/**
* An adaptation of the Zhang-Oles 2001 CLG algorithm by Genkin et al. to
* use the Laplace prior for parameter regularization. On completion,
* optimizes the beta vector to maximize the likelihood of the data set.
*
* @param object $X SparseMatrix representing the training dataset
* @param array $y array of known labels corresponding to the rows of $X
*/
function train($X, $y)
{
$invX = new InvertedData($X);
$this->lambda = $this->estimateLambdaNorm($invX);
$m = $invX->rows();
$n = $invX->columns();
$this->beta = array_fill(0, $n, 0.0);
$beta =& $this->beta;
$lambda = $this->lambda;
$d = array_fill(0, $n, 1.0);
$r = array_fill(0, $m, 0.0);
$converged = false;
$drSum = 0.0;
$rSum = 0.0;
$change = 0.0;
$score = 0.0;
$minDrj = $this->epsilon;
$prevDrj = $this->epsilon;
$schedule = new SplMaxHeap();
$nextSchedule = new SplMaxHeap();
for ($j = 0; $j < $n; $j++) {
$schedule->insert(array($this->epsilon, $j));
}
for ($k = 0; !$converged; $k++) {
$prevR = $r;
$var = 1;
while (!$schedule->isEmpty()) {
list($drj, $j) = $schedule->top();
if ($drj < $minDrj) {
break;
} else {
$schedule->extract();
$prevDrj = $drj;
}
$Xj = $invX->iterateColumn($j);
list($numer, $denom) = $this->computeApproxLikelihood($Xj, $y, $r, $d[$j]);
// Compute tentative step $dvj
if ($beta[$j] == 0) {
$dvj = ($numer - $lambda) / $denom;
if ($dvj <= 0) {
$dvj = ($numer + $lambda) / $denom;
if ($dvj >= 0) {
$dvj = 0;
}
}
} else {
$s = $beta[$j] > 0 ? 1 : -1;
$dvj = ($numer - $s * $lambda) / $denom;
if ($s * ($beta[$j] + $dvj) < 0) {
$dvj = -$beta[$j];
}
}
if ($dvj == 0) {
$d[$j] /= 2;
$nextSchedule->insert(array($this->epsilon, $j, $k));
} else {
// Compute delta for beta[j], constrained to trust region.
$dbetaj = min(max($dvj, -$d[$j]), $d[$j]);
// Update our cached dot product by the delta.
$drj = 0.0;
foreach ($Xj as $cell) {
list($_, $i, $Xij) = $cell;
$dr = $dbetaj * $Xij;
$drj += $dr;
$r[$i] += $dr;
}
$drj = abs($drj);
$nextSchedule->insert(array($drj, $j, $k));
$beta[$j] += $dbetaj;
// Update the trust region.
$d[$j] = max(2 * abs($dbetaj), $d[$j] / 2);
}
if ($this->debug > 1) {
$score = $this->score($r, $y, $beta);
}
$this->log(sprintf("itr = %3d, j = %4d (#%d), score = %6.2f, change = %6.4f", $k + 1, $j, $var, $score, $change));
$var++;
}
// Update $converged
$drSum = 0.0;
$rSum = 0.0;
for ($i = 0; $i < $m; $i++) {
$drSum += abs($r[$i] - $prevR[$i]);
$rSum += abs($r[$i]);
}
$change = $drSum / (1 + $rSum);
$converged = $change <= $this->epsilon;
while (!$schedule->isEmpty()) {
list($drj, $j) = $schedule->extract();
$nextSchedule->insert(array($drj * 4, $j));
}
$tmp = $schedule;
$schedule = $nextSchedule;
$nextSchedule = $tmp;
$minDrj *= 2;
}
}
$heap->insert('3');
$heap->insert('1');
$heap->insert('2');
$heap->insert('5');
$heap->insert('4');
// $heap = new SplPriorityQueue();
// insert($value, $priority)
// 优先队列插入新值
//$heap->insert('A', '3');
//$heap->insert('B', '1');
//$heap->insert('C', '2');
//$heap->insert('D', '5');
//$heap->insert('E', '4');
// isEmpty()
// 判断该堆是否为空
$heap->isEmpty();
// count()
// 获得堆值的数量
$heap->count();
// key()
// 返回当前节点的索引值
$heap->key();
// valid()
// 判断堆中还是否有其他节点
$heap->valid();
// rewind()
// 返回首节点，在堆中是空操作。因为堆是二叉树，rewind始终会在当前位置而不移动
$heap->rewind();
// current()
// 获得当前节点
$heap->current();
<?php $h = new SplMaxHeap(); var_dump($h->isEmpty());
<?php
$data_provider = array(new stdClass(), array(), true, "string", 12345, 1.2345, NULL);
foreach ($data_provider as $input) {
$h = new SplMaxHeap();
var_dump($h->isEmpty($input));
}
<?php $h = new SplMaxHeap(); echo "Checking a new heap is empty: "; var_dump($h->isEmpty()) . "\n"; $h->insert(2); echo "Checking after insert: "; var_dump($h->isEmpty()) . "\n"; $h->extract(); echo "Checking after extract: "; var_dump($h->isEmpty()) . "\n";