func ExamplePGM7_LogisticRegression_Epsilon() {
rand.Seed(98765)
trainFeatures, trainLabels := readTrainData()
benchmarkFeatures, benchmarkLabels := readBenchmarkData()
for _, epsilon := range []float64{1e-1, 1e-2, 1e-3} {
fmt.Println("---\nepsilon: ", epsilon)
classifier := ai.TrainLogisticRegressionClassifier(
trainFeatures,
trainLabels,
0,
&sgrad.RelativeMeanImprovementCrit{},
epsilon)
fmt.Println("train set: ", ai.EvaluateBinaryClassifier(classifier, trainFeatures, trainLabels))
fmt.Println("benchmark set: ", ai.EvaluateBinaryClassifier(classifier, benchmarkFeatures, benchmarkLabels))
}
// Output:
// ---
// epsilon: 0.1
// train set: 0.99
// benchmark set: 0.94
// ---
// epsilon: 0.01
// train set: 0.99
// benchmark set: 0.925
// ---
// epsilon: 0.001
// train set: 1
// benchmark set: 0.93
}
func ExamplePGM7_LogisticRegression_OptimizeLambda() {
rand.Seed(98765)
trainFeatures, trainLabels := readTrainData()
benchmarkFeatures, benchmarkLabels := readBenchmarkData()
goalFunc := func(lambda float64) float64 {
score := ai.HoldoutTestBinaryClassifier(
trainFeatures,
trainLabels,
.1,
ai.NewLogisticRegressionTrainer(
lambda,
&sgrad.NumIterationsCrit{NumIterations: 10},
1e-8))
return -score
}
lambda := gssearh.Minimize(0, 10, goalFunc, &gssearh.AbsoluteErrorTermCrit{}, .1)
fmt.Println("Optimal lambda:", lambda)
classifier := ai.TrainLogisticRegressionClassifier(
trainFeatures,
trainLabels,
lambda,
&sgrad.RelativeMeanImprovementCrit{},
1e-2)
fmt.Println("train set: ", ai.EvaluateBinaryClassifier(classifier, trainFeatures, trainLabels))
fmt.Println("benchmark set: ", ai.EvaluateBinaryClassifier(classifier, benchmarkFeatures, benchmarkLabels))
// Output:
// Optimal lambda: 1.4841053312063623
// train set: 0.895
// benchmark set: 0.9
}
func ExamplePGM7_LogisticRegression_Iterations() {
rand.Seed(98765)
trainFeatures, trainLabels := readTrainData()
benchmarkFeatures, benchmarkLabels := readBenchmarkData()
for _, iterations := range []int{1, 10, 100, 1000} {
fmt.Println("---\niterations: ", iterations)
classifier := ai.TrainLogisticRegressionClassifier(
trainFeatures,
trainLabels,
0,
&sgrad.NumIterationsCrit{NumIterations: iterations},
1e-8)
fmt.Println("train set: ", ai.EvaluateBinaryClassifier(classifier, trainFeatures, trainLabels))
fmt.Println("benchmark set: ", ai.EvaluateBinaryClassifier(classifier, benchmarkFeatures, benchmarkLabels))
}
// Output:
// ---
// iterations: 1
// train set: 0.955
// benchmark set: 0.94
// ---
// iterations: 10
// train set: 0.985
// benchmark set: 0.925
// ---
// iterations: 100
// train set: 1
// benchmark set: 0.925
// ---
// iterations: 1000
// train set: 1
// benchmark set: 0.93
}
func main() {
flag.Parse()
{
f, err := os.Create("digits.prof")
if err != nil {
panic(err)
}
pprof.StartCPUProfile(f)
}
defer pprof.StopCPUProfile()
fmt.Print("Reading training data...")
file, err := os.Open(*trainCsvPath)
if err != nil {
panic(err)
}
csvReader := csv.NewReader(bufio.NewReader(file))
allData, err := csvReader.ReadAll()
if err != nil {
panic(err)
}
allData = allData[1:] // remove header
fmt.Println("done:", len(allData), "rows")
labels := make([]int, len(allData))
pixels := make([]v.F64, len(allData))
// TODO(mike): add bias term
for i, row := range allData {
parsedLabel, err := strconv.ParseInt(row[0], 10, 32)
if err != nil {
panic(err)
}
labels[i] = int(parsedLabel)
pixels[i], err = parseVector(row[1:])
}
binClassifierTrainer := func(features []v.F64, labels []bool) ai.BinaryClassifier {
fmt.Println("Training binary classifier")
classifier := ai.TrainLogisticRegressionClassifier(
features,
labels,
0,
&sgrad.NumIterationsCrit{NumIterations: 10},
1e-8)
return classifier
}
classifier := ai.TrainNominalClassifierFromBinary(pixels, labels, 10, binClassifierTrainer)
fmt.Println(classifier)
}
func ExamplePGM7_LogisticRegression_Lambda() {
rand.Seed(98765)
trainFeatures, trainLabels := readTrainData()
benchmarkFeatures, benchmarkLabels := readBenchmarkData()
for _, lambda := range []float64{0, 0.1, 0.2, 0.4, 0.8, 1} {
fmt.Println("---\nlambda: ", lambda)
classifier := ai.TrainLogisticRegressionClassifier(
trainFeatures,
trainLabels,
lambda,
&sgrad.RelativeMeanImprovementCrit{},
1e-2)
fmt.Println("train set: ", ai.EvaluateBinaryClassifier(classifier, trainFeatures, trainLabels))
fmt.Println("benchmark set: ", ai.EvaluateBinaryClassifier(classifier, benchmarkFeatures, benchmarkLabels))
}
// Output:
// ---
// lambda: 0
// train set: 1
// benchmark set: 0.93
// ---
// lambda: 0.1
// train set: 0.94
// benchmark set: 0.93
// ---
// lambda: 0.2
// train set: 0.94
// benchmark set: 0.925
// ---
// lambda: 0.4
// train set: 0.925
// benchmark set: 0.915
// ---
// lambda: 0.8
// train set: 0.875
// benchmark set: 0.86
// ---
// lambda: 1
// train set: 0.905
// benchmark set: 0.905
}