Here demonstrate the scatter error swaps and column swaps.
More generally need subsasgn(), if that is the function, to work
with permutations (including sub-sampled).
u=AutoDiff(reshape(1:3,3,[]))
f=u.^2
fv=getvalue(f), fD=full(getderivs(f))
% test rows swaps
swap=[1 3 2]
g=f(swap,:) % gather works for vals and derivs
gv=getvalue(g), gD=full(getderivs(g))
g(swap,:)=f % scatter DNW for derivs, xxxxxxxxxxxxxx
gv=getvalue(g), gD=full(getderivs(g))
if size(u,2)>1 %optionally test column swaps
swap=[2 1]
g=f(:,swap) % gather works for vals and derivs
gv=getvalue(g), gD=full(getderivs(g))
g(:,swap)=f % scatter DNW for derivs, xxxxxxxxxxxxx
gv=getvalue(g), gD=full(getderivs(g))
end
Checking row case: just try simple swap of rows 2 and 3.
The first case where LHS = RHS-gather works fine with the
rows of the derivs swapping correctly. The 2nd case where
LHS-gather=RHS does not transform the derivs correctly.
Same for the column swap example.
Here demonstrate the scatter error swaps and column swaps.
More generally need subsasgn(), if that is the function, to work
with permutations (including sub-sampled).
Checking row case: just try simple swap of rows 2 and 3.
The first case where LHS = RHS-gather works fine with the
rows of the derivs swapping correctly. The 2nd case where
LHS-gather=RHS does not transform the derivs correctly.
Same for the column swap example.