavoie [Wed, 25 Apr 2018 23:20:22 +0000 (16:20 -0700)]
Checkpointable: better handling of objects which aren't being restored

initialize_or_restore on a tf.train.Checkpoint status object will now initialize
any variables which aren't being restored, which is closer to the behavior when
executing eagerly (and makes it easier to use).

Fixes a bug where assert_consumed() would miss some Python objects which aren't
part of the object graph being restored. It will now (correctly/as documented)
complain about unmatched Python objects in the dependency graph.

PiperOrigin-RevId: 194315742

6 years ago[XLA] Redesign: migrate tests [m-r].* and vector_ops test.
A. Unique TensorFlower [Wed, 25 Apr 2018 22:29:12 +0000 (15:29 -0700)]
[XLA] Redesign: migrate tests [m-r].* and vector_ops test.

Deleted a map_test case about versioned handle since such case no longer exists.

PiperOrigin-RevId: 194307861

6 years agoDocs
A. Unique TensorFlower [Wed, 25 Apr 2018 22:28:22 +0000 (15:28 -0700)]
Docs

PiperOrigin-RevId: 194307738

6 years agoAutomated g4 rollback of changelist 194261487
A. Unique TensorFlower [Wed, 25 Apr 2018 22:25:15 +0000 (15:25 -0700)]
Automated g4 rollback of changelist 194261487

PiperOrigin-RevId: 194307293

6 years agoRefactoring: Remove a redundant map from model.h.
Yu-Cheng Ling [Wed, 25 Apr 2018 22:21:16 +0000 (15:21 -0700)]
Refactoring: Remove a redundant map from model.h.
PiperOrigin-RevId: 194306629

6 years agoDisable factorization_ops_test on ASAN due to flaky timeouts.
A. Unique TensorFlower [Wed, 25 Apr 2018 21:59:57 +0000 (14:59 -0700)]
Disable factorization_ops_test on ASAN due to flaky timeouts.

PiperOrigin-RevId: 194303178

6 years agoRemove the parallel cpu backend
Sanjoy Das [Wed, 25 Apr 2018 21:36:17 +0000 (14:36 -0700)]
Remove the parallel cpu backend

PiperOrigin-RevId: 194299356

6 years agoMerge pull request #18647 from junpenglao/ordered_bijector
Martin Wicke [Wed, 25 Apr 2018 21:30:39 +0000 (14:30 -0700)]
Merge pull request #18647 from junpenglao/ordered_bijector

Ordered bijector

6 years agoFixing contrib.boosted_trees in Windows (module not linked)
Younghee Kwon [Wed, 25 Apr 2018 21:25:30 +0000 (14:25 -0700)]
Fixing contrib.boosted_trees in Windows (module not linked)
Trying to fix #14292

PiperOrigin-RevId: 194297723

6 years agoAutomated g4 rollback of changelist 193788768
A. Unique TensorFlower [Wed, 25 Apr 2018 21:03:12 +0000 (14:03 -0700)]
Automated g4 rollback of changelist 193788768

PiperOrigin-RevId: 194293938

6 years agoAutomated g4 rollback of changelist 193731341
A. Unique TensorFlower [Wed, 25 Apr 2018 20:58:54 +0000 (13:58 -0700)]
Automated g4 rollback of changelist 193731341

PiperOrigin-RevId: 194293187

6 years agoUse a built-in function shape_n instead of running many array_ops.shape and
A. Unique TensorFlower [Wed, 25 Apr 2018 20:55:23 +0000 (13:55 -0700)]
Use a built-in function shape_n instead of running many array_ops.shape and
stacking the results.

PiperOrigin-RevId: 194292637

6 years agoRemove StreamExecutorUtil::ConvertStatus.
Justin Lebar [Wed, 25 Apr 2018 20:27:30 +0000 (13:27 -0700)]
Remove StreamExecutorUtil::ConvertStatus.

This function is a nop; StreamExecutor's Status is the same as
TensorFlow's Status.

PiperOrigin-RevId: 194288432

6 years agoAdding info to FakeQuant ops in graphviz.
A. Unique TensorFlower [Wed, 25 Apr 2018 19:55:21 +0000 (12:55 -0700)]
Adding info to FakeQuant ops in graphviz.

