deshanadesai / acorns Goto Github PK

Dear devs,

recently I have been trying to differentiate complicated expressions that eventually result in nested calls to pow. I find that the forward mode derivative for these cases is not taken correctly. Here is the MWE:

c_function = """
function_test(double x){
double x2 = pow(x,2);
double H = pow(x2,2);
}

deriv = acorns.autodiff(c_function, 'H', ['x'], func = 'function_test', output_filename = 'test',
     output_func = 'compute_grad_nested_pow')

This gives as the derivative (pow(pow,(2-1)) * (2 * 0 + pow * 0 * log(pow))) which looks like it's using pow as a variable instead of a function.

Computing directly the derivative of

c_function = """
function_test(double x){
double H = pow(x,4);
}

gives the correct result, (pow(x,(4-1)) * (4 * 1 + x * 0 * log(x))). Unfortunately my expressions are complicated enough that no such simplifications are possible.

Dear devs,

First, thanks for the useful library! It hits the perfect sweet spot for the problem I am considering.

I have found that differentiating an expression inside a pow call does not work correctly. Here is a MWE:

c_function = "int function_test(double r, double L){ \
    double blah = pow(L+r,0.5); \
    return 0; \
}"

acorns.autodiff(c_function, 'blah', ['r','L'], func = 'function_test', output_filename = 'test_grad_back',
       output_func = 'compute_grad_back')

This gives as the derivatives:

void compute_grad_back(double values[], int num_points, double ders[]){

	for(int i = 0; i < num_points; ++i)
	{
		double r = values[i* 2 + 0 ];
		double L = values[i* 2 + 1 ];
		ders[i*2+0]= (pow((L + r),(0.5-1)) * (0.5 * 0 + (L + r) * 0 * log((L + r)))); // df/(r) 
		ders[i*2+1]= (pow((L + r),(0.5-1)) * (0.5 * 0 + (L + r) * 0 * log((L + r)))); // df/(L) 
	}
}

i.e. zero for both derivatives, which is of course not the correct answer. It appears that the issue is with the ._forward_diff of the base.

Another issues is when the base of the pow call is negative. Then the current implementation gives nan even when the answer should be sensible, e.g. when differentiating pow(pr,4) where pr is -1e-4.

Could you please fix the license?

It is missing the year and authors.
After that, could you please add the license to be package along with the source distribution?