Proto commits in hughperkins/tf-coriander

These commits are when the Protocol Buffers files have changed: (only the last 100 relevant commits are shown)

2017-06-07

Note that RewriterConfig is not subject to API guarantees

2017-05-16

[XLA] Add F16 support to the Literal protobuf and LiteralUtils class. (#9913) * Add F16 support to the Literal protobuf and LiteralUtils class. No support has been added to any public backend, however the unit tests demonstrate that the literals can store and retreive data correctly. * Changes after code review * Use an alternative form of local initialization * Change a few more C-style casts to C++ casts

2017-05-12

Bugfix for proto file changes introduced in recent while_loop/stop_gradient PR. PiperOrigin-RevId: 155886984

Add timeline support in tfprof This CL mainly adds timeline support in three views of tfprof. It includes a few other small changes: 1. Handle the case that one Op fires multiple kernels. 2. Remove the requirements for CostGraph for easier user adoption, for now. 3. Some speed improvements in graph view. 4. Consolidate the all kinds of tfprof output into one -output option. PiperOrigin-RevId: 155822542

2017-05-11

Automated g4 rollback of changelist 155779520 PiperOrigin-RevId: 155811666

Add timeline support in tfprof This CL mainly adds timeline support in three views of tfprof. It includes a few other small changes: 1. Handle the case that one Op fires multiple kernels. 2. Remove the requirements for CostGraph for easier user adoption, for now. 3. Some speed improvements in graph view. 4. Consolidate the all kinds of tfprof output into one -output option. PiperOrigin-RevId: 155779520

Fix a bug when applying stop_gradient() on the gradient of a while loop. It turned out that I really needed to have tf.gradients() to backprop correctly on the gradient loop even if we have a stop_gradient on it. So, with this CL, I suspect that we are epsilon away to support higher-order gradients for while loops. PiperOrigin-RevId: 155568884

2017-05-08

Extend tfprof to associate op stats with Python codes. It's backward compatible. Stats of a source code line are aggregated from all ops created by that line. A example. _TFProfRoot (0us/22.44ms) model_analyzer_test.py:149:run_filename_as_m...:none (0us/22.44ms) model_analyzer_test.py:33:_run_code_in_main:none (0us/22.44ms) model_analyzer_test.py:208:<module>:test.main() (0us/22.44ms) model_analyzer_test.py:132:testComplexCodeView:x = lib.BuildFull... (0us/22.44ms) model_analyzer_testlib.py:63:BuildFullModel:return sgd_op.min... (0us/21.83ms) model_analyzer_testlib.py:54:BuildFullModel:seq.append(array_... (0us/254us) model_analyzer_testlib.py:42:BuildSmallModel:x = nn_ops.conv2d... (0us/134us) ... model_analyzer_testlib.py:61:BuildFullModel:loss = nn_ops.l2_... (0us/28us) model_analyzer_test.py:134:testComplexCodeView:sess.run(variable... (0us/0us) Change: 155393864

2017-05-06

Specify a different outer class name for the generation of java output. Change: 155296402

Automated rollback of change 155258346 Change: 155292907

Return more complete device information from the GetDevices() method of grappler clusters. Change: 155264843

Extend tfprof to associate op stats with Python codes. It's backward compatible. Stats of a source code line are aggregated from all ops created by that line. A example. _TFProfRoot (0us/22.44ms) model_analyzer_test.py:149:run_filename_as_m...:none (0us/22.44ms) model_analyzer_test.py:33:_run_code_in_main:none (0us/22.44ms) model_analyzer_test.py:208:<module>:test.main() (0us/22.44ms) model_analyzer_test.py:132:testComplexCodeView:x = lib.BuildFull... (0us/22.44ms) model_analyzer_testlib.py:63:BuildFullModel:return sgd_op.min... (0us/21.83ms) model_analyzer_testlib.py:54:BuildFullModel:seq.append(array_... (0us/254us) model_analyzer_testlib.py:42:BuildSmallModel:x = nn_ops.conv2d... (0us/134us) ... model_analyzer_testlib.py:61:BuildFullModel:loss = nn_ops.l2_... (0us/28us) model_analyzer_test.py:134:testComplexCodeView:sess.run(variable... (0us/0us) Change: 155258346

2017-05-05

Merge changes from github. Change: 155209832

Merge commit for internal changes

Implement ClusterSpec Propagation in TF Master ClusterSpec propagation is a capability upgrade for TensorFlow that should make it much easier to (1) build distributed TensorFlow clusters, and (2) handle node failures. The ClusterSpec propagation capability allows TensorFlow workers to be booted independently of each other, and with no knowledge about others. The client can then construct a ClusterDef (ClusterSpec), and then send it to the TF master at session creation. The master in turn then propagates the ClusterDef along to all of the workers. Change: 155159972

2017-05-01

Add ibverbs support for RoCEv2 networks (#9432) * add gid to enable running on roce * fix typo * add gid definition in rdma.h * fix typo * gid changed to snp and iid * typo * increase hop_limit to its upperbound

2017-04-29

Add TFDBG support to GrpcSession * Along the way, unify the way the debugger works in DirectSession (non-distributed Sessions) and MasterSession (for distributed Sessions). * The SummarizDebugTensorWatches method is invoked in DirectSession::GetOrCreateExecutors() and MasterSession::HashBuildGraphOptions() method to generate keys for partition graphs and executors. * The DebugStateInterface::PublishDebugMetadata() method is used to send metadata about the debugged Session::Run() call to debug URLs. This happens in DirectSession::Run() and MasterSession::DoRunWithLocalExecution() respectively. * The DebugGraphDecoratorInterface::DecorateGraph() and DebugGraphDecoratorInterface::PublishGraph() methods are used to insert debug ops to the debugged graph and send the modified graph to debug URLs. This happens in DirectSession::GetOrCreateExecutors() and GraphMgr::InitItem(), respectively. Change: 154631802

2017-04-27

Store the number of trees and layers that were attempted to be built. Useful to terminate training. Also can be used as seeds for random operations. Change: 154472242

2017-04-22

Merge changes from github. Change: 153925676

2017-04-20

Merge commit for internal changes

[XLA] Change types of InfeedRequest and OutfeedRequest configuration fields to "bytes" instead of "string". Change: 153665746

