Commit 28c2dd2c authored by Davis King's avatar Davis King

Changed code for the LM/quasi-newton model around a little to avoid

repeated calculation of things and also added some checks for division
by zero.

--HG--
extra : convert_revision : svn%3Afdd8eb12-d10e-0410-9acb-85c331704f74/trunk%403948
parent 1f73eca7
......@@ -30,6 +30,8 @@ namespace dlib
S = 0;
last_f = 0;
last_f2 = 0;
r.set_size(list.size(),1);
}
const funct_type& f;
......@@ -53,6 +55,8 @@ namespace dlib
for (long i = 0; i < list.size(); ++i)
{
const type temp = f(list(i), x);
// save the residual for later
r(i) = temp;
result += temp*temp;
}
......@@ -66,16 +70,15 @@ namespace dlib
general_matrix& h
) const
{
matrix<type,0,NR,mem_manager_type,layout_type> J(list.size(), x.size());
matrix<type,0,1,mem_manager_type,layout_type> r(list.size(),1);
J.set_size(list.size(), x.size());
// compute the Jacobian
for (long i = 0; i < list.size(); ++i)
{
r(i) = f(list(i), x);
set_rowm(J,i) = trans(der(list(i), x));
}
// Compute the Levenberg-Marquardt gradient and hessian
d = trans(J)*r;
h = trans(J)*J;
......@@ -85,21 +88,26 @@ namespace dlib
S = 0;
}
// If this isn't the first iteration then check if using
// a quasi-newton update helps things.
if (last_r.size() != 0)
{
column_vector s, y, yy, temp;
s = x - last_x;
y = trans(J)*r - trans(last_J)*last_r;
yy = trans(J)*r - trans(last_J)*r;
y = d - last_d;
yy = d - trans(last_J)*r;
const type ys = trans(y)*s;
temp = yy - S*s;
type scale = std::min<type>(1, std::abs(dot(s,yy))/std::abs(trans(s)*S*s));
vtemp = yy - S*s;
const type temp2 = std::abs(trans(s)*S*s);
type scale = (temp2 != 0) ? std::min<type>(1, std::abs(dot(s,yy))/temp2) : 1;
S = scale*S + (temp*trans(y) + y*trans(temp))/(ys) - dot(temp,s)/ys/ys*y*trans(y);
if (ys != 0)
S = scale*S + (vtemp*trans(y) + y*trans(vtemp))/(ys) - dot(vtemp,s)/ys/ys*y*trans(y);
else
S *= scale;
// check how well both the models fit the last change in f we saw
// check how well both the models fit the last change we saw in f()
const type measured_delta = last_f2 - last_f;
s = -s;
const type h_predicted_delta = 0.5*trans(s)*h*s + trans(d)*s;
......@@ -116,22 +124,35 @@ namespace dlib
{
S = 0;
}
// put r into last_r
r.swap(last_r);
}
else
{
// put r into last_r
last_r = r;
}
last_r = r;
last_J = J;
J.swap(last_J);
last_x = x;
last_d = d;
last_f2 = last_f;
}
mutable type last_f; // value of function we saw in last operator()
mutable type last_f2; // value of last_f we saw in get_derivative_and_hessian()
mutable matrix<type,0,1,mem_manager_type,layout_type> r;
mutable column_vector vtemp;
mutable column_vector s, y, yy;
mutable general_matrix S;
mutable column_vector last_x;
mutable column_vector last_d;
mutable matrix<type,0,1,mem_manager_type,layout_type> last_r;
mutable matrix<type,0,NR,mem_manager_type,layout_type> last_J;
mutable matrix<type,0,NR,mem_manager_type,layout_type> J;
};
// ----------------------------------------------------------------------------------------
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment