* conv_grad_ops.cc
* cwise_op_inverse.cc
* cwise_op_squared_difference.cc
* resize_bilinear_op.cc
Change: 123999923

Change: 123998188

Changing API for monitors step_begin to only take step and return list of tensors to run. Support distributed training in EveryN monitor.
Change: 123996408

Change: 123989954

Change: 123989187

Make mean a method in distributions.
Add variance/std/mode as method to the base class.
Make existing distributions conformant.
Add log_likelihood to ContinuousDistribution/DiscreteDistribution.
Change: 123988053

Support for passing params into Estimator that will be passed to the function call of model_fn. Note, params are only passed if model_fn takes 4 arguments, otherwise it works the same way.
Change: 123985980

Change: 123984360

Change: 123983886

Change: 123983409

only registering some types on android and ios (instead of __ANDROID__ macro).
Change: 123983258

Adds an ExtractImagePatchesOp that performs im2col. Forward only implementation for now.  Fixes #2605.
Change: 123983077

Fixing bug in propagating normalized y into the input of the next power iteration, improved eigenvalue estimate by making use of the Rayleigh quotient, and normalizing the random input vector.
Change: 123975933

Change: 123975883

Change: 123975418

Fix incorrect typescript type annotation which was caught by the typescript compiler due to newer typings definitions for lodash library.
Change: 123971815

Change: 123971387

Change: 123971216

gradients, some variants etc.).
Change: 123967787

gradients, some variants etc.).
Change: 123967117

Change: 123964973

It was previously assumed that every op can only execute once for a given session run, however ops can have multiple kernel executions within a single run and within the same device, thus showing up multiple times in the stats pbtxt.

The new "compute time" is determined by the time difference from when the first kernel started to when the last one finished.
Change: 123958417

Change: 123940786

Change: 123939117

Change: 123937085

Change: 123937028

 * ValueError is model_dir / output_dir are empty.
 * Added support for creating parse_example from TensorSignature.
Change: 123935766

Change: 123935296

an op's shape inference function. Include functions for asserting shape rank, dim value, and
for merging shapes and dims.
Change: 123934214

Change: 123931185

Change: 123931055

Change: 123925843

Instead download the whole zip file and extract the single file.
Change: 123924244

Change: 123923043

Change: 123921509

This can be used to provide an ordering on runs that remains the same even as
new runs are added (unlike sorting the run names alphabetically).

We always load the first event whenever we add a run to a multiplexer because
FirstEventTimestamp() and Reload() both grab the same mutex. Since the first
event is just a version specifier, this shouldnt' cause any intensive
computation.
Change: 123921205

Change: 123921139

Change: 123919254

Change: 123913353

Refactoring TensorFlowLinearYYY to use LinearYYY and TensorFlowDNNClassifier / TensorFlowDNNYYY to use DNNYYY. Added DeprecatedMixin to support old interface before we remove them.

* Added dropout to DNN/Linear/DNNLinearCombined.
* Fixed tests and disabled few that are not useful at this point.
* Addressed data feeders to work without n_classes properly (preserving dtype).
* Fixing bug with LinearRegression, where loss was broadcasted into [batch_size, batch_size].
Change: 123912829