2017-04-19

Ibverbs-based RDMA path (#8943) * initial check-in * modifications according to review feedback * fix BUILD file * Update BUILD Add missing comma in verbs/BUILD file * Buildifier fix * Buildifier fix #2 * added #define TENSORFLOW_USE_VERBS * fix defines() * buildifier fix * link option fix * fix TENSORFLOW_USE_VERBS

2017-04-17

Add a GPUOPTIONS option to force all tensors to be gpu_compatible Change: 153386455

2017-04-14

Fold the content a const op into its consumer node. Change: 153204827

2017-04-13

Add serving code for set membership split type. Change: 153101982

Automated rollback of change 153039058 Change: 153045639

Add a GPUOPTIONS option to force all tensors to be gpu_compatible Change: 153039058

2017-04-12

Minor internal change. Change: 152949383

2017-04-10

Add run_mode field to TestResults proto for differentiating between test types. Change: 152740961

tfdbg core: allow gRPC debug server to remotely disable/enable debug ops Synopsis of changes: * The EventReply protobuf is expanded: a new field called "debug_op_state_change" is added to allow the debug server to remotely enable and disable debug ops. * At the end of every debug gRPC stream, the server sends all the queued EventReply protos to the client. The client (i.e., the debugged TF runtime) receives them and toggle the enabled status of the debug ops accordingly. * Added gated_grpc attribute to existing debug ops. This new boolean attribute is set to False by default, ensuring backward compatibility in behavior. If set to True, the debug ops will send the output tensors through grpc:// streams if and only if they are currently enabled. Otherwise we say that the debug op is "gated off" at the grpc:// URL. * If a debug op is gated off at all URLs it possesses, it will perform no expensive computation and instead just emit an empty (size {0}) output tensor. Other detailed change descriptions: * All debug ops now share the same base class "BaseDebugOp" to reduce the amount of boilerplate, which has grown in size due to the new gRPC gating logic. Change: 152733779

