Skip to content

Commit

Permalink
continued sgvector transition
Browse files Browse the repository at this point in the history 
- convert linear machines to use SGVector w instead of double ptr
- remove various unneedef SG_FREEs
  • Loading branch information
Soeren Sonnenburg committed May 5, 2012
1 parent 1f419bb commit eaf5c09
Show file tree
Hide file tree
Showing 17 changed files with 130 additions and 160 deletions.
8 changes: 3 additions & 5 deletions src/shogun/classifier/AveragedPerceptron.cpp
View file
Expand Up @@ -46,12 +46,10 @@ bool CAveragedPerceptron::train(CFeatures* data)
int32_t num_vec=features->get_num_vectors();

ASSERT(num_vec==train_labels.vlen);
SG_FREE(w);
w_dim=num_feat;
w=SG_MALLOC(float64_t, num_feat);
w=SGVector<float64_t>(num_feat);
float64_t* tmp_w=SG_MALLOC(float64_t, num_feat);

float64_t* output=SG_MALLOC(float64_t, num_vec);

//start with uniform w, bias=0, tmp_bias=0
bias=0;
float64_t tmp_bias=0;
Expand All @@ -71,7 +69,7 @@ bool CAveragedPerceptron::train(CFeatures* data)
{
converged=false;
bias+=learn_rate*train_labels.vector[i];
features->add_to_dense_vec(learn_rate*train_labels.vector[i], i, w, w_dim);
features->add_to_dense_vec(learn_rate*train_labels.vector[i], i, w.vector, w.vlen);
}

// Add current w to tmp_w, and current bias to tmp_bias
Expand Down
8 changes: 3 additions & 5 deletions src/shogun/classifier/LDA.cpp
View file
Expand Up @@ -80,9 +80,7 @@ bool CLDA::train_machine(CFeatures* data)
return false;
}

SG_FREE(w);
w=SG_MALLOC(float64_t, num_feat);
w_dim=num_feat;
w=SGVector<float64_t>(num_feat);

float64_t* mean_neg=SG_MALLOC(float64_t, num_feat);
memset(mean_neg,0,num_feat*sizeof(float64_t));
Expand Down Expand Up @@ -177,11 +175,11 @@ bool CLDA::train_machine(CFeatures* data)

bias=0.5*(CMath::dot(w_neg, mean_neg, num_feat)-CMath::dot(w_pos, mean_pos, num_feat));
for (i=0; i<num_feat; i++)
w[i]=w_pos[i]-w_neg[i];
w.vector[i]=w_pos[i]-w_neg[i];

#ifdef DEBUG_LDA
SG_PRINT("bias: %f\n", bias);
CMath::display_vector(w, num_feat, "w");
CMath::display_vector(w.vector, num_feat, "w");
CMath::display_vector(w_pos, num_feat, "w_pos");
CMath::display_vector(w_neg, num_feat, "w_neg");
CMath::display_vector(mean_pos, num_feat, "mean_pos");
Expand Down
9 changes: 3 additions & 6 deletions src/shogun/classifier/Perceptron.cpp
View file
Expand Up @@ -47,18 +47,15 @@ bool CPerceptron::train_machine(CFeatures* data)
int32_t num_vec=features->get_num_vectors();

ASSERT(num_vec==train_labels.vlen);
SG_FREE(w);
w_dim=num_feat;
w=SG_MALLOC(float64_t, num_feat);
w=SGVector<float64_t>(num_feat);
float64_t* output=SG_MALLOC(float64_t, num_vec);

//start with uniform w, bias=0
bias=0;
for (int32_t i=0; i<num_feat; i++)
w[i]=1.0/num_feat;
w.vector[i]=1.0/num_feat;

//loop till we either get everything classified right or reach max_iter

while (!converged && iter<max_iter)
{
converged=true;
Expand All @@ -70,7 +67,7 @@ bool CPerceptron::train_machine(CFeatures* data)
{
converged=false;
bias+=learn_rate*train_labels.vector[i];
features->add_to_dense_vec(learn_rate*train_labels.vector[i], i, w, w_dim);
features->add_to_dense_vec(learn_rate*train_labels.vector[i], i, w.vector, w.vlen);
}
}