PiperOrigin-RevId: 194283908

6 years agoDisable tests for the parallel CPU backend
Sanjoy Das [Wed, 25 Apr 2018 19:36:15 +0000 (12:36 -0700)]
Disable tests for the parallel CPU backend

PiperOrigin-RevId: 194281504

6 years ago[XLA] Make the graph seed to be random when the hlo module seed is 0.
A. Unique TensorFlower [Wed, 25 Apr 2018 19:34:06 +0000 (12:34 -0700)]
[XLA] Make the graph seed to be random when the hlo module seed is 0.

PiperOrigin-RevId: 194281233

6 years agoSwitch tf.keras.Model.save_weights to TensorFlow format for graph networks.
Allen Lavoie [Wed, 25 Apr 2018 19:06:28 +0000 (12:06 -0700)]
Switch tf.keras.Model.save_weights to TensorFlow format for graph networks.

Some tweaks to support loading checkpoints into modified Python programs.

Relaxes the Checkpointable consistency check for object matching: if the same
object in the checkpoint matches two different Python objects, it will just
choose the first one that matches (based on whichever traversal of the Python
dependency graph it's doing). assert_consumed() on the status object will fail,
but this gives users the option of continuing anyway.

Adds a "weight-bearing layer index" dependency to graph networks which skips
Layers without weights, in addition to the regular layer index. This allows
users to add Layers without weights while not breaking checkpoints, as they
could when matching with flattened weights from HDF5 format.

Eventually I'd like to add a dependency structure which matches the topology of
the graph itself (so a Layer would have checkpoint dependencies on other Layers
it outputs to), but there are some subtleties before that's useful (it'd need
something like a secondary check that the Python classes match). I think the
scheme in this CL is robust enough for general use, and adding more dependencies
later can make it more robust now that users won't run into consistency check
errors (previously more dependencies would only make matching more picky).

PiperOrigin-RevId: 194277075

6 years agoRemoving remove_undocumented calls from tensorflow/python.
Anna R [Wed, 25 Apr 2018 18:52:08 +0000 (11:52 -0700)]
Removing remove_undocumented calls from tensorflow/python.

PiperOrigin-RevId: 194274698

6 years agoDisable nasnet_test on MSAN due to flaky timeout.
A. Unique TensorFlower [Wed, 25 Apr 2018 18:48:34 +0000 (11:48 -0700)]
Disable nasnet_test on MSAN due to flaky timeout.

PiperOrigin-RevId: 194274190

6 years agoSet up a basic README with a pointer to the dev summit colab.
Dan Moldovan [Wed, 25 Apr 2018 18:47:01 +0000 (11:47 -0700)]
Set up a basic README with a pointer to the dev summit colab.

PiperOrigin-RevId: 194273919

6 years agoMerge pull request #18868 from av8ramit/fix_hp5y
Amit Patankar [Wed, 25 Apr 2018 18:49:14 +0000 (11:49 -0700)]
Merge pull request #18868 from av8ramit/fix_hp5y

Adding the h5py dependency in devel docker files.

6 years agoSerialize strings properly when using TOCO for model conversion.
A. Unique TensorFlower [Wed, 25 Apr 2018 18:25:10 +0000 (11:25 -0700)]
Serialize strings properly when using TOCO for model conversion.

PiperOrigin-RevId: 194270132

6 years ago[TF:XLA] Fix entry computation layout assignment in the HLO parser
A. Unique TensorFlower [Wed, 25 Apr 2018 18:22:37 +0000 (11:22 -0700)]
[TF:XLA] Fix entry computation layout assignment in the HLO parser

For tuple shapes, the layouts were not correctly attached to the module.

PiperOrigin-RevId: 194269675

6 years agoEnabled TENSORFLOW_USE_ABSL for mobile. TENSORFLOW_USE_ABSL is now enabled in all...
A. Unique TensorFlower [Wed, 25 Apr 2018 18:14:49 +0000 (11:14 -0700)]
Enabled TENSORFLOW_USE_ABSL for mobile. TENSORFLOW_USE_ABSL is now enabled in all cases.

This lets TENSORFLOW_USE_ABSL be removed (as it will always be true), and is necessary to migrate TensorFlow to absl.

PiperOrigin-RevId: 194268101

6 years agoDelete from_stream_executor_status.h.
Justin Lebar [Wed, 25 Apr 2018 17:58:13 +0000 (10:58 -0700)]
Delete from_stream_executor_status.h.

It's dead code.

PiperOrigin-RevId: 194265380

6 years agoAdding install for libhdf5-serial-dev.
Amit Patankar [Wed, 25 Apr 2018 17:57:06 +0000 (10:57 -0700)]
Adding install for libhdf5-serial-dev.

6 years agoAdding the h5py dependency in devel docker files.
Amit Patankar [Wed, 25 Apr 2018 17:49:30 +0000 (10:49 -0700)]
Adding the h5py dependency in devel docker files.

6 years agoAdd eager_service.proto
Akshay Modi [Wed, 25 Apr 2018 17:41:05 +0000 (10:41 -0700)]
Add eager_service.proto

PiperOrigin-RevId: 194262260

6 years agoAutomated g4 rollback of changelist 194186080
A. Unique TensorFlower [Wed, 25 Apr 2018 17:36:13 +0000 (10:36 -0700)]
Automated g4 rollback of changelist 194186080

PiperOrigin-RevId: 194261487

6 years agoImprove look of TensorFlow Lite Demo App
Andrew Selle [Wed, 25 Apr 2018 17:21:14 +0000 (10:21 -0700)]
Improve look of TensorFlow Lite Demo App

- Include logo.
- Fix icon to be tensorflow logo
- Highlight first item as bigger.
- Grey out not very good matches.

PiperOrigin-RevId: 194259027

6 years agoTest more types in ternary_ops_test.
Jacques Pienaar [Wed, 25 Apr 2018 17:09:07 +0000 (10:09 -0700)]
Test more types in ternary_ops_test.

PiperOrigin-RevId: 194256929

6 years agoMerge pull request #18843 from av8ramit/fix_devel_docker
Amit Patankar [Wed, 25 Apr 2018 17:04:03 +0000 (10:04 -0700)]
Merge pull request #18843 from av8ramit/fix_devel_docker

Fixing the mock import error for devel docker.

6 years agofix Non-ASCII character error
Junpeng Lao [Wed, 25 Apr 2018 17:01:38 +0000 (19:01 +0200)]
fix Non-ASCII character error

6 years agoStop using gpu:: as an alias for stream_executor::.
Justin Lebar [Wed, 25 Apr 2018 16:36:24 +0000 (09:36 -0700)]
Stop using gpu:: as an alias for stream_executor::.

Also do a few related namespace cleanups.

PiperOrigin-RevId: 194252437

6 years ago[XLA:GPU] Set default kernel unrolling factor to 4
Benjamin Kramer [Wed, 25 Apr 2018 16:25:25 +0000 (09:25 -0700)]
[XLA:GPU] Set default kernel unrolling factor to 4

NVidia GPUs have 4-wide vector stores, so this is the natural unrolling factor
there. After recent fixes LLVM properly vectorizes the stores, giving a decent
performance win across the board and a huge performance win for F16 kernels.

At the moment the unrolling has no cost model and only applies to loop fusions.
All of the loop fusions I looked at were memory bound, unrolling them is a
clear win. When we add a cost model and/or auto tune later it has to be at
least as good as unrolling all the loop fusions.

PiperOrigin-RevId: 194249951

6 years agoUpdate use of squeeze_dims to axis in call to squeeze
James Keeling [Wed, 25 Apr 2018 16:19:57 +0000 (09:19 -0700)]
Update use of squeeze_dims to axis in call to squeeze

This was causing a warning every time a user created a MonitoredSession.

PiperOrigin-RevId: 194248478

6 years agoBUGFIX: Correct a typo in fisher_factors.py.
A. Unique TensorFlower [Wed, 25 Apr 2018 14:35:23 +0000 (07:35 -0700)]
BUGFIX: Correct a typo in fisher_factors.py.

PiperOrigin-RevId: 194238853

6 years agoAdd a try/finally to context managers.
Mark Daoust [Wed, 25 Apr 2018 12:59:33 +0000 (05:59 -0700)]
Add a try/finally to context managers.

This way the state is always reset if an error occurs after the context manager yields.

PiperOrigin-RevId: 194228925

6 years agoFix reference computation in client_library_test_base for multi device case
A. Unique TensorFlower [Wed, 25 Apr 2018 10:07:12 +0000 (03:07 -0700)]
Fix reference computation in client_library_test_base for multi device case

The constructor taking an execution option to specify multiple devices
didn't created a reference client resoulting in a SEGV and then later it
tried to execute the reference computation on multiple devices on the
interpreter what isn't supported.

PiperOrigin-RevId: 194217186

6 years agoMake CriticalSection work inside a Dataset with eager execution enabled.
Asim Shankar [Wed, 25 Apr 2018 10:07:04 +0000 (03:07 -0700)]
Make CriticalSection work inside a Dataset with eager execution enabled.

tf.colocate_with() might be provided with eager tensors when
constructing TensorFlow functions (like the subgraph for map()
inside a tf.data.Dataset).

Prior to this change, the added test would fail with:
"Tensor.op is meaningless when eager execution is enabled."

PiperOrigin-RevId: 194217166

6 years agoTry to ease the timeout issue in Windows Bazel Build
A. Unique TensorFlower [Wed, 25 Apr 2018 09:50:31 +0000 (02:50 -0700)]
Try to ease the timeout issue in Windows Bazel Build

PiperOrigin-RevId: 194215622

6 years agofix pylint error
Junpeng Lao [Wed, 25 Apr 2018 08:57:46 +0000 (10:57 +0200)]
fix pylint error

6 years agoMake ProcessArgMaxOperator faster by reserving the correct output dims.
A. Unique TensorFlower [Wed, 25 Apr 2018 07:03:28 +0000 (00:03 -0700)]
Make ProcessArgMaxOperator faster by reserving the correct output dims.

PiperOrigin-RevId: 194200078

6 years agoAutomated g4 rollback of changelist 194180248
A. Unique TensorFlower [Wed, 25 Apr 2018 03:15:24 +0000 (20:15 -0700)]
Automated g4 rollback of changelist 194180248

PiperOrigin-RevId: 194186080

6 years agoUpdate ops-related pbtxt files.
A. Unique TensorFlower [Wed, 25 Apr 2018 02:18:48 +0000 (19:18 -0700)]
Update ops-related pbtxt files.

PiperOrigin-RevId: 194182374

6 years agoEnable concat hoisting optimization by default.
A. Unique TensorFlower [Wed, 25 Apr 2018 01:51:45 +0000 (18:51 -0700)]
Enable concat hoisting optimization by default.

PiperOrigin-RevId: 194180248

6 years agoFix header guard.
A. Unique TensorFlower [Wed, 25 Apr 2018 01:33:25 +0000 (18:33 -0700)]
Fix header guard.

PiperOrigin-RevId: 194178828

6 years agoAdd a fuzzer for DecodeWav.
David G. Andersen [Wed, 25 Apr 2018 01:10:22 +0000 (18:10 -0700)]
Add a fuzzer for DecodeWav.
Fix an OOM error in which a wav file could claim to be
larger than it really is, resulting in a large memory
allocation prior to the size mismatch being detected later.

PiperOrigin-RevId: 194176822

6 years agoFix critical bug in metrics computation in Eager with Model.
Francois Chollet [Wed, 25 Apr 2018 00:45:22 +0000 (17:45 -0700)]
Fix critical bug in metrics computation in Eager with Model.

PiperOrigin-RevId: 194174256

6 years ago`PartitionedCallOp`: An op for executing multi-device functions.
Akshay Agrawal [Wed, 25 Apr 2018 00:35:08 +0000 (17:35 -0700)]
`PartitionedCallOp`: An op for executing multi-device functions.

A `PartitionedCallOp` allows for execution of functions across multiple devices
but within a single process. It proceeds by placing and partitioning the graph
underlying a given function body, instantiating for each partitioned subgraph a
function. The yielded function shards, which together are equivalent to the
original function, are then executed.

`PartitionedCallOp` is not part of the public TensorFlow API.

PiperOrigin-RevId: 194173114

6 years agoFixed typo in an error message.
A. Unique TensorFlower [Wed, 25 Apr 2018 00:02:46 +0000 (17:02 -0700)]
Fixed typo in an error message.