Branch 152703253 (#9112) * Improve py_func error handling. Automatically translate some python errors into corresponding TF errors at runtime. Change: 152156821 * Update interaction with libpng so that we use the public API instead of knowledge of the internal libpng data structures. Change: 152167754 * TensorBoard plugins now contain their own name/route prefix. Change: 152167807 * Passes trainable flag to separable_conv2d biases. Change: 152170239 * Saving resource variables with a caching device. Change: 152171539 * Drop loss from estimator_spec.eval_metric_ops, as required by core Estimator. Change: 152179924 * sample_stats.percentile DOCFIX. Change: 152182295 * Added a memory optimizer to grappler. Change: 152184170 * Change default behavior of the tf runs selector: - If there are fewer than 41 runs, enable them all by default - If there are 41 runs or more, disable them all by default This is in response to user complaints that having it enable only the first ten runs by default was confusing, because it was not obvious to users that some runs had been disabled. However, it still solves the initial user complaint that having very many runs simultaneously enabled would lag the UI. I also changed the "toggle all runs" button to try to turn everything off before turning everything on. Also, I improved the logic for detecting when the runs selection is back in the default state, so that we can avoid generating long URI strings wherever possible. Change: 152188948 * Autogenerated Change: Change TensorBoard TAG to 52 Change: 152189000 * Remove warning that only happening with config cuda. Change: 152189205 * Make resource variable shared name consistent with non-resource variables. Remove colocation constraint from resource variable cached value with the variable itself. Change: 152192203 * Add a way to specify the optimization order; refactor and add constant folding to meta optimizer. Change: 152193646 * Backport fixes and improvements from external Keras. Change: 152198296 * Merge changes from github. Change: 152200430 * Go: Update generated wrapper functions for TensorFlow ops. Change: 152200754 * Update ops-related pbtxt files. Change: 152203174 * Make ImportGraphDef() work with functions. In addition to modify graph_constructor.cc, this patch adds some other functionality to enable importing fucntions: * Ability to add FunctionDefLibraries to Graphs and FunctionLibraryDefinitions (in addition to existing functions) * FunctionDefsEqual() utility function Change: 152205258 * Expand contrib test to more than just test targets. Change: 152206822 * Preserve graph version during optimization Change: 152213262 * Exclude enter and exit nodes from shape refiner's constant folding. Change: 152213637 * Allow reshape_mover and algebraic_simplifier to make multiple mutations, by avoiding the short-circuit std::any_of. Change: 152232810 * Fix dynamic_rnn transpose bug (can input/output non-3d tensors). Also a few cleanups to RNN code. Change: 152267628 * Fix flaky tests Change: 152272801 * Add an auto parallelization grappler optimization pass. Change: 152276787 * Change json.decode.JSONDecodeError to ValueError. JSONDecodeError seems to be the exception used in the simplejson module, not the json module. Change: 152278012 * Internal change. Change: 152281471 * [XLA] Force buffer sharing of separate while instructions. Change: 152288540 * replica_device_setter should work for resource variables Change: 152289915 * Fix ./configure script 1. Add %workspace% in .bazelrc file when using import statement 2. Write action_env into bazelrc file for required environment variables for OpenCL support Change: 152290700 * Pointing a number of Tensorboard graph visualization-related help links to the new locations for the correspondent API documentation. Change: 152293459 * Restore most of pull request #8606 Pull request #8606 added str(Label(...)) for most dependencies in tensorflow.bzl, allowing most functions to be used from repositories which include TensorFlow as a submodule. Unfortunately, it broke when pulled into Google and was removed in cl/152200430. This CL restores the change, except for two Android-only functions; these were the only problematic bits. Change: 152297413 * Removed dead code in Estimator. Change: 152297597 * Assert rank is at least equal to new_rank for `_sparse_inner_flatten`. Change: 152303319 * Extend quantization ranges to include 0.0f. Change: 152304380 * Remove Keras config file saving. Change: 152306552 * API backwards compatibility tests. Change: 152310869 * [TF:XLA] Add a test for an R3 -> R4 broadcast. Change: 152313967 * Fix the problem that no enough placeholders for persistent tensor batch delete The deleter_key is always a device_name, hence there is only one of it. Hence, we cannot delete >1 handles at one time. In the fix, it creates delete placeholder on demand, the max number of placeholders is _DEAD_HANDLES_THRESHOLD. Change: 152322770 * [XLA] Add several reduction tests. Change: 152323510 * Added the memory optimizer to the meta optimizer. Change: 152323689 * Started a set of utilities to categorize op types Change: 152329057 * Add AudioSpectrogram op to TensorFlow for audio feature generation Change: 152332221 * Update ops-related pbtxt files. Change: 152332812 * Automated rollback of change 152332221 Change: 152333917 * Call Py_CLEAR on dead fields during TF_RESOURCE-to-ndarray conversion Change: 152338333 * [TF contrib seq2seq] Initial, incomplete implementation of beam search decoder. **DOES NOT WORK, pushed for collaboration only** Change: 152343927 * [XLA] Change HloPassPipeline to disallow Add* calls after Run. Change: 152345578 * Automated rollback of change 152332812 Change: 152349057 * Remove all 64/32 bit compiler warnings from core/ops. Change: 152353506 * libtensorflow.so: Don't export private symbols. With this change, libtensorflow.so will only export functions defined in c_api.h. This also results in a decreased binary size of libtensorflow.so. On Linux the decrease was from roughly 150MB to 67MB. On OS X it was from roughly 101MB to 82MB. Also fixes #8923 Change: 152366053 * Add Elu ops in XLA. Change: 152383201 * Fixed test. ('broadcast_dims' has size 1) Change: 152383633 * Add more detailed error message for rank assertion in _sparse_inner_flatten. Change: 152397909 * tensor_bundle: propagrates errors related to directory creation. Change: 152401909 * matrix_adjoint added to contrib/linalg/linear_operator_util Change: 152404828 * Add an is_active method to plugins This method determines whether a plugin is active. A plugin may be inactive if say it lacks data. This new is_active method allows us to add a route to TensorBoard noting which plugins are active. The frontend could then avoid querying routes of inactive plugins. Change: 152406232 * Replace a gather op for shapes by a stack op so dilated convolutions can be placed on GPU even with strict placing (before the gather went to CPU). Change: 152411159 * [TF:XLA] Implement BatchToSpace, BatchToSpaceND, SpaceToBatch, SpaceToBatchND. Fix crashes in core implementations of the same operators for zero-sized blocks. Change: 152416903 * Estimator saves relative paths in checkpoint. Change: 152420211 * Fix layers_test exception regex matching. Change: 152422855 * Unhide bijectors. Correct TransformedDistribution docstring. Change: 152424418 * Choosing a saner default for min_eval_frequency in the constructor for Experiment for the GCS file system, because the default of 1 causes performance problems. Change: 152439984 * Inherit use_resource from scope for partitioned variables. Change: 152442103 * Support quantized reshape in hexagon runtime Change: 152445539 * tfdbg CLI: add command list_source (ls) + UI fixes and improvements The new list_source (shorthand: ls) command lists Python source files responsible for constructing the nodes and tensors encountered in the run() call. It divides the source files into two categories and list them separately. 