Change: 125834391

Change: 125830950

Change: 125830566

Change: 125829994

- dist/index.html now uses only relative, not absolute, imports
- tf-tensorboard now contains all the demo logic; tf-tensorboard-demo was removed. The vulcanization logic was simplified.
- Fixed serialize_tensorboard, and cleaned it up a bit.
Change: 125827649

Change: 125827302

Change: 125821416

Change: 125816703

Fix linker errors by adding a dependence on constant_op to targets using fill_functor. This is needed because the templates in fill_functor.h are instantiated in constant_op.cc.
Change: 125816082

Change: 125813148

This will allow error checking for subclasses of MeanValue, SampleValue, etc.

For example, for ADVISampleValue, we can ensure that no inputs are descended
from other DistributionTensors with intermediate tensor calculations (because these
require a special Jacobian term added to the score function, which can
only be done by TransformedDistribution objects).
Change: 125812475

the DNNLinearCombinedEstimators work properly.
Change: 125811654

RealValuedColumn.to_weighted_sum didn't use variable scopes, so it was impossible to reuse the values in a graph. This is inconsistent with the rest of layers.

It also manually performed a matrix multiply and this was replaced with a more standard formulation.
Change: 125808727

Change: 125803613

Change: 125799092

Change: 125795153

Change: 125794105

Added Python interfaces to reset resource containers.
pywrap_tensorflow.TF_Reset(target, [containers]) will release resources
in all listed containers, while pywrap_tensorflow.TF_Reset(target) will
release resources in all the containers. In particular, Variables cached
on the devices will be cleared by this call.
Change: 125790643

Change: 125790131

Change: 125789944

StridedSliceGrad op implements the gradient of StridedSlice.
Also implement python benchmark for StridedSlice and simple Slice.
Fix bugs in special case optimizations in StridedSlice.

(Toward resolving bug #206)
Change: 125789921

- Prevent unnecessary tensorflow.Example copies when this method
  is used in serving.
Change: 125788885

- adds support of paths pointing to the root of a bucket, e.g gs://bucket
- adds pagination support for GetChildren
- makes a more optimal HTTP request in GetChildren
Change: 125785863

Change: 125783676

Change: 125783541

Change: 125783336

Change: 125783060

Change: 125780396

Fix a flaky test by computing half-precision gamma samples with float precision intermediate calculations.
Change: 125776264

Change: 125761757

Change: 125752088

Change: 125738602

Change: 125737177

Change: 125735275

Also updated the weight initialization to ensure that the initial weights are likely to be greater than the smaller possible value representable using a 16 bit float (~0.00006).
Change: 125733195

Change: 125729946

Change: 125729599

Change: 125729413

One TODO for DirichletMultinomial.
Change: 125729016

RE the bugfix, I noticed that my tf.slim loss metric is worse than my training loss, and I'm guessing it's because my model is evaluated concurrent with its variables being loaded from a checkpoint.
I'm running now to see if this fixes things.
Feel free to ignore this CL until then.
Change: 125728083