NextCloud/dlib
1
0
Fork 0
mirror of https://github.com/davisking/dlib.git synced 2024-11-01 10:14:53 +08:00
Code Issues Packages Projects Releases Wiki Activity

Changed svm_c_linear_dcd_trainer so that when the user forces the last weight

Browse Source 
to 1 it automatically disables use of the bias since that is almost certainly
what the user will want to do.
This commit is contained in:
Davis King 2013-01-03 22:02:16 -05:00
parent b445ddbd8d
commit 96264d3339
2 changed files with 10 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
dlib/svm/svm_c_linear_dcd_trainer.h
View File

@ -284,7 +284,7 @@ namespace dlib
);
// extend w by the right number of elements
if (have_bias)
if (have_bias && !last_weight_1)
{
// Splice some zeros into the w vector so it will have the
// right length. Here we are being careful to move the bias
@ -316,7 +316,7 @@ namespace dlib
index.reserve(x.size());
Q.reserve(x.size());
if (have_bias)
if (have_bias && !last_weight_1)
w.set_size(dims+1);
else
w.set_size(dims);
@ -328,7 +328,7 @@ namespace dlib
{
Q.push_back(length_squared(x(i)));
if (have_bias)
if (have_bias && !last_weight_1)
{
index.push_back(i);
Q.back() += 1;
@ -570,7 +570,7 @@ namespace dlib
alpha[i] = std::min(std::max(alpha[i] - G/state.Q[i], (scalar_type)0.0), C);
const scalar_type delta = (alpha[i]-alpha_old)*y(i);
add_to(w, x(i), delta);
if (have_bias)
if (have_bias && !last_weight_1)
w(w.size()-1) -= delta;
if (last_weight_1)
@ -618,7 +618,7 @@ namespace dlib
// put the solution into a decision function and then return it
decision_function<kernel_type> df;
if (have_bias)
if (have_bias && !last_weight_1)
df.b = w(w.size()-1);
else
df.b = 0;
@ -639,7 +639,7 @@ namespace dlib
const sample_type& sample
) const
{
if (have_bias)
if (have_bias && !last_weight_1)
{
const long w_size_m1 = w.size()-1;
return dlib::dot(colm(w,0,w_size_m1), sample) - w(w_size_m1);

3
dlib/svm/svm_c_linear_dcd_trainer_abstract.h
View File

@ -111,6 +111,9 @@ namespace dlib
- returns true if this trainer has the constraint that the last weight in
the learned parameter vector must be 1. This is the weight corresponding
to the feature in the training vectors with the highest dimension.
- Forcing the last weight to 1 also disables the bias and therefore the b
field of the learned decision_function will be 0 when forces_last_weight_to_1() == true.
This is true regardless of the setting of #include_bias().
!*/
void force_last_weight_to_1 (