Microsoft Cognitive Toolkit (CNTK), an open source deep-learning toolkit


Chat Windows build status Linux build status
Join the chat at Build Status Build Status

The Microsoft Cognitive Toolkit ( is a unified deep learning toolkit that describes neural networks as a series of computational steps via a directed graph. In this directed graph, leaf nodes represent input values or network parameters, while other nodes represent matrix operations upon their inputs. CNTK allows users to easily realize and combine popular model types such as feed-forward DNNs, convolutional nets (CNNs), and recurrent networks (RNNs/LSTMs). It implements stochastic gradient descent (SGD, error backpropagation) learning with automatic differentiation and parallelization across multiple GPUs and servers. CNTK has been available under an open-source license since April 2015. It is our hope that the community will take advantage of CNTK to share ideas more quickly through the exchange of open source working code.


Installing nightly packages

If you prefer to use latest CNTK bits from master, use one of the CNTK nightly packages:

Learning CNTK

You can learn more about using and contributing to CNTK with the following resources:

More information


CNTK is in active use at Microsoft and constantly evolving. There will be bugs.

Microsoft Open Source Code of Conduct

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact with any additional questions or comments.


You can find more news on the official project feed

Project changelog

2018-04-16. CNTK 2.5.1

Repack CNTK 2.5 with third party libraries included in the bundles (Python wheel packages)

2018-03-15. CNTK 2.5

Change profiler details output format to be chrome://tracing

Enable per-node timing. Working example here

  • per-node timing creates items in profiler details when profiler is enabled.
  • usage in Python:
import cntk as C
C.debugging.start_profiler() # optional
C.debugging.enable_profiler() # optional
#<trainer|evaluator|function> executions

Example profiler details view in chrome://tracing ProfilerDetailWithNodeTiming

CPU inference performance improvements using MKL

  • Accelerates some common tensor ops in Intel CPU inference for float32, especially for fully connected networks
  • Can be turned on/off by cntk.cntk_py.enable_cpueval_optimization()/cntk.cntk_py.disable_cpueval_optimization()

1BitSGD incorporated into CNTK

  • 1BitSGD source code is now available with CNTK license (MIT license) under Source/1BitSGD/
  • 1bitsgd build target was merged into existing gpu target

New loss function: hierarchical softmax

  • Thanks @yaochengji for the contribution!

Distributed Training with Mulitple Learners

  • Trainer now accepts multiple parameter learners for distributed training. With this change, different parameters of a network can be learned by different learners in a single training session. This also facilitates distributed training for GANs. For more information, please refer to the and the


  • Added MeanVarianceNormalization operator.

Bug fixes

  • Fixed convergence issue in Tutorial 201B
  • Fixed pooling/unpooling to support free dimension for sequences
  • Fixed crash in CNTKBinaryFormat deserializer when crossing sweep boundary
  • Fixed shape inference bug in RNN step function for scalar broadcasting
  • Fixed a build bug when mpi=no
  • Improved distributed training aggregation speed by increasing packing threshold, and expose the knob in V2
  • Fixed a memory leak in MKL layout
  • Fixed a bug in cntk.convert API in, which prevents converting complex networks.


  • Updates
    • CNTK exported ONNX models are now ONNX.checker compliant.
    • Added ONNX support for CNTK’s OptimizedRNNStack operator (LSTM only).
    • Added support for LSTM and GRU operators
    • Added support for experimental ONNX op MeanVarianceNormalization.
    • Added support for experimental ONNX op Identity.
    • Added support for exporting CNTK’s LayerNormalization layer using ONNX MeanVarianceNormalization op.
  • Bug or minor fixes:
    • Axis attribute is optional in CNTK’s ONNX Concat operator.
    • Bug fix in ONNX broadcasting for scalars.
    • Bug fix in ONNX ConvTranspose operator.
    • Backward compatibility bug fix in LeakyReLu (argument ‘alpha’ reverted to type double).


  • Added a new API find_by_uid() under cntk.logging.graph.

2018-02-28. CNTK supports nightly build

If you prefer to use latest CNTK bits from master, use one of the CNTK nightly package.

Alternatively, you can also click corresponding build badge to land to nightly build page.

2018-01-31. CNTK 2.4


  • Moved to CUDA9, cuDNN 7 and Visual Studio 2017.
  • Removed Python 3.4 support.
  • Added Volta GPU and FP16 support.
  • Better ONNX support.
  • CPU perf improvement.
  • More OPs.