PiperOrigin-RevId: 194169339

6 years agoDisable UseTowerEstimatorWithoutReplication.test_train_single_tower.
Igor Saprykin [Tue, 24 Apr 2018 23:52:29 +0000 (16:52 -0700)]
Disable UseTowerEstimatorWithoutReplication.test_train_single_tower.

PiperOrigin-RevId: 194168031

6 years agoFixing the mock import error for devel docker.
Amit Patankar [Tue, 24 Apr 2018 23:34:01 +0000 (16:34 -0700)]
Fixing the mock import error for devel docker.

6 years agoAdd support for tensors to numpy array related assertion methods in test_util.Tensorf...
A. Unique TensorFlower [Tue, 24 Apr 2018 22:50:49 +0000 (15:50 -0700)]
Add support for tensors to numpy array related assertion methods in test_util.TensorflowTestCase.

PiperOrigin-RevId: 194159512

6 years agoFix flaky timeouts in metric_ops_test by sharding more.
A. Unique TensorFlower [Tue, 24 Apr 2018 22:49:53 +0000 (15:49 -0700)]
Fix flaky timeouts in metric_ops_test by sharding more.

PiperOrigin-RevId: 194159328

6 years agoMake TF functions work with _USE_C_SHAPES=True.
Skye Wanderman-Milne [Tue, 24 Apr 2018 22:45:50 +0000 (15:45 -0700)]
Make TF functions work with _USE_C_SHAPES=True.

It turns out regular functions need to manually copy handle data in
addition to eager GraphModeFunctions, so I moved the C extensions to
python_api.h from eager/c_api.h.

This also cleans up function_test.py to assume the C API is enabled.

PiperOrigin-RevId: 194158700

6 years agoMerge pull request #18817 from lamberta/fix-install-linux
Billy Lamberta [Tue, 24 Apr 2018 22:48:08 +0000 (15:48 -0700)]
Merge pull request #18817 from lamberta/fix-install-linux

Docs: Clean up install_linux page

6 years agoFix a bug where string::substr is used with wrong position.
gracehoney [Tue, 24 Apr 2018 21:55:47 +0000 (14:55 -0700)]
Fix a bug where string::substr is used with wrong position.

6 years agoAdd deprecation notice to replicate_model_fn.
Igor Saprykin [Tue, 24 Apr 2018 21:52:59 +0000 (14:52 -0700)]
Add deprecation notice to replicate_model_fn.

PiperOrigin-RevId: 194150426

6 years agodocs: install_linux, move GPU section below install procedures.
Billy Lamberta [Tue, 24 Apr 2018 21:52:38 +0000 (14:52 -0700)]
docs: install_linux, move GPU section below install procedures.

6 years agoInvalidate the StatCache as well as the FileBlockCache, as once the file is
Ruoxin Sang [Tue, 24 Apr 2018 21:42:07 +0000 (14:42 -0700)]
Invalidate the StatCache as well as the FileBlockCache, as once the file is
overwritten or removed, the stat will become outdated.

PiperOrigin-RevId: 194148397

6 years agoRun shape inference directly on the graphdef instead of building an intermediate...
Benoit Steiner [Tue, 24 Apr 2018 21:32:39 +0000 (14:32 -0700)]
Run shape inference directly on the graphdef instead of building an intermediate graph.

PiperOrigin-RevId: 194146713

6 years agoshape_tuple in array_ops.stack
Alexandre Passos [Tue, 24 Apr 2018 21:26:21 +0000 (14:26 -0700)]
shape_tuple in array_ops.stack

PiperOrigin-RevId: 194145557

6 years agoRevert #18251 due to the following issue:
Yifei Feng [Tue, 24 Apr 2018 20:58:37 +0000 (13:58 -0700)]
Revert #18251 due to the following issue:
- calling convolution with args instead of kwargs from convolutionXd breaks when called within arg_scope.
- intentional use cases trigger the added dimension error.

PiperOrigin-RevId: 194140820

6 years ago- Add a way to specify custom updater args to updaters in the optimizer.
A. Unique TensorFlower [Tue, 24 Apr 2018 20:47:35 +0000 (13:47 -0700)]
- Add a way to specify custom updater args to updaters in the optimizer.
- Create RegAdagradOptimizer which allows the user to specify whether a gradient update is allowed to update the slot vars.