Expand Down
88 changes: 43 additions & 45 deletions src/shogun/classifier/svm/LibLinear.cpp
View file
Expand Up @@ -116,23 +116,21 @@ bool CLibLinear::train_machine(CFeatures* data)
num_vec, num_train_labels);
}
}
SG_FREE(w);
if (use_bias)
w=SG_MALLOC(float64_t, num_feat+1);
w=SGVector<float64_t>(SG_MALLOC(float64_t, num_feat+1), num_feat);
else
w=SG_MALLOC(float64_t, num_feat+0);
w_dim=num_feat;
w=SGVector<float64_t>(num_feat);

problem prob;
if (use_bias)
{
prob.n=w_dim+1;
memset(w, 0, sizeof(float64_t)*(w_dim+1));
prob.n=w.vlen+1;
memset(w.vector, 0, sizeof(float64_t)*(w.vlen+1));
}
else
{
prob.n=w_dim;
memset(w, 0, sizeof(float64_t)*(w_dim+0));
prob.n=w.vlen;
memset(w.vector, 0, sizeof(float64_t)*(w.vlen+0));
}
prob.l=num_vec;
prob.x=features;
Expand Down Expand Up @@ -163,7 +161,7 @@ bool CLibLinear::train_machine(CFeatures* data)
fun_obj=new l2r_lr_fun(&prob, Cp, Cn);
CTron tron_obj(fun_obj, epsilon*CMath::min(pos,neg)/prob.l, max_iterations);
SG_DEBUG("starting L2R_LR training via tron\n");
tron_obj.tron(w, m_max_train_time);
tron_obj.tron(w.vector, m_max_train_time);
SG_DEBUG("done with tron\n");
delete fun_obj;
break;
Expand All @@ -172,7 +170,7 @@ bool CLibLinear::train_machine(CFeatures* data)
{
fun_obj=new l2r_l2_svc_fun(&prob, Cp, Cn);
CTron tron_obj(fun_obj, epsilon*CMath::min(pos,neg)/prob.l, max_iterations);
tron_obj.tron(w, m_max_train_time);
tron_obj.tron(w.vector, m_max_train_time);
delete fun_obj;
break;
}
Expand Down Expand Up @@ -200,7 +198,7 @@ bool CLibLinear::train_machine(CFeatures* data)
}

if (use_bias)
set_bias(w[w_dim]);
set_bias(w[w.vlen]);
else
set_bias(0);

Expand Down Expand Up @@ -313,9 +311,9 @@ void CLibLinear::solve_l2r_l1l2_svc(
i = index[s];
int32_t yi = y[i];

G = prob->x->dense_dot(i, w, n);
G = prob->x->dense_dot(i, w.vector, n);
if (prob->use_bias)
G+=w[n];
G+=w.vector[n];

if (m_linear_term.vector)
G = G*yi + m_linear_term.vector[i];
Expand Down Expand Up @@ -362,10 +360,10 @@ void CLibLinear::solve_l2r_l1l2_svc(
alpha[i] = CMath::min(CMath::max(alpha[i] - G/QD[i], 0.0), C);
d = (alpha[i] - alpha_old)*yi;

prob->x->add_to_dense_vec(d, i, w, n);
prob->x->add_to_dense_vec(d, i, w.vector, n);

if (prob->use_bias)
w[n]+=d;
w.vector[n]+=d;
}
}

Expand Down Expand Up @@ -406,7 +404,7 @@ void CLibLinear::solve_l2r_l1l2_svc(
double v = 0;
int nSV = 0;
for(i=0; i<w_size; i++)
v += w[i]*w[i];
v += w.vector[i]*w.vector[i];
for(i=0; i<l; i++)
{
v += alpha[i]*(alpha[i]*diag[GETI(i)] - 2);
Expand Down Expand Up @@ -482,7 +480,7 @@ void CLibLinear::solve_l1r_l2_svc(

for(j=0; j<w_size; j++)
{
w[j] = 0;
w.vector[j] = 0;
index[j] = j;
xj_sq[j] = 0;

