logistic reg

regr/src/LogisticRegression.go

@@ -10,7 +10,7 @@ type LogisticRegression struct {
 	Epochs       int
 	Weights      *mat.Dense
 	Bias         float64
-	Losses  []float64
+	LearningRate float64
 }
 
 func sigmoidFunction(x float64) float64 {
@@ -20,54 +20,75 @@ func sigmoidFunction(x float64) float64 {
 	return 1. / (1. + math.Exp(-x))
 }
 
-func (regr *LogisticRegression) backprop(x, y mat.Matrix) float64 {
-	_, c := x.Dims()
-	ry, cy := y.Dims()
-	regr.Bias = 0.1
-	regr.Weights = mat.NewDense(cy, c, nil)
-	coef := &mat.Dense{}
+// binary cross-entropy Loss
+func (regr *LogisticRegression) Loss(yTrue, yPred mat.Matrix) float64 {
+	ep := 1e-9
+	y1 := &mat.Dense{}
+	y1.Apply(func(i, j int, v float64) float64 {
+		return v * math.Log1p(yPred.At(i, j)+ep)
+	}, yTrue)
+	y2 := &mat.Dense{}
+	y2.Apply(func(i, j int, v float64) float64 {
+		return (1. - v) * math.Log1p(1.-yPred.At(i, j)+ep)
+	}, yTrue)
+	sum := &mat.Dense{}
+	sum.Add(y1, y2)
+	w, h := yTrue.Dims()
+	return mat.Sum(sum) / float64(w*h)
+}
 
-	coef.Mul(regr.Weights, x.T())
+func (regr *LogisticRegression) forward(X mat.Matrix) mat.Matrix {
+	coef := &mat.Dense{}
+	coef.Mul(X, regr.Weights)
 	coef.Apply(func(i, j int, v float64) float64 {
 		return sigmoidFunction(v + regr.Bias)
 	}, coef)
-
-	diff := &mat.Dense{}
-	diff.Sub(y.T(), coef)
-
-	w := &mat.Dense{}
-	w.Mul(diff, x)
-	regr.Weights = w
-
-	regr.Bias -= 0.1 * (mat.Sum(diff) / float64(c))
-
-	// Loss
-	yZeroLoss := &mat.Dense{}
-	yZeroLoss.Apply(func(i, j int, v float64) float64 {
-		return v * math.Log1p(coef.At(i, j)+1e-9)
-	}, y.T())
-
-	yOneLoss := &mat.Dense{}
-	yOneLoss.Apply(func(i, j int, v float64) float64 {
-		return (1. - v) * math.Log1p(1.-coef.At(i, j)+1e-9)
-	}, y.T())
-
-	sum := &mat.Dense{}
-	sum.Add(yZeroLoss, yOneLoss)
-	return mat.Sum(sum) / float64(ry+cy)
+	return coef
 }
 
-func (regr *LogisticRegression) Fit(X, Y mat.Matrix, epochs int) error {
-	for i := 0; i < epochs; i++ {
-		loss := regr.backprop(X, Y)
-		regr.Losses = append(regr.Losses, loss)
+func (regr *LogisticRegression) grad(x, yTrue, yPred mat.Matrix) (*mat.Dense, float64) {
+	nSamples, _ := x.Dims()
+	deriv := &mat.Dense{}
+	deriv.Sub(yPred, yTrue)
+	dw := &mat.Dense{}
+	dw.Mul(x.T(), deriv)
+	dw.Apply(func(i, j int, v float64) float64 {
+		return 1. / float64(nSamples) * v
+	}, dw)
+	db := (1. / float64(nSamples)) * mat.Sum(deriv)
+	return dw, db
+}
+
+func (regr *LogisticRegression) backprop(x, y mat.Matrix) float64 {
+	_, c := x.Dims()
+	_, cy := y.Dims()
+	if regr.Weights == nil {
+		regr.Weights = mat.NewDense(c, cy, nil)
 	}
-	return nil
+	if regr.LearningRate == 0 {
+		regr.LearningRate = 0.01
+	}
+	yPred := regr.forward(x)
+	loss := regr.Loss(y, yPred)
+	dw, db := regr.grad(x, y, yPred)
+	regr.Weights.Sub(regr.Weights, dw)
+	regr.Bias -= regr.LearningRate * db
+	return loss
+}
+
+func (regr *LogisticRegression) Fit(X, Y mat.Matrix, epochs int, losses *[]float64) {
+	for i := 0; i < epochs; i++ {
+		regr.backprop(X, Y)
+		if losses != nil {
+			*losses = append(*losses, regr.Loss(Y, regr.forward(X)))
+		}
+	}
+
 }
 
 func (regr *LogisticRegression) Predict(X mat.Matrix) mat.Matrix {
 	coef := &mat.Dense{}
-	coef.Mul(X, regr.Weights.T())
+	coef.Mul(X, regr.Weights)
 	coef.Apply(func(i, j int, v float64) float64 {
 		p := sigmoidFunction(v + regr.Bias)
 		if p > .5 {

regr/src/LogisticRegression_test.go

@@ -6,16 +6,16 @@ import (
 )
 
 func TestLogisticRegression(t *testing.T) {
-	X := [][]float64{{10.1, 10.1, 10.1}, {2.1, 2.1, 2.1}, {10.2, 10.2, 10.2}, {2.2, 2.2, 2.2}}
+	X := [][]float64{{.1, .1, .1}, {.2, .2, .2}, {.1, .1, .1}, {.2, .2, .2}}
 	Y := [][]float64{{0}, {1}, {0}, {1}}
 	XDense := Array2DToDense(X)
 	YDense := Array2DToDense(Y)
-	epochs := 10
-	regr := &LogisticRegression{}
-	err := regr.Fit(XDense, YDense, epochs)
-	if err != nil {
-		t.Error(err)
+	epochs := 1000
+	regr := &LogisticRegression{
+		LearningRate: .1,
 	}
-	fmt.Println(regr.Weights, regr.Bias, regr.Losses)
-	fmt.Println(YDense, regr.Predict(XDense))
+	regr.Fit(XDense, YDense, epochs, nil)
+	fmt.Println(regr.Weights, regr.Bias)
+	yPred := regr.Predict(XDense)
+	fmt.Println(YDense, yPred, regr.Loss(YDense, yPred))
 }