microsoft/CNTK

Represents neural networks as directed acyclic graphs of computational operations, enabling automatic differentiation and efficient backpropagation across distributed GPU clusters. Supports multiple architectures including DNNs, CNNs, and RNNs/LSTMs with unified training via stochastic gradient descent. Integrates with Keras as a backend and exports models to ONNX format for cross-framework interoperability.

17,610 stars. No commits in the last 6 months.