Expand Down Expand Up @@ -558,7 +556,7 @@ void CLibLinear::solve_l1r_l2_svc(
double Gp = G+1;
double Gn = G-1;
double violation = 0;
if(w[j] == 0)
if(w.vector[j] == 0)
{
if(Gp < 0)
violation = -Gp;
Expand All @@ -572,31 +570,31 @@ void CLibLinear::solve_l1r_l2_svc(
continue;
}
}
else if(w[j] > 0)
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[j])
if(Gp <= H*w.vector[j])
d = -Gp/H;
else if(Gn >= H*w[j])
else if(Gn >= H*w.vector[j])
d = -Gn/H;
else
d = -w[j];
d = -w.vector[j];

if(fabs(d) < 1.0e-12)
continue;

double delta = fabs(w[j]+d)-fabs(w[j]) + G*d;
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[j]+d)-fabs(w[j]) - sigma*delta;
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)
Expand Down Expand Up @@ -688,7 +686,7 @@ void CLibLinear::solve_l1r_l2_svc(
}
}

w[j] += d;
w.vector[j] += d;

// recompute b[] if line search takes too many steps
if(num_linesearch >= max_num_linesearch)
Expand All @@ -699,19 +697,19 @@ void CLibLinear::solve_l1r_l2_svc(

for(int i=0; i<n; i++)
{
if(w[i]==0)
if(w.vector[i]==0)
continue;

iterator=x->get_feature_iterator(i);
while (x->get_next_feature(ind, val, iterator))
b[ind] -= w[i]*val*y[ind];
b[ind] -= w.vector[i]*val*y[ind];
x->free_feature_iterator(iterator);
}

if (use_bias && w[n])
if (use_bias && w.vector[n])
{
for (ind=0; ind<l; ind++)
b[ind] -= w[n]*y[ind];
b[ind] -= w.vector[n]*y[ind];
}
}
}
Expand Down Expand Up @@ -748,9 +746,9 @@ void CLibLinear::solve_l1r_l2_svc(
int nnz = 0;
for(j=0; j<w_size; j++)
{
if(w[j] != 0)
if(w.vector[j] != 0)
{
v += fabs(w[j]);
v += fabs(w.vector[j]);
nnz++;
}
}
Expand Down Expand Up @@ -832,7 +830,7 @@ void CLibLinear::solve_l1r_lr(
}
for(j=0; j<w_size; j++)
{
w[j] = 0;
w.vector[j] = 0;
index[j] = j;
xj_max[j] = 0;
C_sum[j] = 0;
Expand Down Expand Up @@ -924,7 +922,7 @@ void CLibLinear::solve_l1r_lr(
double Gp = G+1;
double Gn = G-1;
double violation = 0;
if(w[j] == 0)
if(w.vector[j] == 0)
{
if(Gp < 0)
violation = -Gp;
Expand All @@ -938,31 +936,31 @@ void CLibLinear::solve_l1r_lr(
continue;
}
}
else if(w[j] > 0)
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[j])
if(Gp <= H*w.vector[j])
d = -Gp/H;
else if(Gn >= H*w[j])
else if(Gn >= H*w.vector[j])
d = -Gn/H;
else
d = -w[j];
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[j]+d)-fabs(w[j]) + G*d;
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[j]+d)-fabs(w[j]) - sigma*delta;
cond = fabs(w.vector[j]+d)-fabs(w.vector[j]) - sigma*delta;

if(x_min >= 0)
{
Expand Down Expand Up @@ -1046,7 +1044,7 @@ void CLibLinear::solve_l1r_lr(
}
}

w[j] += d;
w.vector[j] += d;

// recompute exp_wTx[] if line search takes too many steps
if(num_linesearch >= max_num_linesearch)
Expand All @@ -1057,18 +1055,18 @@ void CLibLinear::solve_l1r_lr(

for(int i=0; i<w_size; i++)
{
if(w[i]==0) continue;
if(w.vector[i]==0) continue;

if (use_bias && i==n)
{
for (ind=0; ind<l; ind++)
exp_wTx[ind] += w[i];
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[i]*val;
exp_wTx[ind] += w.vector[i]*val;
x->free_feature_iterator(iterator);
}
}
Expand Down Expand Up @@ -1108,9 +1106,9 @@ void CLibLinear::solve_l1r_lr(
double v = 0;
int nnz = 0;
for(j=0; j<w_size; j++)
if(w[j] != 0)
if(w.vector[j] != 0)
{
v += fabs(w[j]);
v += fabs(w.vector[j]);
nnz++;
}
for(j=0; j<l; j++)
Expand Down

0 comments on commit eaf5c09

Please sign in to comment.