Change: 137701824

Change: 137701036

Change: 137699654

Change: 137695685

Change: 137692532

Change: 137689806

Change: 137682365

Change: 137675348

Change: 137661695

Change: 137660281

Change: 137658899

Change: 137658764

Change: 137657699

It failed because:
- assertSameElements doesn't exist in OSS
- json_format.MessageToDict doesn't exist in the OSS.

Replaced with the correct ones.
Change: 137602799

Change: 137587827

Change: 137584302

Change: 137582456

Change: 137581645

Change: 137577421

Change: 137576487

Change: 137575131

Added test to demomstrate how to compose larger graphs from several scoped
subgraph.
Change: 137574327

Change: 137572074

Change: 137572042

Change: 137568969

Change: 137568328

Change: 137566553

Change: 137564920

Adding CPU kernels for tf.scatter_nd(), tf.scatter_nd_update(), tf.scatter_nd_add(), tf.scatter_nd_sub(), tf.scatter_nd_mul() and tf.scatter_nd_div() as well as gradient functions for tf.scatter_nd() and tf.gather_nd()
Change: 137564676

Change: 137559864

Change: 137559198

Change: 137558823

Change: 137552350

Change: 137549301

Change: 137548941

1) I added support for time-based (as opposed to step-based logging) to
   LoggingTensorHook.
2) I also clarified the documentation to make it clear that LoggingTensorHook
   is triggered once every N local steps on the current worker, whereas most
   of the other hooks are triggered once every N global steps.
3) I removed the default value for 'every_n_iter' in order to remain consistent
   with CheckpointSaverHook and SummarySaverHook. A similar change was recently
   made to SummarySaverHook, so I think that sets a good precedent.
Change: 137546456

Fix asserts to pass expected before actual values.
Change: 137544983

move.
Change: 137544724

Change: 137543783

Change: 137541047