/**
* 57% of time spent in LogConditionalObjectiveFunction.calculateCLBatch() 22% spent in
* constructing datums (expensive)
*
* <p>Single threaded, 4100 ms Multi threaded, 600 ms
*
* <p>With same data, seed 42, 52 ms With reordered accesses for cacheing, 38 ms Down to 73% of
* the time
*
* <p>with 8 cpus, a 6.8x speedup -- basically the same as with RVFDatum
*/
public static void benchmarkLogisticRegression() {
Dataset<String, String> data = new Dataset<>();
for (int i = 0; i < 10000; i++) {
Random r = new Random(42);
Set<String> features = new HashSet<>();
boolean cl = r.nextBoolean();
for (int j = 0; j < 1000; j++) {
if (cl && i % 2 == 0) {
if (r.nextDouble() > 0.3) {
features.add("f:" + j + ":true");
} else {
features.add("f:" + j + ":false");
}
} else {
if (r.nextDouble() > 0.3) {
features.add("f:" + j + ":false");
} else {
features.add("f:" + j + ":false");
}
}
}
data.add(new BasicDatum<String, String>(features, "target:" + cl));
}
LinearClassifierFactory<String, String> factory = new LinearClassifierFactory<>();
long msStart = System.currentTimeMillis();
factory.trainClassifier(data);
long delay = System.currentTimeMillis() - msStart;
System.out.println("Training took " + delay + " ms");
}
/**
* 67% of time spent in LogConditionalObjectiveFunction.rvfcalculate() 29% of time spent in
* dataset construction (11% in RVFDataset.addFeatures(), 7% rvf incrementCount(), 11% rest)
*
* <p>Single threaded, 4700 ms Multi threaded, 700 ms
*
* <p>With same data, seed 42, 245 ms With reordered accesses for cacheing, 195 ms Down to 80% of
* the time, not huge but a win nonetheless
*
* <p>with 8 cpus, a 6.7x speedup -- almost, but not quite linear, pretty good
*/
public static void benchmarkRVFLogisticRegression() {
RVFDataset<String, String> data = new RVFDataset<>();
for (int i = 0; i < 10000; i++) {
Random r = new Random(42);
Counter<String> features = new ClassicCounter<>();
boolean cl = r.nextBoolean();
for (int j = 0; j < 1000; j++) {
double value;
if (cl && i % 2 == 0) {
value = (r.nextDouble() * 2.0) - 0.6;
} else {
value = (r.nextDouble() * 2.0) - 1.4;
}
features.incrementCount("f" + j, value);
}
data.add(new RVFDatum<>(features, "target:" + cl));
}
LinearClassifierFactory<String, String> factory = new LinearClassifierFactory<>();
long msStart = System.currentTimeMillis();
factory.trainClassifier(data);
long delay = System.currentTimeMillis() - msStart;
System.out.println("Training took " + delay + " ms");
}
public static void benchmarkDatum() {
long msStart = System.currentTimeMillis();
Dataset<String, String> data = new Dataset<>();
for (int i = 0; i < 10000; i++) {
Random r = new Random(42);
Set<String> features = new HashSet<>();
boolean cl = r.nextBoolean();
for (int j = 0; j < 1000; j++) {
if (cl && i % 2 == 0) {
if (r.nextDouble() > 0.3) {
features.add("f:" + j + ":true");
} else {
features.add("f:" + j + ":false");
}
} else {
if (r.nextDouble() > 0.3) {
features.add("f:" + j + ":false");
} else {
features.add("f:" + j + ":false");
}
}
}
data.add(new BasicDatum<String, String>(features, "target:" + cl));
}
long delay = System.currentTimeMillis() - msStart;
System.out.println("Dataset construction took " + delay + " ms");
msStart = System.currentTimeMillis();
for (int i = 0; i < 10000; i++) {
Random r = new Random(42);
Set<String> features = new HashSet<>();
boolean cl = r.nextBoolean();
for (int j = 0; j < 1000; j++) {
if (cl && i % 2 == 0) {
if (r.nextDouble() > 0.3) {
} else {
}
} else {
if (r.nextDouble() > 0.3) {
} else {
}
}
}
}
delay = System.currentTimeMillis() - msStart;
System.out.println("MultiVector took " + delay + " ms");
}
/**
* 29% in FactorTable.getValue() 28% in CRFCliqueTree.getCalibratedCliqueTree() 12.6% waiting for
* threads
*
* <p>Single threaded: 15000 ms - 26000 ms Multi threaded: 4500 ms - 7000 ms
*
* <p>with 8 cpus, 3.3x - 3.7x speedup, around 800% utilization
*/
public static void benchmarkCRF() {
Properties props = new Properties();
props.setProperty("macro", "true"); // use a generic CRF configuration
props.setProperty("useIfInteger", "true");
props.setProperty("featureFactory", "edu.stanford.nlp.benchmarks.BenchmarkFeatureFactory");
props.setProperty("saveFeatureIndexToDisk", "false");
CRFClassifier<CoreLabel> crf = new CRFClassifier<CoreLabel>(props);
Random r = new Random(42);
List<List<CoreLabel>> data = new ArrayList<>();
for (int i = 0; i < 100; i++) {
List<CoreLabel> sentence = new ArrayList<>();
for (int j = 0; j < 20; j++) {
CoreLabel l = new CoreLabel();
l.setWord("j:" + j);
boolean tag = j % 2 == 0 ^ (r.nextDouble() > 0.7);
l.set(CoreAnnotations.AnswerAnnotation.class, "target:" + tag);
sentence.add(l);
}
data.add(sentence);
}
long msStart = System.currentTimeMillis();
crf.train(data);
long delay = System.currentTimeMillis() - msStart;
System.out.println("Training took " + delay + " ms");
}
public static void benchmarkSGD() {
Dataset<String, String> data = new Dataset<>();
for (int i = 0; i < 10000; i++) {
Random r = new Random(42);
Set<String> features = new HashSet<>();
boolean cl = r.nextBoolean();
for (int j = 0; j < 1000; j++) {
if (cl && i % 2 == 0) {
if (r.nextDouble() > 0.3) {
features.add("f:" + j + ":true");
} else {
features.add("f:" + j + ":false");
}
} else {
if (r.nextDouble() > 0.3) {
features.add("f:" + j + ":false");
} else {
features.add("f:" + j + ":false");
}
}
}
data.add(new BasicDatum<String, String>(features, "target:" + cl));
}
LinearClassifierFactory<String, String> factory = new LinearClassifierFactory<>();
factory.setMinimizerCreator(
new Factory<Minimizer<DiffFunction>>() {
@Override
public Minimizer<DiffFunction> create() {
return new SGDMinimizer<DiffFunction>(0.1, 100, 0, 1000);
}
});
long msStart = System.currentTimeMillis();
factory.trainClassifier(data);
long delay = System.currentTimeMillis() - msStart;
System.out.println("Training took " + delay + " ms");
}