  • top_k operation: in the forward pass it computes the top (largest) k values and corresponding indices along the specified axis. In the backward pass the gradient is scattered to the top k elements (an element not in the top k gets a zero gradient).
  • gather operation now supports an axis argument
  • squeeze and expand_dims operations for easily removing and adding singleton axes
  • zeros_like and ones_like operations. In many situations you can just rely on CNTK correctly broadcasting a simple 0 or 1 but sometimes you need the actual tensor.
  • depth_to_space: Rearranges elements in the input tensor from the depth dimension into spatial blocks. Typical use of this operation is for implementing sub-pixel convolution for some image super-resolution models.
  • space_to_depth: Rearranges elements in the input tensor from the spatial dimensions to the depth dimension. It is largely the inverse of DepthToSpace.
  • sum operation: Create a new Function instance that computes element-wise sum of input tensors.
  • softsign operation: Create a new Function instance that computes the element-wise softsign of a input tensor.
  • asinh operation: Create a new Function instance that computes the element-wise asinh of a input tensor.
  • log_softmax operation: Create a new Function instance that computes the logsoftmax normalized values of a input tensor.
  • hard_sigmoid operation: Create a new Function instance that computes the hard_sigmoid normalized values of a input tensor.
  • element_and, element_not, element_or, element_xor element-wise logic operations
  • reduce_l1 operation: Computes the L1 norm of the input tensor's element along the provided axes.
  • reduce_l2 operation: Computes the L2 norm of the input tensor's element along the provided axes.
  • reduce_sum_square operation: Computes the sum square of the input tensor's element along the provided axes.
  • image_scaler operation: Alteration of image by scaling its individual values.


  • There have been several improvements to ONNX support in CNTK.
  • Updates
    • Updated ONNX Reshape op to handle InferredDimension.
    • Adding producer_name and producer_version fields to ONNX models.
    • Handling the case when neither auto_pad nor pads atrribute is specified in ONNX Conv op.
  • Bug fixes
    • Fixed bug in ONNX Pooling op serialization
    • Bug fix to create ONNX InputVariable with only one batch axis.
    • Bug fixes and updates to implementation of ONNX Transpose op to match updated spec.
    • Bug fixes and updates to implementation of ONNX Conv, ConvTranspose, and Pooling ops to match updated spec.


  • Group convolution
    • Fixed bug in group convolution. Output of CNTK Convolution op will change for groups > 1. More optimized implementation of group convolution is expected in the next release.
    • Better error reporting for group convolution in Convolution layer.

Halide Binary Convolution

  • The CNTK build can now use optional Halide libraries to build library that can be used with the netopt module. The library contains optimized binary convolution operators that perform better than the python based binarized convolution operators. To enable Halide in the build, please download Halide release and set HALIDE_PATH environment varibale before starting a build. In Linux, you can use ./configure --with-halide[=directory] to enable it. For more information on how to use this feature, please refer to How_to_use_network_optimization.

See more in the Release Notes. Get the Release from the CNTK Releases page.

2018-01-22. CNTK support for CUDA 9

CNTK now supports CUDA 9/cuDNN 7. This requires an update to build environment to Ubuntu 16/GCC 5 for Linux, and Visual Studio 2017/VCTools 14.11 for Windows. With CUDA 9, CNTK also added a preview for 16-bit floating point (a.k.a FP16) computation.

Please check out the example of FP16 in ResNet50 here

Notes on FP16 preview:

  • FP16 implementation on CPU is not optimized, and it's not supposed to be used in CPU inference directly. User needs to convert the model to 32-bit floating point before running on CPU.
  • Loss/Criterion for FP16 training needs to be 32bit for accumulation without overflow, using cast function. Please check the example above.
  • Readers do not have FP16 output unless using numpy to feed data, cast from FP32 to FP16 is needed. Please check the example above.
  • FP16 gradient aggregation is currently only implemented on GPU using NCCL2. Distributed training with FP16 with MPI is not supported.
  • FP16 math is a subset of current FP32 implementation. Some model may get Feature Not Implemented exception using FP16.
  • FP16 is currently not supported in BrainScript. Please use Python for FP16.

To setup build and runtime environment on Windows:

  • Install Visual Studio 2017 with following workloads and components. From command line (use Community version installer as example): vs_community.exe --add Microsoft.VisualStudio.Workload.NativeDesktop --add Microsoft.VisualStudio.Workload.ManagedDesktop --add Microsoft.VisualStudio.Workload.Universal --add Microsoft.Component.PythonTools --add Microsoft.VisualStudio.Component.VC.Tools.14.11
  • Install NVidia CUDA 9
  • From PowerShell, run: DevInstall.ps1
  • Start VCTools 14.11 command line, run: cmd /k "%VS2017INSTALLDIR%\VC\Auxiliary\Build\vcvarsall.bat" x64 --vcvars_ver=14.11
  • Open CNTK.sln from the VCTools 14.11 command line. Note that starting CNTK.sln other than VCTools 14.11 command line, would causes CUDA 9 build error.

To setup build and runtime environment on Linux using docker, please build Unbuntu 16.04 docker image using Dockerfiles here. For other Linux systems, please refer to the Dockerfiles to setup dependent libraries for CNTK.

2017-12-05. CNTK 2.3.1 Release of Cognitive Toolkit v.2.3.1.

CNTK support for ONNX format is now out of preview mode.

If you want to try ONNX, you can build from master or pip install one of the below wheels that matches your Python environment.

For Windows CPU-Only:

For Windows GPU:

Linux CPU-Only:

Linux GPU:

You can also try one of the below NuGet package.

  • 最近提交:09-15
  • 创建时间:2015-11-26