sadCpp

Simple Automatic Differentiation (in C++)

Expressions and Variables

There is one basic class: sad::Expression<numType>. The way instances are declared determine their function. Declaring an instance as

sad::Expression<double> x; 

declares a variable that, when evaluated (see later), returns a double. We should note that the default templte argument is double, so it may be omitted. Then, type-deduction allows us to skip all other template arguments.

Declaring an instance as

sad::Expression<double> c=2; 

declares a constant that, holds the value (double)2. Note that you can skip the template parameter here!

Declaring an instance through another Expression

sad::Expression<double> f=x+c; 

declares a general expression.

This library is all about manipulating such expressions!

Assignment

We can assign expressions to other expressions. Example:

sad::Expression<double> x; 
sad::Expression<double> y; 
sad::Expression f=x+y;

f=f*x; //fine. f becomes (x+y)*x

evaluate and derivative

There are two functions that the user needs to know; template<typename numType> sad::evaluateand template<typename numType> sad::derivative.

The first, sad::evaluate takes an instance of the Expression, and a std::map<unsigned int, numType> class and returns its value setting the variables at the values described in the map.

The sad::derivative function takes at least two instances of the Expression class and returns a new expression which is the derivative of the first with respect to the others. For example,

sad::Expression<double> x,y; 
sad::Expression f=x+y;

std::map at=sad::at(std::vector{x,y},std::vector{6,2}) ;

evaluate(f,at);//returns the value of x+y

derivative(f,x);// returns \partialf/\partialx 
derivative(f,x,y);// returns \partial^2f/\partialx \partialy 
derivative(f,x,y,y);// returns \partial^3f/\partialx \partialy \partialy 

If we differentiate with respect to an expression, we get an expression that has the value of $0$. That is, in the previous example, derivative(f,f); returns an instance of Expression with the value 0.

Note: you only need to explicitely use a template parameter for variables. All other instances can deduct the type since c++17.

Code generation

There are two function that can print expressions as functions or classes.

To-do:

• Implement as many functions as possible.
• Code generation for faster execution.
• Compute derivatives using backprop (and caching?).