Introducing the ModularML Framework - A transparent and modular machine learning framework made as a tool for research and education

Abstract

This thesis explores the process of creating a highly modular machine learning framework in C++, without performance compromises. The framework can parse ONNX models into dynamic C++ objects, modify the implementation of core computational functions (like GEMM), and use the models to perform inference. The framework reproduces results achieved in peer frameworks like PyTorch and TensorFlow. The usefulness of this framework stems from its pure C++ implementation, with no API layers to compiled modules or other black-box functionality. This makes it highly suitable for use in education and research, where debuggability, modularity, and ease of use are paramount.