microsoft/cntk

★ 17,597C++Audience · researcherComplexity · 5/5Setup · hard

Mindmap

mindmap
  root((CNTK))
    What it does
      Neural network training
      Distributed GPU
      Auto differentiation
    Model Types
      Convolutional networks
      LSTM recurrent
      Feed-forward
    APIs
      Python
      C++ C# Java
      Keras backend
    Status
      Final release v2.7
      Use ONNX Runtime

mindmap root((CNTK)) What it does Neural network training Distributed GPU Auto differentiation Model Types Convolutional networks LSTM recurrent Feed-forward APIs Python C++ C# Java Keras backend Status Final release v2.7 Use ONNX Runtime

Click or tap to explore — scroll the page freely

Things people build with this

USE CASE 1

Train convolutional neural networks for image classification across multiple GPUs in parallel.

USE CASE 2

Build and train LSTM recurrent networks for text or audio sequence tasks with automatic differentiation.

USE CASE 3

Use the Keras backend to write models in the familiar Keras API while running CNTK's distributed training.

Tech stack

C++PythonC#.NETJavaCUDAKerasONNX

Getting it running

Difficulty · hard Time to first run · 1day+

Requires CUDA-capable GPU and complex build setup, project is no longer maintained, use ONNX Runtime for model deployment.

In plain English

Microsoft's Cognitive Toolkit, called CNTK, is an open-source deep learning toolkit. Deep learning is the branch of AI that uses neural networks, software loosely inspired by how brain cells connect, to learn patterns from data. CNTK gives engineers and researchers a way to design, train, and run those networks. The toolkit describes a neural network as a directed graph: a flowchart where input data flows through a series of mathematical steps and produces a result at the other end. Each node in the graph either holds input values, holds parameters that the network learns, or performs a matrix operation on whatever comes in. From this building block, CNTK lets people combine common model types, including feed-forward networks, convolutional networks often used for images, and recurrent networks including LSTMs often used for sequences such as text or audio. Training uses stochastic gradient descent, a standard method that nudges parameters in small steps to reduce mistakes, with automatic differentiation to compute those adjustments. It can spread the work across multiple GPUs and multiple servers in parallel. You would reach for CNTK if you were building a deep learning model and wanted a framework with parallel, distributed training. APIs are available in Python, C++, C#/.NET, and Java, and there is a backend for Keras. One important note from the project itself: the README states that the 2.7 release is the last main release of CNTK, that there are no plans for new feature development beyond it, and that the team recommends operationalizing models through ONNX and the ONNX Runtime instead. The full README is longer than what was provided.

Copy-paste prompts

Prompt 1

Show me how to define and train a simple LSTM network for text classification using the CNTK Python API.

Prompt 2

How do I configure CNTK to distribute training of a convolutional network across two GPUs on one server?

Prompt 3

I have a trained CNTK model and want to export it to ONNX format for deployment with ONNX Runtime, show me the Python steps.

Prompt 4

Compare how to define the same feed-forward network in CNTK Python vs the Keras backend, which one is simpler for my use case?

Open on GitHub → Explain another repo

← microsoft on gitmyhub — every repo by this author, as a profile.

Verify against the repo before relying on details.