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

Change: 125724750

Change: 125722665

Change: 125720510

Added max_steps into training. Distributed training will use max_steps instead of incremental steps.
Change: 125719014

Change: 125714059

Change: 125713863

Change: 125712724

Change: 125704382

Add more unit tests for BaseEstimator/Estimator (and one for tensor_signature). Fix bug if no targets used in training. Factor out some repeated code for better test coverage.
Change: 125701861

Change: 125701630

Upgraded Eigen to the latest version that provides new scan operations. This will enable the implementation of the cumsum operation in TensorFlow
Change: 125697517

Write code to build and register entrywise unary transforms. Create transforms for unary TF arithmetic ops.
Change: 125691817

In some cases TensorFlow reads the data via RandomAccessFile in really small
chunks, which doesn't work very efficiently with HTTP requests. Adding a
read-ahead cache significantly boosts the performance.
Change: 125691397

non-deterministic.
Change: 125689946

Change: 125684645

Change: 125684232

Change: 125669565

Change: 125660054

Change: 125644391

Change: 125639609

Add tests to assert all update_op tensors returned from metric_ops produce the proper metric value, and fix streaming_auc to do so.
Change: 125639012

Change: 125637583

Change: 125636045

Change: 125624938

Change: 125624711

Moved Multinomial-related code to its own files and BUILD target.
Motivation is to keep random_ops small.
Change: 125624016

Adds a crop_and_resize_op. Crops multiple boxes from a single image and resizes them to fixed output size by bilinear interpolation.
Change: 125623274

Fix some issues with scope and variable names.
Change: 125621539

Make sure the zeros tensor for gradient accumulation for a loop invariant is initialized with the right shape. This is needed because this zeros tensor can get passed downstream when the loop exits with no iterations.
Change: 125620906

Change: 125620641

Change: 125610456

Change: 125610051

Change: 125608314

Change: 125604719