A C library for automatic foward and backward differentiation.
Use RAD by dynamically allocating expressions, or RAD functions, typed rad_func *. There are several functions available for this, such as rad_const, rad_input, and rad_multiply.
Each function has inputs indexed by an unsigned int. For example, rad_input(n) creates a RAD function which outputs the input with index n.
The functions rad_eval, rad_forward_grad, rad_forward_diff, and rad_backward_diff are used for evaluating and computing derivatives of RAD functions, and they accept buffers for derivatives indexed by the inputs.
RAD comes with a built-in reference counter to assist with garbage collection.
All library functions except for rad_copy and rad_deep_copy consume each input RAD function, so a rad_func * value should not be reused after being passed as an argument.
By instead passing the output of rad_copy as an argument, the user can indicate to the library that they plan on continuing to use the RAD function.
rad_discard may be used to indicate that the user no longer needs a RAD function, and the library will free memory if there are no other references to the RAD function.
An example program neurons.c is included. In less than 100 lines, the program uses RAD to create a neural network which may be optimized using backpropogation.
The neural network is then optimized for 100000 epochs to evaluate XOR.
- Allow the user to specify their own functions to allocate memory. Currently, the library uses
malloc. - Fix an edge case involving compositions of rad functions. Currently, if
fandgarerad_func *, then differentiation of the functionrad_add(rad_copy(f), rad_composition(f, 1, g))will work incorrectly, whilerad_add(f, rad_composition(rad_deep_copy(f), 1, g))works correctly.