CNTK_2_0_Beta_10_Release_Notes

This page has migrated to our new site. Please update any bookmarks.

CNTK v.2.0 Beta 10 Release Notes

This is a summary of new features delivered with the Beta 10 release of CNTK V.2.0.

New Examples and Tutorials

• A Python version of the deconvolution layer and image auto encoder example was added (Example 07_Deconvolution in Image - Getting Started).
• Python distributed training example for image classification using AlexNet
• Basic implementation of Generative Adversarial Networks (GAN) networks
• Training with Sampled Softmax

Python API

The following updates are introduced to Python API:

• Operators can now be implemented in pure Python by means of UserFunctions, cf. here

  • The API is still experimental and subject to change.

• Plotting the CNTK graph Plotting a CNTK graph is now as easy as calling

  cntk.graph.plot(node, 'node.png').

  Prerequisites are that the Python module pydot_ng is installed (via pip) and that GraphViz has been installed from graphviz.org and its executable is in the system's path.

• API support for object detection using Fast R-CNN was added

  • See the description and code for the Fast R-CNN example.

• Tensorboard Event Generation

  • Initial support for generating Tensorboard events from the ProgressPrinter class was added, cf. here

Bug fixes

• Speed up TimesNodeBase for sparse by avoiding unroll. This improves speed of CrossEntropyWithSoftmax and ClassificationError for sparse labels. The language understanding example (LanguageUnderstanding.py) got four times faster now.

CNTK Evaluation library. NuGet package

A new set of NuGet Packages is provided with this Release.

IMPORTANT! In Visual Studio Manage Nuget Packages Window change the default option Stable Only to Include Prerelease. After selecting the appropriate NuGet package to install, use the version selector on the right to explicitly select package version 2.0.0-beta10.