This is a fairly minimal change, some functions have been cleaned up slightly
because the new summary ops dont need manual namespacing.

Behavior is approximately identical.
Change: 136745220

Change: 136745074

Change: 136744992

Change: 136743848

Change: 136743547

I originally intended to replace all the summary ops with implementations that use the TensorSummary kernel. However, I realized this would produce a lot of work without any user-visible benefit, so I decided not to do so.
Although I had already created a TensorSummary wrapper for tf.summary.scalar, now that replacing all summary implementations is a non-goal, it is better to keep using ScalarSummary for consistency and backcompat.

The behavior wrt. TensorBoard and the EventAccumulator is unchanged.
Change: 136742625

Change: 136742352

Redefine `parameters` property to be the distribution construction arguments and add new `graph_parents` property to fulfill existing role of `parameters`. Also add validate_args to {Transformed,Quantized}Distribution ctors.
Change: 136741700

Change: 136732417

separate ops for finding the min/max and doing the requantization step.
Change: 136727927

not have 'hdrs'.
Change: 136723996

Change: 136693563

Change: 136690791

Change: 136688415

Change: 136680849

Change: 136678928

Change: 136676914

Change: 136676591

This does not implement the Python API change yet for forward compatibility.
e.g. eventually tf.reverse(a, [0,1,-1]) will be the same as the old API of
tf.reverse(a, [True, True, False, False, False, True]) for a 6 dimensional
tensor a.
Change: 136675570

Change: 136672574

Change: 136668713

Adds unit tests for handling negative arguments to tf.sqrt.
Followup to cl/135111959.
Change: 136665526

headers from "core:framework" target like we do for framework_headers_lib,
so that it is available in the pip package consistently.

Tested by building the wheel with this change and validating that
the files were in site-packages/tensorflow/include/tensorflow/core/framework,
etc.

Fixes #4996.
Change: 136664737

nomsan and notsan, and concat_op_test passes in msan and tsan.
Change: 136664499

Change: 136664255

This change adapts some unit tests, that previously would fail when V2 becomes
the default.  They now pass with both V1 or V2 being the default.

This is in preparation for the upcoming switch to V2 (~end-of-month).
Change: 136663731

topological neighborhood of a device.
Change: 136663586

Change: 136663053

scalar shapes during construction.  Placholder/Variable currently
treat [] as "unknown", because it predated scalar shapes.  So
we introduce new versions that are the same implementation underneath,
but can differentiate between unknown and scalar.
Change: 136659818

On mac 64-bit, the library paths is actually at /usr/local/cuda/lib by default. The assumption that it is always at /usr/local/cuda/lib64 has led to build failures, e.g.,
http://ci.tensorflow.org/view/Nightly/job/nightly-matrix-mac-gpu/100/TF_BUILD_IS_OPT=OPT,TF_BUILD_IS_PIP=PIP,TF_BUILD_PYTHON_VERSION=PYTHON2,label=gpu-mac/console
Change: 136653210

Before this CL, Device::Sync() was called only when the executor graph contains a SINK node. However the SINK node in the graph is guaranteed to be reachable from SOURCE. This CL makes sure Device::Sync() is always called before the executor exits.

As a separate issue, we may consider if we want the executor graph to be "well-formed" with a single SOURCE and a single SINK.
Change: 136652586

Before this change, the EventAccumulator parses the new scalar_summaries as
a one-element np array like this: np.array(3, dtype=np.float32).

This is the same as a number wrt equality checking (i.e. expression above is
equal to 3.0), but it isn't json serializable, which causes problems for
TensorBoard.
Change: 136647129

Change: 136645313

Change: 136644858

Change: 136644641

Change: 136643977

Change: 136641403

the second best split in finished_nodes_op.
Change: 136640840

Change: 136637890