void CLibLinear::solve_l1r_lr(
const problem *prob_col, double eps,
double Cp, double Cn)
{
int l = prob_col->l;
int w_size = prob_col->n;
int j, s, iter = 0;
int active_size = w_size;
int max_num_linesearch = 20;
double x_min = 0;
double sigma = 0.01;
double d, G, H;
double Gmax_old = CMath::INFTY;
double Gmax_new;
double Gmax_init=0;
double sum1, appxcond1;
double sum2, appxcond2;
double cond;
int *index = SG_MALLOC(int, w_size);
int32_t *y = SG_MALLOC(int32_t, l);
double *exp_wTx = SG_MALLOC(double, l);
double *exp_wTx_new = SG_MALLOC(double, l);
double *xj_max = SG_MALLOC(double, w_size);
double *C_sum = SG_MALLOC(double, w_size);
double *xjneg_sum = SG_MALLOC(double, w_size);
double *xjpos_sum = SG_MALLOC(double, w_size);
CDotFeatures* x = prob_col->x;
void* iterator;
int ind;
double val;
double C[3] = {Cn,0,Cp};
int n = prob_col->n;
if (prob_col->use_bias)
n--;
for(j=0; j<l; j++)
{
exp_wTx[j] = 1;
if(prob_col->y[j] > 0)
y[j] = 1;
else
y[j] = -1;
}
for(j=0; j<w_size; j++)
{
w.vector[j] = 0;
index[j] = j;
xj_max[j] = 0;
C_sum[j] = 0;
xjneg_sum[j] = 0;
xjpos_sum[j] = 0;
if (use_bias && j==n)
{
for (ind=0; ind<l; ind++)
{
x_min = CMath::min(x_min, 1.0);
xj_max[j] = CMath::max(xj_max[j], 1.0);
C_sum[j] += C[GETI(ind)];
if(y[ind] == -1)
xjneg_sum[j] += C[GETI(ind)];
else
xjpos_sum[j] += C[GETI(ind)];
}
}
else
{
iterator=x->get_feature_iterator(j);
while (x->get_next_feature(ind, val, iterator))
{
x_min = CMath::min(x_min, val);
xj_max[j] = CMath::max(xj_max[j], val);
C_sum[j] += C[GETI(ind)];
if(y[ind] == -1)
xjneg_sum[j] += C[GETI(ind)]*val;
else
xjpos_sum[j] += C[GETI(ind)]*val;
}
x->free_feature_iterator(iterator);
}
}
CTime start_time;
while (iter < max_iterations && !CSignal::cancel_computations())
{
if (m_max_train_time > 0 && start_time.cur_time_diff() > m_max_train_time)
break;
Gmax_new = 0;
for(j=0; j<active_size; j++)
{
int i = j+rand()%(active_size-j);
CMath::swap(index[i], index[j]);
}
for(s=0; s<active_size; s++)
{
j = index[s];
sum1 = 0;
sum2 = 0;
H = 0;
if (use_bias && j==n)
{
for (ind=0; ind<l; ind++)
{
double exp_wTxind = exp_wTx[ind];
double tmp1 = 1.0/(1+exp_wTxind);
double tmp2 = C[GETI(ind)]*tmp1;
double tmp3 = tmp2*exp_wTxind;
sum2 += tmp2;
sum1 += tmp3;
H += tmp1*tmp3;
}
}
else
{
iterator=x->get_feature_iterator(j);
while (x->get_next_feature(ind, val, iterator))
{
double exp_wTxind = exp_wTx[ind];
double tmp1 = val/(1+exp_wTxind);
double tmp2 = C[GETI(ind)]*tmp1;
double tmp3 = tmp2*exp_wTxind;
sum2 += tmp2;
sum1 += tmp3;
H += tmp1*tmp3;
}
x->free_feature_iterator(iterator);
}
G = -sum2 + xjneg_sum[j];
double Gp = G+1;
double Gn = G-1;
double violation = 0;
if(w.vector[j] == 0)
{
if(Gp < 0)
violation = -Gp;
else if(Gn > 0)
violation = Gn;
else if(Gp>Gmax_old/l && Gn<-Gmax_old/l)
{
active_size--;
CMath::swap(index[s], index[active_size]);
s--;
continue;
}
}
else if(w.vector[j] > 0)
violation = fabs(Gp);
else
violation = fabs(Gn);
