About

CNum is a CPU-optimized machine learning library for C++ that provides user-friendly ML model interfaces for training and inference, while exposing more low-level features that allow for control over thread management and memory allocation. CNum currently offers a Gradient Boosting model abstraction which can be trained with any type of tree booster to solve regression and classification problems. Currently the only type of tree booster the CNum library offers is an XGBoost-style tree booster. CNum also offers an extension for creating REST APIs for inference. The design principles behind the creation of CNum were maintaining numerical stability and model accuracy while optimizing for efficiency and cache-friendliness. These principles are demonstrated in CNum's use of deterministic random number generators and algorithms such as Kahan's summation algorithm which aid in numerical stability and reproducibility. They are also demonstrated in the use of row-wise features, and thread-local arena allocators for cache-friendliness. CNum is currently still in the early stages of development, but in the future will also contain deep learning tools like an autograd and tensor engine, neurons, and layers.