umx.ncnn

This project is an ncnn/Eigen port of Open-Unmix-PyTorch, with code inspired by sevagh/umx.cpp.

Features

• Based on the ncnn inference framework, supporting CPU and Vulkan GPU inference
• Efficient matrix operations using Eigen
• Supports Open-Unmix LSTM models
• Cross-platform (Linux/macOS/Android)

Dependencies

• C++17
• ncnn
• Eigen3
• libnyquist (for audio I/O)

Patching ncnn

Before building, you need to apply the patches in vendor/patches to the ncnn source code:

cd vendor/ncnn
git apply ../../patches/*.patch

Build

git submodule update --init --recursive
mkdir build && cd build
cmake