Gmax_new = CMath::max(Gmax_new, violation);
// obtain Newton direction d
if(Gp <= H*w.vector[j])
d = -Gp/H;
else if(Gn >= H*w.vector[j])
d = -Gn/H;
else
d = -w.vector[j];
if(fabs(d) < 1.0e-12)
continue;
d = CMath::min(CMath::max(d,-10.0),10.0);
double delta = fabs(w.vector[j]+d)-fabs(w.vector[j]) + G*d;
int num_linesearch;
for(num_linesearch=0; num_linesearch < max_num_linesearch; num_linesearch++)
{
cond = fabs(w.vector[j]+d)-fabs(w.vector[j]) - sigma*delta;
if(x_min >= 0)
{
double tmp = exp(d*xj_max[j]);
appxcond1 = log(1+sum1*(tmp-1)/xj_max[j]/C_sum[j])*C_sum[j] + cond - d*xjpos_sum[j];
appxcond2 = log(1+sum2*(1/tmp-1)/xj_max[j]/C_sum[j])*C_sum[j] + cond + d*xjneg_sum[j];
if(CMath::min(appxcond1,appxcond2) <= 0)
{
if (use_bias && j==n)
{
for (ind=0; ind<l; ind++)
exp_wTx[ind] *= exp(d);
}
else
{
iterator=x->get_feature_iterator(j);
while (x->get_next_feature(ind, val, iterator))
exp_wTx[ind] *= exp(d*val);
x->free_feature_iterator(iterator);
}
break;
}
}
cond += d*xjneg_sum[j];
int i = 0;
if (use_bias && j==n)
{
for (ind=0; ind<l; ind++)
{
double exp_dx = exp(d);
exp_wTx_new[i] = exp_wTx[ind]*exp_dx;
cond += C[GETI(ind)]*log((1+exp_wTx_new[i])/(exp_dx+exp_wTx_new[i]));
i++;
}
}
else
{
iterator=x->get_feature_iterator(j);
while (x->get_next_feature(ind, val, iterator))
{
double exp_dx = exp(d*val);
exp_wTx_new[i] = exp_wTx[ind]*exp_dx;
cond += C[GETI(ind)]*log((1+exp_wTx_new[i])/(exp_dx+exp_wTx_new[i]));
i++;
}
x->free_feature_iterator(iterator);
}
if(cond <= 0)
{
i = 0;
if (use_bias && j==n)
{
for (ind=0; ind<l; ind++)
{
exp_wTx[ind] = exp_wTx_new[i];
i++;
}
}
else
{
iterator=x->get_feature_iterator(j);
while (x->get_next_feature(ind, val, iterator))
{
exp_wTx[ind] = exp_wTx_new[i];
i++;
}
x->free_feature_iterator(iterator);
}
break;
}
else
{
d *= 0.5;
delta *= 0.5;
}
}
w.vector[j] += d;
// recompute exp_wTx[] if line search takes too many steps
if(num_linesearch >= max_num_linesearch)
{
SG_INFO("#");
for(int i=0; i<l; i++)
exp_wTx[i] = 0;
for(int i=0; i<w_size; i++)
{
if(w.vector[i]==0) continue;
if (use_bias && i==n)
{
for (ind=0; ind<l; ind++)
exp_wTx[ind] += w.vector[i];
}
else
{
iterator=x->get_feature_iterator(i);
while (x->get_next_feature(ind, val, iterator))
exp_wTx[ind] += w.vector[i]*val;
x->free_feature_iterator(iterator);
}
}
for(int i=0; i<l; i++)
exp_wTx[i] = exp(exp_wTx[i]);
}
}
if(iter == 0)
Gmax_init = Gmax_new;
iter++;
SG_SABS_PROGRESS(Gmax_new, -CMath::log10(Gmax_new), -CMath::log10(Gmax_init), -CMath::log10(eps*Gmax_init), 6);
if(Gmax_new <= eps*Gmax_init)
{
if(active_size == w_size)
break;
else
{
active_size = w_size;
Gmax_old = CMath::INFTY;
continue;
}
}
Gmax_old = Gmax_new;
}
SG_DONE();
SG_INFO("optimization finished, #iter = %d\n", iter);
if(iter >= max_iterations)
SG_WARNING("\nWARNING: reaching max number of iterations\n");
// calculate objective value
double v = 0;
int nnz = 0;
for(j=0; j<w_size; j++)
if(w.vector[j] != 0)
