comment/reformat in examples.
Change: 125374843

Refactor Get2dOutputSizes/Get2dOutputSizesVerbose/Get3dOutputSizes to share a common 1-dimensional GetWindowedOutputSize/GetWindowedOutputSizeVerbose.

The output sizes and padding of each dimension of a windowed operation (such as convolution or pooling) are orthogonal and can be computed independently. We can simplify the code by providing a 1D size computation and calling it for each dimension.

Also remove special cases for 1x1 spatial convolutions in dimension calculations; they add complexity and are a case that the general code handles correctly.

In general, 2D convolutions and their gradients have a lot of shape calculation code that is duplicated for each spatial dimension. This CL is a step in the direction of treating spatial dimensions the same so we can share more code.
Change: 125360639

Change: 125359897

This change should fix issues where (e.g.) a node is added after queue
runners are started, and an error like `NotFoundError: FetchOutputs
node input_producer_Dequeue:0: not found` is reported.

The cause of the error was an incorrect multiple read of the
`Graph.version` property. A graph has a `version` property that is
(currently) defined to be the count of nodes in the graph. To keep the
front-end and session graphs in sync, the `tf.Session` maintains a
`_current_version` property, which is intended to be the last version
at which the graph was synced to the backend (via
TF_ExtendGraph). However, the logic was wrong in the face of
concurrent writes to the graph:

1. x = read `Session._current_version`.
2. y = read `Session._graph.version`.
3. If y > x:
   - Call `TF_ExtendGraph(Session._graph.as_graph_def(from_version=x))`
4. z = read `Session._graph.version`.
5. Set `Session._current_version` to z.

Written this way, the race condition is clear: a node added after step
3 and before step 4 would never be included in an TF_ExtendGraph call,
and so would never be added to the graph.

The fix involves returning the current graph version atomically with
the GraphDef, using a new internal method called
`Graph._as_graph_def()`. These are protected by the graph's mutex, to
ensure that no node can be added between creating the GraphDef and
reading the version.

Fixes #2957.
Change: 125356853

Change: 125246026

Change: 125225689

Change: 125225345

Update BaseDistribution API for NaN handling.  Doc only change in BaseDistribution (which does not specify __init__ kwargs).
Change: 125224922

Change: 125215862

Change: 125215192

Change: 125215108

Change: 125208073

Change: 125206796

Change: 125205961

Change: 125205938

Change: 125190689

Change: 125190301

Change: 125189009

Change: 125180641

Change: 125178681

Change: 125134620

Change: 125133388

Change: 125133220

Change: 125132823

unique_ptr<OpRegistrationData>, so that the OpDef building can be
deferred until an call to Lookup.
Change: 125132276

Change: 125131602

Change: 125130833

tf.rank() is recently patched to work on SparseTensor.
Change: 125127586

Switches to return a std::unique_ptr<> to convey transfer of ownership.
Keeps existing functions temporarily for backwards compatibility.

Change file_system.h interface to switch entirely to the new interface,
change all callers. If this breaks someone in the public, the interface
change should be straightforward.
Change: 125127028

Change: 125126928

Change: 125125512

Change: 125120055

Change: 125115298

Change: 125114422

Change: 125113500

Change: 125113173

Change: 125112939

for comparison, instead of copying to strings.
Change: 125112408

Change: 125110928

- TensorArray shape inference now works correctly for scalar elements
- TensorArrays each now get a unique name at runtime, per step.  This means
  that they can be used in nested functional ops (e.g. tf.scan(tf.scan(...)))
END_PUBLIC

TESTED:
  - opensource_build: passed (http://ci.tensorflow.org/job/tensorflow-cl-presubmit-multijob/1778/)
Change: 125110643