Skip to content

CNTK 2.0 Beta Highlights

Jump to bottom
Alexey Orlov edited this page Mar 31, 2017 · 6 revisions

In this article we have summarized new features, changes, and improvements that were introduced to CNTK during multiple Beta Releases of CNTK v.2.0 (October 2016 - March 2017).

For the detailed information on all Beta releases, please see the correspondent Release Notes - you will find links at the CNTK Releases Page.

What's new in CNTK v.2

CNTK as a Library. Python, C++, and C# / .NET Managed API

Version 2 enables using CNTK as a software library with multiple APIs in addition to running CNTK executable and programming the Networks with BrainScript.

The following programming languages are supported:

  • Python
    • Python versions supported are 2.7, 3.4, and 3.5
  • C++
  • C#/.NET Managed

We have introduced a lot of features for each of these API and provided many examples illustrating their usage. Use the following resources:

Python Examples and Tutorials (Jupyter Notebooks)

Recognizing the importance of Python in Deep Learning we have prepared a set of Python Examples and Tutorials (the latter are implemented as Jupyter Notebooks). You will find all information at these locations:

You may also use CNTK Docker Containers to run Jupyter Notebooks tutorials.

CNTK Evaluation library

The CNTK Library API allows to evaluate both CNTK model-v1 and model-V2 format.

In CNTK v.2 the Evaluation Library can be used on Windows and Linux, in CPU and GPU configurations, and utilizing C++, Python, as well as C# and other .NET languages.

See more on CNTK Evaluation Library here.

CNTK new features

We have introduced a lot of new features during the v.2.0 Beta period. The list below highlights some of them:

  • Support of object recognition using Fast R-CNN algorithm.
  • Integration with NVIDIA NCCL, a stand-alone library of standard collective communication routines, such as all-gather, reduce, broadcast, etc., that have been optimized to achieve high bandwidth over PCIe. See how to enable NCCL in the CNTK Wiki.
  • Support of Asynchronous Stochastic Gradient Descent (ASGD)/Hogwild! training parallelization support using Microsoft’s Parameter Server (Project Multiverso).
  • Support of Distributed scenarios in Python API. See more in the sections on Distributed scenarios in ConvNet and ResNet examples.
  • Support for training on one-hot and sparse arrays via NumPy.
  • Lambda rank and NDCG at 1 are accessible from Python for real this time.
  • New Python and BrainScript for VGG16 and 19.
  • Performance Profiler for BrainScript and Python.
  • Support in training session for cross validation and preservation of all checkpoints.
  • Ability to write your own optimizer in Python by inheriting from UserLearner and overriding the update method.
  • Ability to implement Learners in Python using UserLearner Read more here.
  • All deserializers are exposed in C++.
    • HTK deserializers are also exposed in Python.
  • Support for TensorBoard output in both Python and BrainScript. Read more here.
  • Support for Model debugging, that can be done conveniently similarly to gdb/pdb, by wrapping the model with debug_model() and training/evaluating it.

We have also implemented a lot of performance improvements.Read more here.

CNTK Installation and Usage

New automated installation procedures

We have enabled different ways to install CNTK in an automated manner:

  • Installation scripts for binary distribution.
  • Installation scripts for the users who would like to work with CNTK source code.
  • Installation via pip install procedure.

Read more about different ways to install CNTK here.

CNTK at Docker Hub and as self-built Docker Images

CNTK is now available as Docker Images at Docker Hub.

You may also build your own Docker Images using pre-configured Docker files from CNTK Code base. See this Wiki page on using CNTK as Docker Images and Containers.

CNTK as NuGet packages

Windows developers using C++ and C# / .NET are welcome to use multiple CNTK NuGet packages supporting both v.1 and v.2 model formats.

See more on CNTK NuGet packages here.

New Documentation Site

Iteration Plans

Clone this wiki locally