1) files that are not part of the TensorFlow Python library, and 2) files that are a part of it. The list contains information about how many nodes, tensors and dumps of tensors the files is responsible for. The file paths contain clickable links to the existing print_source/ps command. The list_source/ls command supports filtering by file-path and node-name regex patterns. UI fixes: * Fixed inconsistent black vs. transparent background color that made the layout look messy on some terminal types. Now using the transparent color for default font color consistently. * In the print_source command output, add clickable links to expand source lines and graph elements. Change: 152446002 * tfcompile: Be a little more verbose about missing required flags. Fixes #9014 Change: 152446338 * Disable failing test cases in pooling_ops_test. Change: 152447322 * Register more types for tf.image_crop_and_resize(). Resolves #9020. Change: 152448160 * Automated rollback of change 152439984 Change: 152450929 * Add a route to TensorBoard for fetching plugin names Specifically, we add a /data/plugins_listing route to the TensorBoard application. This route responds with an object mapping the name of each initialized plugin to whether it is active. This route could help the frontend avoid issuing requests to inactive plugins. Ordered the listing of routes within application.py so there is a little more organization. Refactored the test for application to use a fake plugin. Change: 152451390 * Added the ability to retrieve the amount of usable gpu memory Change: 152453470 * Allow to set session ConfigProto in RunConfig and use it in Estimator. Change: 152454548 * Colocate ResourceVariable reads with their handles. Change: 152455939 * tfdbg: update doc for new command list_source/ls Change: 152456128 * Make rnn directions slightly easier to follow. Change: 152456296 * Internal change Change: 152458104 * Adds batch renormalization. NOTE: if you use renormalization, you might want to use faster moving average updates, i.e. lower `decay` values. Change: 152458872 * When using ImportGraphDef with a passed in ShapeRefiner, use the producer version of the GraphDef when importing; the ShapeRefiner may be initialized with a different graph_def_version, so we need to be able to override it. The test failed without the change to graph_constructor and passes with it. The test uses a legacy graph that is supported (reduction shape). Change: 152459169 * Allow any iterable for `export_strategies` arg. Change: 152461826 * Log steps/sec every 100 steps in MonitoredSession, as before. Change: 152465320 * Fixes documentation to note that the in case of ties the identity of the return value of ArgMin and ArgMaxis not guaranteed . Change: 152465346 * Automated rollback of change 152465346 Change: 152465844 * Fix shape inference fn on _ParallelConcatStart. Change: 152466076 * Fix getting started guide Explain numerical differences in loss fix one example to print Change: 152466119 * Remove superfluous mode argument. Change: 152467334 * Add a tool that converts HLO computations to tensorflow GraphDef which can be visualized on Tensorboard. This CL defines basic tensorflow::OpDef for each HLO instruction/node. More attributes (e.g. shapes, colors) will be added in the future. Change: 152477918 * [TF:XLA] Increase shard count of //third_party/tensorflow/compiler/tests:spacetobatch_test to reduce flakiness when built under ASAN. Change: 152496244 * Make projector plugin backend read assets saved via the PluginAssets API. At the same time, keep backwards compatibility with the old way of looking up assets. Change: 152504793 * Move MNIST pointers to mirror hosted by the CVDF on Google Cloud. Fixes: #9031 Change: 152504901 * Merge changes from github. Change: 152508170 * Update API after changing default step couter frequency before. Change: 152517535 * Move a few random op helper functions to header files 1. shape_inference::RandomShape 2. OpKernel::MakeShape(Tensor, TensorShape*) Change: 152522156 * addresses the divide by zero bug Change: 152522488 * Clarify doc on tf.assign. Change: 152523909 * Sparse adam for resource variables. Change: 152525327 * Automated rollback of change 152310869 Change: 152528732 * Add an env_var tf_sync_on_finish_bool that block until device has finished all queued operations in a step if true. Change: 152533676 * Add more node attributes for HloInstruction on Tensorboard e.g. shape and layout etc. Change: 152534472 * Add tf.complex64 GPU support to tf.gather. Also add ldg specializations for std::complex. Change: 152537848 * Formatting changes Change: 152544842 * Upgrade TensorBoard TypeScript to 2.2.1 See also: #8326 Change: 152545950 * TEST: Getting reasonable test sizes on linalg library, removing need for sharding. Change: 152546409 * Disabling _testSourceUtilModuleReturnsTrue as its causing opensource issues. Change: 152548721 * Fix race due to unsafe buffer forwarding in maxpooling second order gradients added in #6664. Re-enable previously flaky tests. Clean up a few minor things in maxpooling_op_gpu.cu.cc Change: 152550050 * LinearOperator: adjoint_arg kwarg added to all operators. Now, operator.apply(x, adjoint_arg=True) means that the adjoint of 'x' is taken before application of operator. Sometimes this is done more efficiently than simply taking adjoint. Change: 152560471 * Adds weighted_average_loss metric key. Change: 152560999 * Documentation: Fix bug in manual device placement example Change: 152563392 * Change for internal compatibility. * Use std::vector for storage instead of map. Do the sorting inplace and return the same vector to avoid any copies. On larger streams it is about 50% faster. Change: 152576112 * Add tf.add_n GPU support for complex64/complex128. Also adds a unit test for tf.add_n. Change: 152577190 * - Adds support for nested types in tf.case and tf.cond. - Adds a "strict" mode which disables silent unpacking of singleton lists. - Adds shape inference to tf.case. - Adds a lot of unit tests. Change: 152581097 * [XLA] Add support for folding transpose into convolution Change: 152581336 * Add a smoke test to ensure that the doc generator runs. Change: 152592164 * Add tensorboard to the _do_not_descend_map of the PublicAPIVisitor. Change: 152592268 * Add auto parallelization to meta optimizer. Enable MetaOptimizer if any one of the optimizers is on. Change: 152598517 * Update ops-related pbtxt files. Change: 152629248 * Prevent the renorm_weight from being updated too early. Change: 152631776 * Automated rollback of change 152528732 Change: 152652473 * Construct TensorBoard dashboards in a JS list Previously, adding a dashboard to TensorBoard involved changing logic in several places. As part of this effort, added constructors to dashboards. Tweaked logic in various dashboards to preserve original behavior. For instance, the graph dashboard can only perform fitting after the dashboard is attached to the DOM. Change: 152658532 * Make CheckpointSaverListener visible next to CheckpointSaverHook. Change: 152662945 * tfdbg CLI: minor bug fixes 1: The calculation of the scroll command in the scroll bar didn't take into account that the y-coordinate of the scroll block is in the ScrollBar coordinate system, while the mouse click y-coordinate is in the screen coordinate system. 2: The y position of the ScrollBar was off by one. 3: The command box is not re-created after mouse-triggered commands, leading to strange-looking cursor position. Change: 152684294 * Remove obsolete use of validate_indices from embedding_ops.py validate_indices is ignored, so it shouldn't appear in new code. Change: 152691948 * Preparation of using GMock matchers in XLA tests. Change: 152691970 * Replace RuntimeException by RuntimeError in coordinator documentation. Change: 152697758 * Move the TensorBoard debugger plugin to be internal. This feature is currently not open-source anyway. Change: 152700267 * Add a single-machine tf.learn Estimator implementation for the WALS solver. Change: 152700915 * Add tf.contrib.training.python_input -- making it easy to feed data into TensorFlow from python coroutines. Change: 152701623 * Show that QuantizeToFloat consistently introduces a small error. The error is equal to range_min - round(range_min / range_scale) * range_scale Change: 152702015 * Internal Changes Change: 152703253 * Remove tensorflow/tensorboard/plugins/debugger, as part of merge resolution.

