Edit

CNTK_2_0_Beta_12_Release_Notes

  • Article

This is a summary of new features delivered with the Beta 12 release of CNTK V.2.0.

Breaking changes

This release contains the following breaking changes:

  • crop, scale, mean, and color transforms are no longer static methods of ImageDeserializer and have been moved to cntk.io.transforms.
  • ReaderConfig is renamed to _ReaderConfig (this helper class is intended for internal use only)
  • Starting from this Release CNTK libraries from CNTK NuGet Packages will be signed using sha256 Digest algorithm only. sha1-based signatures will no longer be used.

New and updated features

  • More built-in activation functions: Leaky ReLU, ELU and Parametric ReLU.
  • Argmax and Argmin are now supported.
  • It is now possible to write your own optimizer in Python by inheriting from UserLearner and overriding the update method.
  • HTK deserializers are now exposed in Python. All deserializers are exposed in C++.
  • The memory pool implementation of CNTK has been updated with a new global optimization algorithm. Hyper memory compression has been removed.
  • Significantly improved performance of numpy interop
  • Improved progress logging API and performance. Progress tracking code moved from Python to a new C++ ProgressWriter class. Users can now pass progress writers upon instantiating a Trainer, and do not have to use them explicitly in their training loop.
  • New functionality of existing operators
    • PastValue() and FutureValue() now accept a data-dependent initial state, for sequence-to-sequence modeling.
    • Where() and gather() now interprets the flags as a repeat factor, which allows to duplicate steps. Factors can be fractional.
    • ReconcileDynamicAxis() now accepts a data input without dynamic axis, to easily broadcast a constant.
  • Improved graph.plot():
    • Improved visual quality and compactness.
    • New output formats: PDF and SVG.
  • Introducing stop_gradient operator that acts as an identity in the forward pass but stops gradients from flowing back to its inputs in the backward pass.
  • New features in CNTK NuGet Package. See section on NuGet Package below.

CNTK Python API

We improved the installation of CNTK on Python. As a first step you will find some simple instructions how to pip install the CPU version of CNTK on Anaconda 3 on this page. Similar improvements for Linux as well as CNTK GPU versions will follow shortly.

CNTK C++ API

The following is added to CNTK C++ API:

  • CreateBatch, CreateSequence, CreateBatchOfSequence - create Value object representing batch, sequence and batch of sequence.
  • CopyVariableValueTo - copy data out of Value object.
  • FindByName - find a function with the given name in the Function graph.
  • FindAllWithName - find a list of functions with the given name in the Function graph
  • Progress tracking code moved from Python to a new C++ ProgressWriter class. Users can now pass progress writers upon instantiating a Trainer, and do not have to use them explicitly in their training loop.

New Examples and Tutorials

CNTK NuGet package

A new set of NuGet Packages is provided with this Release. The packages have the following new features:

  • CNTK V2 C++ Library is now published as a part of NuGet packages.
  • CNTK.GPU and CNTK.CPUOnly NuGet packages now contain header files, libraries and all dependent dll libraries required to use the C++ library for training and evaluation. Both debug and release builds are supported.

IMPORTANT! In Visual Studio Manage NuGet Packages Window change the default option Stable Only to Include Prerelease. Otherwise the packages will not be visible. The Package version should be 2.0-beta12.

IMPORTANT! NuGet package in this Release contains a breaking change related to the way CNTK libraries are digitally signed. See Breaking changes section in the beginning of these Release Notes.