/**
* This is the function that should be called to get the
* classifier_trainer to start training a logistic regression instance for
* a particular classifier. The class label corresponding to the
* classifier to be finalized should be passed as the second command-line
* argument.
*/
function start()
{
global $argv;
CrawlDaemon::init($argv, "classifier_trainer");
$label = $argv[2];
crawlLog("Initializing classifier trainer log..", $label . '-classifier_trainer', true);
$classifier = Classifier::getClassifier($label);
$classifier->prepareToFinalize();
$classifier->finalize();
Classifier::setClassifier($classifier);
crawlLog("Training complete.\n");
CrawlDaemon::stop('classifier_trainer', $label);
}
/**
* Creates a new classifier for a label, first deleting any existing
* classifier with the same label.
*
* @param string $label class label of the new classifier
* @return object created classifier instance
*/
function makeFreshClassifier($label)
{
if ($classifier = Classifier::getClassifier($label)) {
$this->deleteClassifier($label);
}
$classifier = new Classifier($label, $this->options['cls']);
Classifier::setClassifier($classifier);
return $classifier;
}
/**
* Finds the next document for which to request a label, sometimes first
* recording the label that the user selected for the last document. This
* method should only be called via an XmlHttpRequest initiated by the edit
* classifier JavaScript, and consequently it always writes out
* JSON-encoded data, which is easily decoded by the page JavaScript.
*/
function classify()
{
$arg = $this->clean($_REQUEST['arg'], 'string');
$label = $this->clean($_REQUEST['label'], 'string');
if (isset($_REQUEST['index'])) {
$index = $this->clean($_REQUEST['index'], 'int');
if (intval($index) == 1) {
$index = $this->model("crawl")->getCurrentIndexDatabaseName();
}
$source_type = $this->clean($_REQUEST['type'], 'string');
$keywords = $this->clean($_REQUEST['keywords'], 'string');
}
/*
The call to prepareToLabel is important; it loads all of the data
required to manage the training set from disk, and also determines
what will be saved *back* to disk later.
*/
$classifier = Classifier::getClassifier($label);
$classifier->prepareToLabel();
$data = array();
switch ($arg) {
case 'getdocs':
/*
Load documents in from a user-specified index, and find the
next best one to label (for 'manual' source type), or label
them all with a single label (for either the 'positive' or
'negative' source types).
*/
$mix_iterator = $this->buildClassifierCrawlMix($label, $index, $keywords);
if ($source_type == 'manual') {
$num_docs = $classifier->initBuffer($mix_iterator);
$classifier->computeBufferDensities();
$data['num_docs'] = $num_docs;
list($new_doc, $disagreement) = $classifier->findNextDocumentToLabel();
if ($new_doc) {
$score = $classifier->classify($new_doc);
$data['new_doc'] = $this->prepareUnlabelledDocument($new_doc, $score, $disagreement, $index, $keywords);
}
Classifier::setClassifier($classifier);
} else {
if ($source_type == 'positive' || $source_type == 'negative') {
$doc_label = $source_type == 'positive' ? 1 : -1;
$add_count = $classifier->addAllDocuments($mix_iterator, $doc_label);
if ($add_count > 0) {
/*
Pass true to always update accuracy after adding a
batch of documents all at once.
*/
$classifier->train(true);
Classifier::setClassifier($classifier);
}
$data['add_count'] = $add_count;
}
}
break;
case 'addlabel':
/*
First label the last candidate document presented to the
user (potentially skipping it instead of actually applying a
label), then pick the next best candidate for labeling.
When skipping a document instead of adding a label, avoid
re-training since the training set hasn't actually changed.
*/
$doc = $_REQUEST['doc_to_label'];
$docid = $this->clean($doc['docid'], 'int');
$key = webdecode($this->clean($doc['key'], 'string'));
$doc_label = $this->clean($doc['label'], 'int');
$mix_iterator = $this->retrieveClassifierCrawlMix($label);
$labels_changed = $classifier->labelDocument($key, $doc_label);
$num_docs = $classifier->refreshBuffer($mix_iterator);
$classifier->computeBufferDensities();
$data['num_docs'] = $num_docs;
if ($labels_changed) {
$update_accuracy = $classifier->total > 0 && $classifier->total % 10 == 0;
$classifier->train($update_accuracy);
}
list($new_doc, $disagreement) = $classifier->findNextDocumentToLabel();
if ($new_doc) {
$score = $classifier->classify($new_doc);
$data['new_doc'] = $this->prepareUnlabelledDocument($new_doc, $score, $disagreement, $index, $keywords);
}
Classifier::setClassifier($classifier);
break;
case 'updateaccuracy':
/*
Don't do anything other than re-compute the accuracy for the
current training set.
*/
$classifier->updateAccuracy();
Classifier::setClassifier($classifier);
break;
}
/*
No matter which activity we ended up carrying out, always include
the statistics that *might* have changed so that the client can just
naively keep them up to date.
*/
$data['positive'] = $classifier->positive;
$data['negative'] = $classifier->negative;
$data['total'] = $classifier->total;
$data['accuracy'] = $classifier->accuracy;
/*
Pass along a new authentication token so that the client can make a
new authenticated request after this one.
*/
$data['authTime'] = strval(time());
$data['authSession'] = md5($data['authTime'] . AUTH_KEY);
$response = json_encode($data);
header("Content-Type: application/json");
header("Content-Length: " . strlen($response));
echo $response;
}