Automated rollback of change 152528732 Change: 152652473

2017-04-08

Add auto parallelization to meta optimizer. Enable MetaOptimizer if any one of the optimizers is on. Change: 152598517

Branch 152550050 (#9059) * Improve py_func error handling. Automatically translate some python errors into corresponding TF errors at runtime. Change: 152156821 * Update interaction with libpng so that we use the public API instead of knowledge of the internal libpng data structures. Change: 152167754 * TensorBoard plugins now contain their own name/route prefix. Change: 152167807 * Passes trainable flag to separable_conv2d biases. Change: 152170239 * Saving resource variables with a caching device. Change: 152171539 * Drop loss from estimator_spec.eval_metric_ops, as required by core Estimator. Change: 152179924 * sample_stats.percentile DOCFIX. Change: 152182295 * Added a memory optimizer to grappler. Change: 152184170 * Change default behavior of the tf runs selector: - If there are fewer than 41 runs, enable them all by default - If there are 41 runs or more, disable them all by default This is in response to user complaints that having it enable only the first ten runs by default was confusing, because it was not obvious to users that some runs had been disabled. However, it still solves the initial user complaint that having very many runs simultaneously enabled would lag the UI. I also changed the "toggle all runs" button to try to turn everything off before turning everything on. Also, I improved the logic for detecting when the runs selection is back in the default state, so that we can avoid generating long URI strings wherever possible. Change: 152188948 * Autogenerated Change: Change TensorBoard TAG to 52 Change: 152189000 * Remove warning that only happening with config cuda. Change: 152189205 * Make resource variable shared name consistent with non-resource variables. Remove colocation constraint from resource variable cached value with the variable itself. Change: 152192203 * Add a way to specify the optimization order; refactor and add constant folding to meta optimizer. Change: 152193646 * Backport fixes and improvements from external Keras. Change: 152198296 * Merge changes from github. Change: 152200430 * Go: Update generated wrapper functions for TensorFlow ops. Change: 152200754 * Update ops-related pbtxt files. Change: 152203174 * Make ImportGraphDef() work with functions. In addition to modify graph_constructor.cc, this patch adds some other functionality to enable importing fucntions: * Ability to add FunctionDefLibraries to Graphs and FunctionLibraryDefinitions (in addition to existing functions) * FunctionDefsEqual() utility function Change: 152205258 * Expand contrib test to more than just test targets. Change: 152206822 * Preserve graph version during optimization Change: 152213262 * Exclude enter and exit nodes from shape refiner's constant folding. Change: 152213637 * Allow reshape_mover and algebraic_simplifier to make multiple mutations, by avoiding the short-circuit std::any_of. Change: 152232810 * Fix dynamic_rnn transpose bug (can input/output non-3d tensors). Also a few cleanups to RNN code. Change: 152267628 * Fix flaky tests Change: 152272801 * Add an auto parallelization grappler optimization pass. Change: 152276787 * Change json.decode.JSONDecodeError to ValueError. JSONDecodeError seems to be the exception used in the simplejson module, not the json module. Change: 152278012 * Internal change. Change: 152281471 * [XLA] Force buffer sharing of separate while instructions. Change: 152288540 * replica_device_setter should work for resource variables Change: 152289915 * Fix ./configure script 1. Add %workspace% in .bazelrc file when using import statement 2. Write action_env into bazelrc file for required environment variables for OpenCL support Change: 152290700 * Pointing a number of Tensorboard graph visualization-related help links to the new locations for the correspondent API documentation. Change: 152293459 * Restore most of pull request #8606 Pull request #8606 added str(Label(...)) for most dependencies in tensorflow.bzl, allowing most functions to be used from repositories which include TensorFlow as a submodule. Unfortunately, it broke when pulled into Google and was removed in cl/152200430. This CL restores the change, except for two Android-only functions; these were the only problematic bits. Change: 152297413 * Removed dead code in Estimator. Change: 152297597 * Assert rank is at least equal to new_rank for `_sparse_inner_flatten`. Change: 152303319 * Extend quantization ranges to include 0.0f. Change: 152304380 * Remove Keras config file saving. Change: 152306552 * API backwards compatibility tests. Change: 152310869 * [TF:XLA] Add a test for an R3 -> R4 broadcast. Change: 152313967 * Fix the problem that no enough placeholders for persistent tensor batch delete The deleter_key is always a device_name, hence there is only one of it. Hence, we cannot delete >1 handles at one time. In the fix, it creates delete placeholder on demand, the max number of placeholders is _DEAD_HANDLES_THRESHOLD. Change: 152322770 * [XLA] Add several reduction tests. Change: 152323510 * Added the memory optimizer to the meta optimizer. Change: 152323689 * Started a set of utilities to categorize op types Change: 152329057 * Add AudioSpectrogram op to TensorFlow for audio feature generation Change: 152332221 * Update ops-related pbtxt files. Change: 152332812 * Automated rollback of change 152332221 Change: 152333917 * Call Py_CLEAR on dead fields during TF_RESOURCE-to-ndarray conversion Change: 152338333 * [TF contrib seq2seq] Initial, incomplete implementation of beam search decoder. **DOES NOT WORK, pushed for collaboration only** Change: 152343927 * [XLA] Change HloPassPipeline to disallow Add* calls after Run. Change: 152345578 * Automated rollback of change 152332812 Change: 152349057 * Remove all 64/32 bit compiler warnings from core/ops. Change: 152353506 * libtensorflow.so: Don't export private symbols. With this change, libtensorflow.so will only export functions defined in c_api.h. This also results in a decreased binary size of libtensorflow.so. On Linux the decrease was from roughly 150MB to 67MB. On OS X it was from roughly 101MB to 82MB. Also fixes #8923 Change: 152366053 * Add Elu ops in XLA. Change: 152383201 * Fixed test. ('broadcast_dims' has size 1) Change: 152383633 * Add more detailed error message for rank assertion in _sparse_inner_flatten. Change: 152397909 * tensor_bundle: propagrates errors related to directory creation. Change: 152401909 * matrix_adjoint added to contrib/linalg/linear_operator_util Change: 152404828 * Add an is_active method to plugins This method determines whether a plugin is active. A plugin may be inactive if say it lacks data. This new is_active method allows us to add a route to TensorBoard noting which plugins are active. The frontend could then avoid querying routes of inactive plugins. Change: 152406232 * Replace a gather op for shapes by a stack op so dilated convolutions can be placed on GPU even with strict placing (before the gather went to CPU). Change: 152411159 * [TF:XLA] Implement BatchToSpace, BatchToSpaceND, SpaceToBatch, SpaceToBatchND. Fix crashes in core implementations of the same operators for zero-sized blocks. Change: 152416903 * Estimator saves relative paths in checkpoint. Change: 152420211 * Fix layers_test exception regex matching. Change: 152422855 * Unhide bijectors. Correct TransformedDistribution docstring. Change: 152424418 * Choosing a saner default for min_eval_frequency in the constructor for Experiment for the GCS file system, because the default of 1 causes performance problems. Change: 152439984 * Inherit use_resource from scope for partitioned variables. Change: 152442103 * Support quantized reshape in hexagon runtime Change: 152445539 * tfdbg CLI: add command list_source (ls) + UI fixes and improvements The new list_source (shorthand: ls) command lists Python source files responsible for constructing the nodes and tensors encountered in the run() call. It divides the source files into two categories and list them separately. 1) files that are not part of the TensorFlow Python library, and 2) files that are a part of it. The list contains information about how many nodes, tensors and dumps of tensors the files is responsible for. The file paths contain clickable links to the existing print_source/ps command. The list_source/ls command supports filtering by file-path and node-name regex patterns. UI fixes: * Fixed inconsistent black vs. transparent background color that made the layout look messy on some terminal types. Now using the transparent color for default font color consistently. * In the print_source command output, add clickable links to expand source lines and graph elements. Change: 152446002 * tfcompile: Be a little more verbose about missing required flags. Fixes #9014 Change: 152446338 * Disable failing test cases in pooling_ops_test. Change: 152447322 * Register more types for tf.image_crop_and_resize(). Resolves #9020. Change: 152448160 * Automated rollback of change 152439984 Change: 152450929 * Add a route to TensorBoard for fetching plugin names Specifically, we add a /data/plugins_listing route to the TensorBoard application. This route responds with an object mapping the name of each initialized plugin to whether it is active. This route could help the frontend avoid issuing requests to inactive plugins. Ordered the listing of routes within application.py so there is a little more organization. Refactored the test for application to use a fake plugin. Change: 152451390 * Added the ability to retrieve the amount of usable gpu memory Change: 152453470 * Allow to set session ConfigProto in RunConfig and use it in Estimator. Change: 152454548 * Colocate ResourceVariable reads with their handles. Change: 152455939 * tfdbg: update doc for new command list_source/ls Change: 152456128 * Make rnn directions slightly easier to follow. Change: 152456296 * Internal change Change: 152458104 * Adds batch renormalization. NOTE: if you use renormalization, you might want to use faster moving average updates, i.e. lower `decay` values. Change: 152458872 * When using ImportGraphDef with a passed in ShapeRefiner, use the producer version of the GraphDef when importing; the ShapeRefiner may be initialized with a different graph_def_version, so we need to be able to override it. The test failed without the change to graph_constructor and passes with it. The test uses a legacy graph that is supported (reduction shape). Change: 152459169 * Allow any iterable for `export_strategies` arg. Change: 152461826 * Log steps/sec every 100 steps in MonitoredSession, as before. Change: 152465320 * Fixes documentation to note that the in case of ties the identity of the return value of ArgMin and ArgMaxis not guaranteed . Change: 152465346 * Automated rollback of change 152465346 Change: 152465844 * Fix shape inference fn on _ParallelConcatStart. Change: 152466076 * Fix getting started guide Explain numerical differences in loss fix one example to print Change: 152466119 * Remove superfluous mode argument. Change: 152467334 * Add a tool that converts HLO computations to tensorflow GraphDef which can be visualized on Tensorboard. This CL defines basic tensorflow::OpDef for each HLO instruction/node. More attributes (e.g. shapes, colors) will be added in the future. Change: 152477918 * [TF:XLA] Increase shard count of //third_party/tensorflow/compiler/tests:spacetobatch_test to reduce flakiness when built under ASAN. Change: 152496244 * Make projector plugin backend read assets saved via the PluginAssets API. At the same time, keep backwards compatibility with the old way of looking up assets. Change: 152504793 * Move MNIST pointers to mirror hosted by the CVDF on Google Cloud. Fixes: #9031 Change: 152504901 * Merge changes from github. Change: 152508170 * Update API after changing default step couter frequency before. Change: 152517535 * Move a few random op helper functions to header files 1. shape_inference::RandomShape 2. OpKernel::MakeShape(Tensor, TensorShape*) Change: 152522156 * addresses the divide by zero bug Change: 152522488 * Clarify doc on tf.assign. Change: 152523909 * Sparse adam for resource variables. Change: 152525327 * Automated rollback of change 152310869 Change: 152528732 * Add an env_var tf_sync_on_finish_bool that block until device has finished all queued operations in a step if true. Change: 152533676 * Add more node attributes for HloInstruction on Tensorboard e.g. shape and layout etc. Change: 152534472 * Add tf.complex64 GPU support to tf.gather. Also add ldg specializations for std::complex. Change: 152537848 * Formatting changes Change: 152544842 * Upgrade TensorBoard TypeScript to 2.2.1 See also: #8326 Change: 152545950 * TEST: Getting reasonable test sizes on linalg library, removing need for sharding. Change: 152546409 * Disabling _testSourceUtilModuleReturnsTrue as its causing opensource issues. Change: 152548721 * Fix race due to unsafe buffer forwarding in maxpooling second order gradients added in #6664. Re-enable previously flaky tests. Clean up a few minor things in maxpooling_op_gpu.cu.cc Change: 152550050

