void Gradient::apply_bwd (const ExprApply& a, ExprLabel** x, const ExprLabel& y) {
Array<Domain> d(a.nb_args);
Array<Domain> g(a.nb_args);
int n=0;
for (int i=0; i<a.nb_args; i++) {
d.set_ref(i,*x[i]->d);
g.set_ref(i,*x[i]->g);
n+=x[i]->d->dim.size();
}
// we unvectorize the components of the gradient.
IntervalVector old_g(n);
load(old_g, g);
IntervalVector tmp_g(n);
if (a.func.expr().dim.is_scalar()) {
gradient(a.func,d,tmp_g);
//cout << "tmp-g=" << tmp_g << endl;
tmp_g *= y.g->i(); // pre-multiplication by y.g
tmp_g += old_g; // addition to the old value of g
load(g,tmp_g);
} else {
assert(a.func.expr().dim.is_vector()); // matrix-valued function not implemented...
int m=a.func.expr().dim.vec_size();
IntervalMatrix J(m,n);
jacobian(a.func,d,J);
tmp_g = y.g->v()*J; // pre-multiplication by y.g
tmp_g += old_g;
load(g,tmp_g);
}
}
void Gradient::apply_bwd(int* x, int y) {
const ExprApply& a = (const ExprApply&) f.node(y);
Array<Domain> d2(a.func.nb_arg());
Array<Domain> g2(a.nb_args);
int n=0;
for (int i=0; i<a.func.nb_arg(); i++) {
d2.set_ref(i,d[x[i]]);
g2.set_ref(i,g[x[i]]);
n+=d[x[i]].dim.size();
}
// we unvectorize the components of the gradient.
IntervalVector old_g(n);
load(old_g, g2);
IntervalVector tmp_g(n);
if (a.func.expr().dim.is_scalar()) {
a.func.deriv_calculator().gradient(d2,tmp_g);
//cout << "tmp-g=" << tmp_g << endl;
tmp_g *= g[y].i(); // pre-multiplication by y.g
tmp_g += old_g; // addition to the old value of g
load(g2,tmp_g);
} else {
if (!a.func.expr().dim.is_vector())
not_implemented("automatic differentiation of matrix-valued function");
int m=a.func.expr().dim.vec_size();
IntervalMatrix J(m,n);
a.func.deriv_calculator().jacobian(d2,J);
tmp_g = g[y].v()*J; // pre-multiplication by y.g
tmp_g += old_g;
load(g2,tmp_g);
}
}
