Added tf-storage component for storing TensorBoard URI state and modified tf-tensorboard.html to use it. Added tf-globals component to hold TensorBoard global variables.
Change: 125489418

Change: 125487857

improves by 56%: from 148 to 231 in the forward-backward pass.

Support both the fastest algorithm, and fall back to the fastest algorithm
without using any scratch memory, if the first algorithm fails scratch memory
allocation.

Soumith's conv-benchmarks stay the same before and after this change. But now
it can run with bigger batch size.
Change: 125484122

Change: 125476357

Also, if the TAG is not found, log an info rather than warning.
Change: 125476271

Change: 125475913

Change: 125472018

Also update RELEASE.md to forward port from 0.9 branch, and add
this new 'breaking change to the API.'
Change: 125465285

Change: 125463524

Change: 125460867

Change: 125458571

Change: 125456104

Change: 125453662

Change: 125452820

Adjustments to contrib/ffmpeg to allow the ops to run inside google3. Second try: this time, the ops aren't linked by default in google3.
Change: 125447159

Change: 125430508

Added parse_fn callback from read_batch_examples to provide a way to parse single example at a time (and push parsing before queue).
Change: 125427676

Change: 125427521

* Consume Aborted exception, because it happens on PS restarts. Keep training the model.
* Make sure to stop session before coordinator stop, to abort the queues blocked on enqueue op.
* Consume exception when coordinator is asked to stop, if threads didn't stop yet.
Change: 125427252

* optimize_loss expects 0d Tensor.
* Estimator.get_eval_ops requires to have `targets` not None.
* Adding asserts for shapes in rnn_cells.
* Added error check in rnn.rnn.
Change: 125425807

Change: 125421501

For local run, support to run ValidationMonitor while training is running if `local_eval_frequency` specified.
Change: 125412804

This reduces the number of unnecessary NumPy array allocations when
feeding values into a TensorFlow session, and can reduce heap
fragmentation.

Fixes #2942 (as far as possible). Note that we recommend that you use
tcmalloc when running TensorFlow, as it is less susceptible to heap
fragmentation with mmany large objects.
Change: 125407787

Update feature_column to use string_to_hash_bucket_fast instead of the deprecated and not stable string_to_hash_bucket.
Change: 125401852

Change: 125400736

Change: 125400645

Handled: Crossing multi-dimensional bucketized feature should assign unique id's for different dimensions.
Change: 125400598

Change: 125399142

Change: 125397022

Change: 125394884

public section.
Change: 125393890

Change: 125386366

Change: 125381550

Change: 125380874

Change: 125380336

shapes.

Also restored math_ops_test and shape_inference_testutil_test in BUILD file
that were dropped from original change somehow.
Change: 125376193

Change: 125375538

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