2017-04-07

Automated rollback of change 152310869 Change: 152528732

2017-04-06

Added the memory optimizer to the meta optimizer. Change: 152323689

2017-04-05

API backwards compatibility tests. Change: 152310869

Branch 152232810 (#8988) * Improve py_func error handling. Automatically translate some python errors into corresponding TF errors at runtime. Change: 152156821 * Update interaction with libpng so that we use the public API instead of knowledge of the internal libpng data structures. Change: 152167754 * TensorBoard plugins now contain their own name/route prefix. Change: 152167807 * Passes trainable flag to separable_conv2d biases. Change: 152170239 * Saving resource variables with a caching device. Change: 152171539 * Drop loss from estimator_spec.eval_metric_ops, as required by core Estimator. Change: 152179924 * sample_stats.percentile DOCFIX. Change: 152182295 * Added a memory optimizer to grappler. Change: 152184170 * Change default behavior of the tf runs selector: - If there are fewer than 41 runs, enable them all by default - If there are 41 runs or more, disable them all by default This is in response to user complaints that having it enable only the first ten runs by default was confusing, because it was not obvious to users that some runs had been disabled. However, it still solves the initial user complaint that having very many runs simultaneously enabled would lag the UI. I also changed the "toggle all runs" button to try to turn everything off before turning everything on. Also, I improved the logic for detecting when the runs selection is back in the default state, so that we can avoid generating long URI strings wherever possible. Change: 152188948 * Autogenerated Change: Change TensorBoard TAG to 52 Change: 152189000 * Remove warning that only happening with config cuda. Change: 152189205 * Make resource variable shared name consistent with non-resource variables. Remove colocation constraint from resource variable cached value with the variable itself. Change: 152192203 * Add a way to specify the optimization order; refactor and add constant folding to meta optimizer. Change: 152193646 * Backport fixes and improvements from external Keras. Change: 152198296 * Merge changes from github. Change: 152200430 * Go: Update generated wrapper functions for TensorFlow ops. Change: 152200754 * Update ops-related pbtxt files. Change: 152203174 * Make ImportGraphDef() work with functions. In addition to modify graph_constructor.cc, this patch adds some other functionality to enable importing fucntions: * Ability to add FunctionDefLibraries to Graphs and FunctionLibraryDefinitions (in addition to existing functions) * FunctionDefsEqual() utility function Change: 152205258 * Expand contrib test to more than just test targets. Change: 152206822 * Preserve graph version during optimization Change: 152213262 * Exclude enter and exit nodes from shape refiner's constant folding. Change: 152213637 * Allow reshape_mover and algebraic_simplifier to make multiple mutations, by avoiding the short-circuit std::any_of. Change: 152232810 * fixing workspace.bzl * workspace.bzl further fixes * fixing tensorflow.bzl merge conflicts * fixing typo in dnn.h * fixing bad merge for dnn.h

2017-04-04

Add a way to specify the optimization order; refactor and add constant folding to meta optimizer. Change: 152193646

2017-04-03

Fix the circular dependency tf/contrib/tensorboard <--> tf/tensorboard. tf/tensorboard should not depend on tf/contrib/.. This is resolved by moving projector_config.proto from tf/contrib/tensorboard/plugins/projector to tf/tensorboard/plugins/projector. The user-facing API hasn't changed. `ProjectorConfig` is still exposed as `tensorflow.contrib.tensorboard.plugins.projector.ProjectorConfig`. Change: 152030381

2017-03-31

Migrate trees, models, testutils, and resources libraries to boosted_trees. Change: 151832033

Remove stale documtation about node colocation Change: 151778994

Move all parameters of remote graph execution to NodeDef instead of additional serialized data Change: 151771414

2017-03-25

Consolidate worker state behind a session-centric abstraction. State in workers is currently splayed across graph_mgr, rendezvous_mgr, and additional components. This has resulted in it being difficult to ensure proper cleanup and shut down of the worker components. In addition to paving the way for a more reliable shut down, this CL also sets up the beginnings of ClusterSpec propagation. ClusterSpec propagation is a capability upgrade for TensorFlow that should make it much easier to (1) build distributed TensorFlow clusters, and (2) handle node failures. After the ClusterSpec propagation capability is fully implemented, the TensorFlow workers can be booted independently of each other, and with no knowledge about others. A client can then query a central cluster scheduler or other API to find all of the workers, and then send the ClusterDef (ClusterSpec) to the TF master, which then propagates that along to all of the workers. This change is only the first of a sequence to fully implement ClusterSpec propagation in TensorFlow. Change: 151229111

Added a new OpPerformanceList message Change: 151191936

2017-03-24

Added a config option to control model pruning Change: 151130707

2017-03-22

Migrate utils and protos to contrib/boosted_trees. Change: 150897748

2017-03-20

Adds java package and outer class name to rewriter_config.proto. Change: 150664674

Adds java package and outer class name to rewriter_config.proto. Change: 150664674

Created a proto to configure the amount of graph rewriting taking place. Change: 150648084

2017-03-15

[XLA] Give Transpose its own Request, rather than piggybacking on ReshapeRequest. Avoids building unnecessary Reshape operators when Transpose was called by the client. Also avoids building Transpose operators when Reshape has identity transpose dimensions, for example when the client called the variant of ComputationBuilder::Reshape() that does not transpose. Makes the HLO graph emitted by the TF bridge more readable. Change: 150253949

2017-03-14

First version of BigQuery Reader. Change: 150016997

2017-03-10

Attempt to specify tensorflow::Reset(). Change: 149714094

2017-03-09

[TF:XLA] Add debug metadata to HLO ops. In order to support end-to-end debugging and performance profiling tooling for the TensorFlow::XLA toolchain, this change adds a DebugMetadata proto to the HloInstruction class, and pipes it through the tf2xla stack. Change: 149703349

2017-03-08

Started to open source Grappler. First application is the GPU layout optimizer. Change: 149558284

2017-03-03

Factor-out dryrun functionalities from graph_transfer_utils to remote_fused_graph_execute_op in order to cache shapes in RemtoeFusedGraphExecuteInfo Change: 149143066

2017-03-02

Add BenchmarkType to TestResults proto and fix logging of C++ microbenchmarks. C++ microbenchmarks results are now normalized by iters (like python benchmarks). Change: 149045367

2017-03-01

Add interfaces in OpKernelContext and ResourceBase for tracking persistent storage. Record persistent tensor memory and persistent memory (originally auxiliary memory) in the cost model. Change: 148920117

Add optional transport_options to RecvTensorRequest. Change: 148827175

2017-02-28

tfdbg: add option to tolerate debug op creation failures Change: 148788753

2017-02-27

[XLA] Add basic outfeed support. Change: 148699787

2017-02-24

Add remote_fused_graph_execute_info to abstract hexagon specific parameters Change: 148489487

[XLA] Add an IsFinite operation that tests elementwise whether values are finite (i.e., not NaN or Inf). Change: 148485205

Open source HParams object and proto. Change: 148474914

2017-02-23

tensorflow::Reset() now honors the device_filter in SessionOptions the same way as does tensorflow::NewSession(). This allows Reset() to succeed, even if an unrelated worker happens to be down. Change: 148343973

2017-02-22

Move tensorflow::ReaderBase to tensorflow/core/framework so that it can be used in custom op libraries and contrib. Change: 148181260

[XLA] Properly version outfeed and send operations in UserComputation. Previously outfeed and send operations were unconditionally emitted during UserComputation lowering even if the outfeed/send was not in the requested version (computation snapshot). This CL versions these operations. Also, opportunistically improve logging in UserComputation, Service, and ComputationTracker which was used to root cause the underlying bug. Change: 148170893

2017-02-17

Add compute_time and memory_time to CostGraphDef::Node. Change: 147871989

2017-02-10

Record persistent memory in tracking allocator and cost model. Change: 147123056

2017-02-09

Tolerate unrecognized top-level fields in fast Example parsing. Some internal serialization routines use unrecognized field numbers. Change: 147039073

Record kernel's host memory and device memory in cost model. Change: 146997688

2017-02-08

tfdbg core: add core metadata to debugger data stream + better support of concurrent debugged runs * Let the debugger send/dump an Event proto holding a JSON string in its log_message.message field. The JSON metadata includes, 1) An optional, client-specified global_step field that defaults to -1 if not supplied 2) A session run count 3) An executor invocation step count 4) Input names (feed keys) 5) Output names (fetched Tensor names) 6) Target node names * grpc_debug_server.EventListenerBaseServicer now requires a constructor of the type EventListenerBaseStreamHandler and will construct a new handler object from it, for every stream. This leads to better support of concurrent debugged Session::Run() calls. * Add support for path names in grpc:// URLs, such as "grpc://localhost:6000/thread1". Different path names will lead to separate gRPC streams being opened to the same server:port, supporting concurrent debugged Session::Run() calls. Change: 146896481

