Change: 138079985

 * tf.all_variables to tf.global_variables
 * tf.VARIABLES -> tf.GLOBAL_VARIABLES
 * tf.initialize_all_variables -> tf.global_variable_initializers
Change: 138078998

Change: 138078360

C++ Gradients: Adds gradient functions for some nn ops. Adds functionality to gradient checker to enable explicit initial values for some tensors.
Change: 138078095

Benchmarks:

Benchmark                                   Time(ns)    CPU(ns) Iterations
--------------------------------------------------------------------------
BM_PTruncatedNormal_gpu_1000_1000            4632369    5175938        100  184.251M items/s
BM_PTruncatedNormal_2SD_gpu_10000_100        2849437    3368804        206  283.090M items/s
BM_PTruncatedNormal_OneTail_gpu_10000_100    3300317    3905713        179  244.174M items/s
Change: 138074670

a) EIGEN_USE_NONBLOCKING_THREAD_POOL no longer exist. The non-blocking threadpool is the default (It can be disabled by the symbol EIGEN_USE_SIMPLE_THREAD_POOL).

b) An eigen_initialize.cc file is no longer needed due to recent fixes for thread safety in Eigen.
Change: 138074328

Change: 138037407

Right now, the names of all of the checked runs are stored in the url bar.
This means that when TensorBoard first loads, it generates a (sometimes very long) url.

With this change, TensorBoard stores the names of all the disabled runs. This has identical behavior except that the url bar will be short if all the runs are checked, and long if all the runs are unchecked.

As a general invariant, I think loading a TensorBoard without touching the UI should always generate a short url (just /scalars maybe), and the url should grow long only if you add in settings.
Change: 138032618

Change: 138030455

- Move the URI parsing and creation functions to path.h so that we don't
  depend on core/platform from core/lib/io.
- Fixes https://github.com/tensorflow/tensorflow/issues/5316
Change: 138023841

Make a doc generator method public so that it can be reused in different versions of the doc generator.
Change: 138023534

Change: 138021741

The Stepper API enables continue-to actions (i.e., the cont() call) on a TF graph. Stepper API and the already-existing Analyzer API constitute the base of the tfdbg Python API.

Features of the Stepper:
1) Uses TensorHandles to cache results from previous cont() calls and avoid unnecessary recomputation during subsequent cont() calls.

2) Supports cont() calls on backward optimization runs() by keeping track of which Variables are in a dirty state (updated by previous cont() calls) and restoring them if necessary. Also provides an option to disable the restoration.

3) Allows clients to override intermediate tensor values before further calls of cont().

Current limitations that will be addressed in a later CL:

1) Does not obtain intermediate tensors via the tfdbg Analyzer API. This will soon be addressed by connecting the Analyzer with the Stepper.

2) Feeding tensor handles involves an eval() call of the handle. This will be improved once TensorHandles becomes a first-class citizen in tensor types and can be fed directly.

3) Supports only a single fetch target. Support for lists, tuples and dicts of fetches will be added soon.
Change: 138019857

Remove automatically generated (Distribution)Tensors.
Change: 138018661

Change: 138017961

Change: 138017353

Change: 138015167

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: 138013328

Move the last skflow examples into learn examples. Remove skflow examples folder. Fixes #2606. Deleted usage of learn.models.logistic_regression, learn.ops.*.
Change: 138010275

Change: 138010189

Change: 138008252

Allowed images in the Images panel to approach being pixelated after being scaled in Firefox. This solution is a brief CSS change.
Change: 138005373

Change: 138001884

Change: 138001077

Change: 137996377

Change: 137995278

Fixes #5360
Change: 137993461

Change: 137990808

This change completes support for distributed partial run.
RPC/Proto changes:
- PartialRunSetup RPC is added to the Master interface.
- A is_partial bool is added to the RunStepRequest.

MasterSession implementation:
- When the MasterSession gets a PartialRunSetupRequest, it constructs the graph and registers the partitions.
- When a is_partial RunStepRequest is received, the feeds and fetches are validated, and  corresponding is_partial RunGraphRequests are sent to the workers.
Change: 137989990

Change: 137989913

Change: 137989251

Change: 137988710

Change: 137974467

Change: 137973576

Includes tf.contrib.integrate.odeint, an integrator for ODEs patterned off of
scipy.integrate.odeint.
Change: 137973526

Change: 137969509

Change: 137964954

This makes "Smart Reply 5K" the first in the list of tensors, and we now choose the first
tensor in the embeddings array as the default tensor.
Change: 137964537

Reload all saveable objects from the graph.
Change: 137964298

Change: 137961957