PiperOrigin-RevId: 194139121

6 years agoImplement hoisting of common prefix of unary ops to concat.
A. Unique TensorFlower [Tue, 24 Apr 2018 20:21:49 +0000 (13:21 -0700)]
Implement hoisting of common prefix of unary ops to concat.

PiperOrigin-RevId: 194135148

6 years agoFixing concurrency issues in RPC factory.
Jiri Simsa [Tue, 24 Apr 2018 20:13:18 +0000 (13:13 -0700)]
Fixing concurrency issues in RPC factory.

PiperOrigin-RevId: 194133903

6 years agoembedding_lookup_sparse documentation change.
A. Unique TensorFlower [Tue, 24 Apr 2018 20:08:51 +0000 (13:08 -0700)]
embedding_lookup_sparse documentation change.
Remove "(typically from FeatureValueToId)" from args descriptions.
This appears to have been an obsolete reference from an ancestor
implementation.

PiperOrigin-RevId: 194133212

6 years agoExtract OptimizeGraph function in meta-optimizer.
A. Unique TensorFlower [Tue, 24 Apr 2018 19:45:45 +0000 (12:45 -0700)]
Extract OptimizeGraph function in meta-optimizer.

PiperOrigin-RevId: 194129729

6 years agoMerge pull request #18836 from tensorflow/av8ramit-patch-3
Martin Wicke [Tue, 24 Apr 2018 19:16:18 +0000 (12:16 -0700)]
Merge pull request #18836 from tensorflow/av8ramit-patch-3

Update README.md

6 years agoClarify error encountered when serializing critical_section_executions is a warning.
A. Unique TensorFlower [Tue, 24 Apr 2018 18:49:35 +0000 (11:49 -0700)]
Clarify error encountered when serializing critical_section_executions is a warning.

PiperOrigin-RevId: 194121508

6 years agoGo: Update generated wrapper functions for TensorFlow ops.
A. Unique TensorFlower [Tue, 24 Apr 2018 18:46:17 +0000 (11:46 -0700)]
Go: Update generated wrapper functions for TensorFlow ops.
PiperOrigin-RevId: 194120868

6 years agoUpdating tests in constant_folding_test.cc so that the tests evaluate the original...
A. Unique TensorFlower [Tue, 24 Apr 2018 18:43:35 +0000 (11:43 -0700)]
Updating tests in constant_folding_test.cc so that the tests evaluate the original and optimized graphs and check that the output is the same.

PiperOrigin-RevId: 194120424

6 years agoSmall refactor of tf.keras aiming at centralizing reusable utilities in `utils`.
Francois Chollet [Tue, 24 Apr 2018 18:29:43 +0000 (11:29 -0700)]
Small refactor of tf.keras aiming at centralizing reusable utilities in `utils`.

PiperOrigin-RevId: 194118244

6 years agoClarify error message when importing a GraphDef with unknown ops.
Malcolm Reynolds [Tue, 24 Apr 2018 18:26:26 +0000 (11:26 -0700)]
Clarify error message when importing a GraphDef with unknown ops.

This should make the situation from github.com/tensorflow/tensorflow/issues/17014 less confusing.

PiperOrigin-RevId: 194117660

6 years agoFix typo in event field name.
Shashi Shekhar [Tue, 24 Apr 2018 18:24:37 +0000 (11:24 -0700)]
Fix typo in event field name.

PiperOrigin-RevId: 194117352

6 years ago show breakdown of execution cost with compute and memory cost for op summarization
A. Unique TensorFlower [Tue, 24 Apr 2018 18:22:49 +0000 (11:22 -0700)]
 show breakdown of execution cost with compute and memory cost for op summarization

PiperOrigin-RevId: 194117030

6 years ago Improve handling of scopes in folding unfused batch norms. This change allows foldin...
Raghuraman Krishnamoorthi [Tue, 24 Apr 2018 18:20:04 +0000 (11:20 -0700)]
 Improve handling of scopes in folding unfused batch norms. This change allows folding to work for MobilenetV2 with unfused batch norms

PiperOrigin-RevId: 194116535

6 years agoUpdate ops-related pbtxt files.
A. Unique TensorFlower [Tue, 24 Apr 2018 18:18:47 +0000 (11:18 -0700)]
Update ops-related pbtxt files.

PiperOrigin-RevId: 194116315