Attempt to create simple to use AD mechanism, which could be easily integrated into existing code, formulas and algorithms, essentially making them differentiable. It also contains basic Gradient Descent implementation, essentially turning library into non-linear problem solver.
Originally it was developed as part of another Tiny PGA project, however it was extracted as being useful in other projects too.
Assume you have an equation (or function), depending on one or more variables, for instance something like this:
float function(float x, float y)
{
return 0.5F * x * x + 2.0F * x * y + 4.F * sin(y);
}
...
float result = function(0.1F, 0.2F);
You may turn it into a differentiable version by redefining numerical type like that:
using Float = tiny_autodf::AutoDf<float>;
template<typename type>
type function(type x, type y)
{
return 0.5F * x * x + 2.0F * x * y + 4.F * sin(y);
}
Float x = 0.1F;
Float y = 0.2F;
Float result = function<Float>(x, y);
Now your result is "differentiable" over the x and 'y'.
In order to obtain gradient in current point (0.1, 0.2) you'd need to call eval() member function:
auto eval = result.eval();
std::cout << "dx = " << eval.derivatives[x.ID()] << ", dy = " << eval.derivatives[y.ID()] << std::endl;