Release 0.2.0

In this release we have:

  • Added support for incremental compilation.
  • Added support for Horovod.
  • Added support for timelines to allow for easy profiling of TensorFlow graphs.
  • Fixed a major memory leak (issue #87).
  • Updated the JNI bindings to be compatible with the TensorFlow 1.9.0 release.
  • Added support for obtaining the list of available devices from within Scala.
  • Fixed bugs for some control flow ops.
  • Added support for tf.cases.
  • Added support for the RMSProp optimizer, the lazy Adam optimizer, the AMSGrad optimizer, the lazy AMSGrad optimizer, and the YellowFin optimizer.
  • Added more learning rate decay schemes:
    • Cosine decay.
    • Cycle-linear 10x decay.
    • Square-root decay.
    • More warm-up decay schedules.
  • Added support for dataset interleave ops.
  • Fixed some bugs related to variable scopes and variable sharing.
  • Fixed some bugs related to functional ops.
  • Added support for some new image-related ops, under the namespace tf.image.
  • Improved consistency for the creation of initializer ops.
  • Added support for the tf.initializer op creation context.
  • Exposed part of the TensorArray API.
  • Exposed tf.Op.Builder in the public API.
  • Improvements to the learn API:
    • Refactored mode into an implicit argument.
    • Improved the evaluator hook.
    • Removed the layer creation context mechanism, to be refactored later. It was causing some issues due to bad design and unclear semantics. The plan is to implement this, in the near future, as wrapper creation context layers.
    • Improved the Model class.
    • Fixed a bug that was causing some issues related to inference hooks in the in-memory estimator.
    • Improved logging.
  • Added support for reading and writing numpy (i.e., .npy) files.
  • Added a logo. :)
0.5.1*