{
v += fabs(w.vector[j]);
nnz++;
}
for(j=0; j<l; j++)
if(y[j] == 1)
v += C[GETI(j)]*log(1+1/exp_wTx[j]);
else
v += C[GETI(j)]*log(1+exp_wTx[j]);
SG_INFO("Objective value = %lf\n", v);
SG_INFO("#nonzeros/#features = %d/%d\n", nnz, w_size);
delete [] index;
delete [] y;
delete [] exp_wTx;
delete [] exp_wTx_new;
delete [] xj_max;
delete [] C_sum;
delete [] xjneg_sum;
delete [] xjpos_sum;
}
void CLibLinear::solve_l1r_l2_svc(
problem *prob_col, double eps, double Cp, double Cn)
{
int l = prob_col->l;
int w_size = prob_col->n;
int j, s, iter = 0;
int active_size = w_size;
int max_num_linesearch = 20;
double sigma = 0.01;
double d, G_loss, G, H;
double Gmax_old = CMath::INFTY;
double Gmax_new;
double Gmax_init=0;
double d_old, d_diff;
double loss_old=0, loss_new;
double appxcond, cond;
int *index = SG_MALLOC(int, w_size);
int32_t *y = SG_MALLOC(int32_t, l);
double *b = SG_MALLOC(double, l); // b = 1-ywTx
double *xj_sq = SG_MALLOC(double, w_size);
CDotFeatures* x = (CDotFeatures*) prob_col->x;
void* iterator;
int32_t ind;
float64_t val;
double C[3] = {Cn,0,Cp};
int n = prob_col->n;
if (prob_col->use_bias)
n--;
for(j=0; j<l; j++)
{
b[j] = 1;
if(prob_col->y[j] > 0)
y[j] = 1;
else
y[j] = -1;
}
for(j=0; j<w_size; j++)
{
w.vector[j] = 0;
index[j] = j;
xj_sq[j] = 0;
if (use_bias && j==n)
{
for (ind=0; ind<l; ind++)
xj_sq[n] += C[GETI(ind)];
}
else
{
iterator=x->get_feature_iterator(j);
while (x->get_next_feature(ind, val, iterator))
xj_sq[j] += C[GETI(ind)]*val*val;
x->free_feature_iterator(iterator);
}
}
CTime start_time;
while (iter < max_iterations && !CSignal::cancel_computations())
{
if (m_max_train_time > 0 && start_time.cur_time_diff() > m_max_train_time)
break;
Gmax_new = 0;
for(j=0; j<active_size; j++)
{
int i = j+rand()%(active_size-j);
CMath::swap(index[i], index[j]);
}
for(s=0; s<active_size; s++)
{
j = index[s];
G_loss = 0;
H = 0;
if (use_bias && j==n)
{
for (ind=0; ind<l; ind++)
{
if(b[ind] > 0)
{
double tmp = C[GETI(ind)]*y[ind];
G_loss -= tmp*b[ind];
H += tmp*y[ind];
}
}
}
else
{
iterator=x->get_feature_iterator(j);
while (x->get_next_feature(ind, val, iterator))
{
if(b[ind] > 0)
{
double tmp = C[GETI(ind)]*val*y[ind];
G_loss -= tmp*b[ind];
H += tmp*val*y[ind];
}
}
x->free_feature_iterator(iterator);
}
G_loss *= 2;
G = G_loss;
H *= 2;
H = CMath::max(H, 1e-12);
double Gp = G+1;
double Gn = G-1;
double violation = 0;
if(w.vector[j] == 0)
{
if(Gp < 0)
violation = -Gp;
else if(Gn > 0)
violation = Gn;
else if(Gp>Gmax_old/l && Gn<-Gmax_old/l)
{
active_size--;
CMath::swap(index[s], index[active_size]);
s--;
continue;
}
}
else if(w.vector[j] > 0)
violation = fabs(Gp);
else
violation = fabs(Gn);
Gmax_new = CMath::max(Gmax_new, violation);
// obtain Newton direction d
if(Gp <= H*w.vector[j])
d = -Gp/H;
else if(Gn >= H*w.vector[j])
d = -Gn/H;
else
d = -w.vector[j];
if(fabs(d) < 1.0e-12)
continue;
double delta = fabs(w.vector[j]+d)-fabs(w.vector[j]) + G*d;