Rename the DEBUG logging level to DEBUGGING. Some software projects compile C/C++ code with -DDEBUG (equivalent to -DDEBUG=1) in debug builds. Notably, this is done by default in debug builds on iOS. Therefore, we should avoid using DEBUG as an identifier in C/C++ code. Change: 146840049

2017-02-07

Rename the DEBUG logging level to DEBUGGING. Some software projects compile C/C++ code with -DDEBUG (equivalent to -DDEBUG=1) in debug builds. Notably, this is done by default in debug builds on iOS. Therefore, we should avoid using DEBUG as an identifier in C/C++ code. Change: 146840049

[XLA] Update comment on RNG_UNIFORM in xla_data.proto Change: 145810242

[XLA] Rename CompilationOptions proto to ExecutionOptions. We want to put fields in this proto that aren't strictly related to compilation. Change: 146477500

Added optional string argument to infeed HLO op. Change: 145188452

Addition of Outfeed HLO op. Change: 145772331

[XLA] Move shape_with_output_layout and seed fields into ExecutionOptions. This simplifies the execution API by getting rid of two default params. Also change HloModuleConfig so it stores each of the fields of ExecutionOptions individually, instead of keeping an instance of the ExecutionOptions proto. This is necessary because HloModuleConfig already has a field derived from shape_with_output_layout -- if we stored the ExecutionOptions proto in HloModuleConfig, its shape wouldn't necessarily match the shape we already have. Change: 146477669

[XLA] Move fast-math flags into HLO module config. Previously, XLA controlled the presence/absence of fast-math flags (FMF) via a command-line flag. This patch changes things so we use a new CompileOptions proto instead. This proto lives in HloModuleConfig, and is passed to the service via ExecuteRequest. This change lets us entirely remove llvm_backend_flags.{h,cc}. In addition, this change takes us from two to one fast-math flags. Previously we tried to control "unsafe FP transformations" separately from "full fast math". It turns out that LLVM is misleadingly inconsistent in how it handles these. In the backend, they are indeed two separate options that can be enabled/disabled independently. In the frontend, however, unsafe-fp-math implies all the other FMFs. As a result, it doesn't really make sense for XLA to attempt to split out these two flags, at least not until LLVM changes how it handles them. Change: 146183994

2017-02-06

Metagraph round-tripping for resource variables. Change: 146519414

Add an op to execute a subgraph fused and transferred by GraphTransferer in a remote processor Change: 146527671

2017-02-04

Add an op to execute a subgraph fused and transferred by GraphTransferer in a remote processor Change: 146527671

2017-02-03

Metagraph round-tripping for resource variables. Change: 146519414

[XLA] Move shape_with_output_layout and seed fields into ExecutionOptions. This simplifies the execution API by getting rid of two default params. Also change HloModuleConfig so it stores each of the fields of ExecutionOptions individually, instead of keeping an instance of the ExecutionOptions proto. This is necessary because HloModuleConfig already has a field derived from shape_with_output_layout -- if we stored the ExecutionOptions proto in HloModuleConfig, its shape wouldn't necessarily match the shape we already have. Change: 146477669

[XLA] Rename CompilationOptions proto to ExecutionOptions. We want to put fields in this proto that aren't strictly related to compilation. Change: 146477500

Add a mechanism for hiding, skipping, and modifying the generated op functions for C++. A souped-up version of the hidden_ops mechanism in Python, the intent is to use this for most or all of the client languages, with a common list of changes to make in a common file and per-language overrides. Also: * include the documentation for outputs in the generated comments * several updates to C++ API to match Python * fix C++ shape function for ConcatV2 now that we use it by default * split op_gen_lib out of core:framework, since it is only used by the op generators, and I don't want to add another proto to mobile builds Change: 146267344

Use graph transfer info proto in GraphTransferer Change: 146421389

Document the unit used to store times in the cost graph. Change: 146416639

2017-02-02

Add proto for graph transfer to serialize soc node Change: 146409586

2017-02-01

Remove the "Experimental" markers of RunOptions and RunMetadata protos. This seems like a reasonable thing to do, as (1) they have been in existence for ~1 year, (2) people have added new fields to them since their inception, and (3) the C API supports them. Change: 146271432

Add a mechanism for hiding, skipping, and modifying the generated op functions for C++. A souped-up version of the hidden_ops mechanism in Python, the intent is to use this for most or all of the client languages, with a common list of changes to make in a common file and per-language overrides. Also: * include the documentation for outputs in the generated comments * several updates to C++ API to match Python * fix C++ shape function for ConcatV2 now that we use it by default * split op_gen_lib out of core:framework, since it is only used by the op generators, and I don't want to add another proto to mobile builds Change: 146267344

[XLA] Move fast-math flags into HLO module config. Previously, XLA controlled the presence/absence of fast-math flags (FMF) via a command-line flag. This patch changes things so we use a new CompileOptions proto instead. This proto lives in HloModuleConfig, and is passed to the service via ExecuteRequest. This change lets us entirely remove llvm_backend_flags.{h,cc}. In addition, this change takes us from two to one fast-math flags. Previously we tried to control "unsafe FP transformations" separately from "full fast math". It turns out that LLVM is misleadingly inconsistent in how it handles these. In the backend, they are indeed two separate options that can be enabled/disabled independently. In the frontend, however, unsafe-fp-math implies all the other FMFs. As a result, it doesn't really make sense for XLA to attempt to split out these two flags, at least not until LLVM changes how it handles them. Change: 146183994

2017-01-27

[XLA] Update comment on RNG_UNIFORM in xla_data.proto Change: 145810242

Addition of Outfeed HLO op. Change: 145772331

2017-01-21

Added optional string argument to infeed HLO op. Change: 145188452

2017-01-19

Expose the two constants that control GPU Event polling pacing as GPUOptions values that are user settable. Change: 144992308

2017-01-18

Android demo: read MultiBox priors from a txt file in Java rather than reading from a proto file in C++ code, in order to reduce code/build complexity. New model file archive with corresponding changes has been uploaded to https://storage.googleapis.com/download.tensorflow.org/models/mobile_multibox_v1a.zip Resolves #6670 Change: 144872035

2017-01-12

Delete the FunctionDef.Node code paths, now that we have switched to the NodeDef representation. Change: 144281952

2017-01-11

Add support for list(func) AttrValues. Change: 144211572