Zhetapi is a modern C++ machine learning and numerical analysis library with an emphasis on intuitive usage.

Linear Algebra

Zhetapi provides a basic linear algebra interface using C++ 20 concepts. The Field structure enforces a contraint on types to ensure they behave like algebraic fields. As a result these structures support basic arithmetic along with other useful methods.

For now, the following template classes are provided: Tensor, Matrix, Vector.

Auto Differentiation

The auto differentiation facilities in Zhetapi belong in the zhetapi::autograd namespace. All operations that depend on autodiff use floats as the underlying type; in particular Constant is a Tensor <float> and is the basis of all numerical values in this module.

To provide a seamless, operator based interface into the autodiff facilities, two notable classes are provided, Variable and Function. As one can expect, Variables can store arbitrary Constant values, and Functions are compositions of Variables under varying operations. For example:

Function f = x + y;
Function g = x * y;

// f and g are now functions of *two* variables
Constant a = f(1, 2);
Constant b = g(1, 2);

// Composition of functions is done likewise
Function h = f(x, g(x, y)); // NOTE: h is still a function of two variables

Symbolic Differentiation

Backward Pass

Currently, only backward mode is enabled for autograd.

Building

Zhetapi is primarily a header-only library, but for now there are some examples that one can play around with in the experimental directory.

This project is developed using C++ 20. Additional dependenies include PNG (libpng-dev on Ubuntu systems), OpenMP (Optional) and CUDA (Optional).

Generate the build configuration using CMake as follows:

$ cd zhetapi
$ mkdir build && cd build
$ cmake -DZHETAPI_ENABLE_CUDA=<ON|OFF> # ON by default

And build the targets as one would usually do (e.g. make or ninja).