regular-dev/nevermind-neu

Machine learning library written in rust with OpenCL support

nevermind-neu

Machine learning library and tool with terminal user interface written in rust. It supports OpenCL layers and CPU layers. Core math matrix library is ndarray which use matrixmultiply crate for CPU matrix multiplication.

Design goals

• Fast optimized computations
• User-friendly API
• Provide utility terminal application to create, train, manage models with user-friendly terminal interface

OpenCL

OpenCL support is based on ocl crate. It is optional feature and enabled by default.

Features

• FullyConnected layer
• Euclidean Loss, Softmax Loss
• Optimizers: Adam, RMSProp, AdaGrad, AdaDelta
• Async parallel data loading
• Protobuf, CSV dataloaders
• (De)Serializing neural network state to protobuf
• (De)Serializing neural network configuration net yaml file
• Activation functions : sigmoid, tanh, relu, leaky_relu

Terminal user interface tool

Examples

XOR function approximation

cargo run --release --example xor

XOR function approximation with OpenCL support

cargo run --release --example xor_ocl

For MNIST digits classification example read MNIST example readme

Roadmap

• Conv2D layer
• RNN + LSTM
• Residual block
• OpenCL optimization

License

Apache License Version 2.0