d_old = 0;
int num_linesearch;
for(num_linesearch=0; num_linesearch < max_num_linesearch; num_linesearch++)
{
d_diff = d_old - d;
cond = fabs(w.vector[j]+d)-fabs(w.vector[j]) - sigma*delta;
appxcond = xj_sq[j]*d*d + G_loss*d + cond;
if(appxcond <= 0)
{
if (use_bias && j==n)
{
for (ind=0; ind<l; ind++)
b[ind] += d_diff*y[ind];
break;
}
else
{
iterator=x->get_feature_iterator(j);
while (x->get_next_feature(ind, val, iterator))
b[ind] += d_diff*val*y[ind];
x->free_feature_iterator(iterator);
break;
}
}
if(num_linesearch == 0)
{
loss_old = 0;
loss_new = 0;
if (use_bias && j==n)
{
for (ind=0; ind<l; ind++)
{
if(b[ind] > 0)
loss_old += C[GETI(ind)]*b[ind]*b[ind];
double b_new = b[ind] + d_diff*y[ind];
b[ind] = b_new;
if(b_new > 0)
loss_new += C[GETI(ind)]*b_new*b_new;
}
}
else
{
iterator=x->get_feature_iterator(j);
while (x->get_next_feature(ind, val, iterator))
{
if(b[ind] > 0)
loss_old += C[GETI(ind)]*b[ind]*b[ind];
double b_new = b[ind] + d_diff*val*y[ind];
b[ind] = b_new;
if(b_new > 0)
loss_new += C[GETI(ind)]*b_new*b_new;
}
x->free_feature_iterator(iterator);
}
}
else
{
loss_new = 0;
if (use_bias && j==n)
{
for (ind=0; ind<l; ind++)
{
double b_new = b[ind] + d_diff*y[ind];
b[ind] = b_new;
if(b_new > 0)
loss_new += C[GETI(ind)]*b_new*b_new;
}
}
else
{
iterator=x->get_feature_iterator(j);
while (x->get_next_feature(ind, val, iterator))
{
double b_new = b[ind] + d_diff*val*y[ind];
b[ind] = b_new;
if(b_new > 0)
loss_new += C[GETI(ind)]*b_new*b_new;
}
x->free_feature_iterator(iterator);
}
}
cond = cond + loss_new - loss_old;
if(cond <= 0)
break;
else
{
d_old = d;
d *= 0.5;
delta *= 0.5;
}
}
w.vector[j] += d;
// recompute b[] if line search takes too many steps
if(num_linesearch >= max_num_linesearch)
{
SG_INFO("#");
for(int i=0; i<l; i++)
b[i] = 1;
for(int i=0; i<n; i++)
{
if(w.vector[i]==0)
continue;
iterator=x->get_feature_iterator(i);
while (x->get_next_feature(ind, val, iterator))
b[ind] -= w.vector[i]*val*y[ind];
x->free_feature_iterator(iterator);
}
if (use_bias && w.vector[n])
{
for (ind=0; ind<l; ind++)
b[ind] -= w.vector[n]*y[ind];
}
}
}
if(iter == 0)
Gmax_init = Gmax_new;
iter++;
SG_SABS_PROGRESS(Gmax_new, -CMath::log10(Gmax_new), -CMath::log10(Gmax_init), -CMath::log10(eps*Gmax_init), 6);
if(Gmax_new <= eps*Gmax_init)
{
if(active_size == w_size)
break;
else
{
active_size = w_size;
Gmax_old = CMath::INFTY;
continue;
}
}
Gmax_old = Gmax_new;
}
SG_DONE();
SG_INFO("optimization finished, #iter = %d\n", iter);
if(iter >= max_iterations)
SG_WARNING("\nWARNING: reaching max number of iterations\n");
// calculate objective value
double v = 0;
int nnz = 0;
for(j=0; j<w_size; j++)
{
if(w.vector[j] != 0)
{
v += fabs(w.vector[j]);
nnz++;
}
}
for(j=0; j<l; j++)
if(b[j] > 0)
v += C[GETI(j)]*b[j]*b[j];
SG_INFO("Objective value = %lf\n", v);
SG_INFO("#nonzeros/#features = %d/%d\n", nnz, w_size);
delete [] index;
delete [] y;
delete [] b;
delete [] xj_sq;
}
