void GBDT::fit(Problem const &Tr, Problem const &Va)
{
bias = calc_bias(Tr.Y);
std::vector<float> F_Tr(Tr.nr_instance, bias), F_Va(Va.nr_instance, bias);
Timer timer;
printf("iter time tr_loss va_loss\n");
for(uint32_t t = 0; t < trees.size(); ++t)
{
timer.tic();
std::vector<float> const &Y = Tr.Y;
std::vector<float> R(Tr.nr_instance), F1(Tr.nr_instance);
#pragma omp parallel for schedule(static)
for(uint32_t i = 0; i < Tr.nr_instance; ++i){
double tmp = Y[i]*F_Tr[i];
//tmp = (tmp < 1e-12)? 1e-12 : tmp ;
//tmp = (tmp > 1e12)? 1e12 : tmp ;
//R[i] = static_cast<float>(Y[i]/(1+exp(Y[i]*F_Tr[i])));
R[i] = static_cast<float>(Y[i]/(1+exp(tmp)));
}
trees[t].fit(Tr, R, F1);
double Tr_loss = 0;
#pragma omp parallel for schedule(static) reduction(+: Tr_loss)
for(uint32_t i = 0; i < Tr.nr_instance; ++i)
{
F_Tr[i] += F1[i];
Tr_loss += log(1+exp(-Y[i]*F_Tr[i]));
}
Tr_loss /= static_cast<double>(Tr.nr_instance);
#pragma omp parallel for schedule(static)
for(uint32_t i = 0; i < Va.nr_instance; ++i)
{
std::vector<float> x = construct_instance(Va, i);
F_Va[i] += trees[t].predict(x.data()).second;
}
double Va_loss = 0;
#pragma omp parallel for schedule(static) reduction(+: Va_loss)
for(uint32_t i = 0; i < Va.nr_instance; ++i)
Va_loss += log(1+exp(-Va.Y[i]*F_Va[i]));
Va_loss /= static_cast<double>(Va.nr_instance);
printf("%4d %8.1f %10.5f %10.5f\n", t, timer.toc(), Tr_loss, Va_loss);
fflush(stdout);
}
}
void GBDT::fit(Problem const &Tr, Problem const &Va)
{
bias = calc_bias(Tr.Y);
std::vector<float> F_Tr(Tr.nr_instance, bias), F_Va(Va.nr_instance, bias);
Timer timer;
printf("iter time tr_loss va_loss\n");
// 感觉这里使用的是 logloss，即logit boost算法。
// 整体算法取自 The Elements of Statistical Learning Algorithm 10.3
for(uint32_t t = 0; t < trees.size(); ++t)
{
timer.tic();
std::vector<float> const &Y = Tr.Y;
std::vector<float> R(Tr.nr_instance), F1(Tr.nr_instance);
// 计算残差，也就是loss function的梯度下降
#pragma omp parallel for schedule(static)
for(uint32_t i = 0; i < Tr.nr_instance; ++i)
R[i] = static_cast<float>(Y[i]/(1+exp(Y[i]*F_Tr[i])));
// regression tree，去fit targets R[i]
trees[t].fit(Tr, R, F1);
// 计算整体的loss，并累加F_Tr(sum{y})
// 可以看到，这里没有使用权重alpha，也就是没有使用bootstrap re-sample
double Tr_loss = 0;
#pragma omp parallel for schedule(static) reduction(+: Tr_loss)
for(uint32_t i = 0; i < Tr.nr_instance; ++i)
{
F_Tr[i] += F1[i];
Tr_loss += log(1+exp(-Y[i]*F_Tr[i]));
}
Tr_loss /= static_cast<double>(Tr.nr_instance);
// 计算validation loss
#pragma omp parallel for schedule(static)
for(uint32_t i = 0; i < Va.nr_instance; ++i)
{
std::vector<float> x = construct_instance(Va, i);
F_Va[i] += trees[t].predict(x.data()).second;
}
double Va_loss = 0;
#pragma omp parallel for schedule(static) reduction(+: Va_loss)
for(uint32_t i = 0; i < Va.nr_instance; ++i)
Va_loss += log(1+exp(-Va.Y[i]*F_Va[i]));
Va_loss /= static_cast<double>(Va.nr_instance);
printf("%4d %8.1f %10.5f %10.5f\n", t, timer.toc(), Tr_loss, Va_loss);
fflush(stdout);
}
}
