public void train() {
double mu;
int qid, i, j, pSize;
for (int n = 0; n < m_maxIter; n++) {
Utils.shuffle(m_order, m_trainingSize);
qid = 0;
while (qid < m_trainingSize) {
pSize = 0;
Arrays.fill(m_g, 0.0);
for (j = 0; j < m_windowSize; j++) { // collect the gradients in mini-batch
pSize += gradientUpdate(m_queries.get(m_order[qid % m_trainingSize]));
qid++;
}
// Step 4: gradient from regularization
for (i = 0; i < m_weight.length; i++) m_g[i] = m_g[i] / pSize + m_lambda * m_weight[i];
mu = Math.random() * m_step;
for (i = 0; i < m_weight.length; i++) m_weight[i] -= mu * m_g[i];
}
m_step *= m_shrinkage;
}
}
