Change: 150680902

This is calling Square<int32>(65280), the result doesn't fit into an int32.
LLVM takes advantage of this and changes the behavior of the code, making the
test fail when optimization is enabled (e.g. in msan and tsan mode).
Change: 150672767

Change: 150671385

Change: 150670512

Older versions of Windows (e.g., Windows 7) have lower temporal resolution in timestamps than other platforms. On such Windows machines, multiple dumps of the same or different tensors may end up with the same timestamp, causing
1) failures in the causality check in the constructor of DebugDumpDir,
2) occasional overwriting of previous tensor dumps and core metadata files,
3) loss of ordering information for debug watches of a tensor that is dumped multiple times during a Session::Run() call, e.g., due to While loops.

This CL addresses those two issues:
1) DebugDumpDir._validate_dump_with_graphs now examines tensor dumps of the same timestamp to take into account the possibility of timestamp collision during dependency validation.
2) During the dumping of tensors, file path collisions are checked and if they occur, an index suffix (e.g., "-1", "-2") is appended to the file path.
3) The core metadata Events files now include session_run_count in their names, to avoid collision due to same timestamps.
Change: 150670298

Added three attributes to the debug op "DebugNumericSummary" used in tfdbg-based TensorBoard health pills:
1) lower_bound (type: float)
2) upper_bound (type: float)
3) mute_if_healthy (type: bool)

lower_bound and upper_bound make it possible to customize thresholds beyond which tensor elements are counted as -inf or inf. mute_if_healthy makes it possible to mute a DebugNumericSummary op unless there are nan, -inf or inf elements in the watched tensor, which is useful for reducing the amount of health pill data.

Changes are made in the C++ DebugNodeInserter class, so that these attributes can be directly set from Python methods such as tf_debug.watch_graph() using the following syntax in the debug_ops argument:
debug_ops=["DebugNumericSummary(attribute_name=attribute_value)"]

e.g.,
debug_ops=["DebugNumericSummary(lower_bound=-100.0; mute_if_healthy=true)"]

Currently, string, float, int, and bool attribute value types are supported.
Change: 150665493

Change: 150664674

Change: 150664185

Change: 150663008

Change: 150659388

Branch 150655275

Resolved conflicts in :
tensorflow/compiler/xla/service/call_graph.cc
tensorflow/compiler/xla/service/call_graph.h
tensorflow/compiler/xla/service/call_graph_test.cc
tensorflow/core/BUILD
tensorflow/python/ops/image_ops_test.py

Change: 150655275

Change: 150653208

Change: 150652150

Change: 150651883

Change: 150650257

Change: 150648734

Change: 150648084

Change: 150646570

Change: 150645199

Change: 150642610

Include option to run TensorForest eval locally even if training is distributed, which is a common setup.
Change: 150640056

Change: 150633544

Fixes: #8047
Change: 150631022

Change: 150630961

Change: 150616846

special_math_ops:  Fix handling of empty arrays in lbeta.  We now follow the (standard) convention that the sum over the empty set is 0, and the product is 1.  This same standard is followed elsewhere in TF, so this is a fix.

As a side benefit, no conditionals or asserts are needed.
Also, this fixes a Github issue. #8391#issuecomment-286648477
which is addressed by the test  test_four_dimensional_arg_with_partial_shape_dynamic
Fixes #8391
Change: 150614295

Change: 150570754

This makes the Backspace and Delete keys work consistently in the Curses-based UI on Linux and Mac.

Previously, for example, Backspace didn't work on some Macs and users had to use Ctrl-H to replace it.
Change: 150542283

Fix misalign of code blocks in `tensorflow/tools/ci_build/README.md`

The code blocks in `tensorflow/tools/ci_build/README.md` are misaligned,
due to the unaligned pair of "```bash" and "```".

This fix fixes the misalign.

Change: 150523210

Change: 150511737

Change: 150510103

The bug was that the parsing code generated blocks of output, where each block may contain 0 or more var-len batch entries.  I had originally assumed that each block corresponded to a single minibatch entry.  The existing unit tests were small enough that this always ended up to be true because for small batch sizes, the thread pool just gave each minibatch entry its own block.  But for any reasonable batch size, this assumption broke down and I would accidentally skip copying some minibatch data to the output Tensors.

This has now been fixed.
Change: 150509644

Change: 150506154

Lower amount of unrolling so runs on k20 without __launch_bounds__.
Change: 150505830

Let the user view audio summaries at previous steps.
Change: 150504867