multiclass log reg

regr/src/LogisticRegression.go

@@ -34,7 +34,7 @@ func (regr *LogisticRegression) Loss(yTrue, yPred mat.Matrix) float64 {
 	sum := &mat.Dense{}
 	sum.Add(y1, y2)
 	w, h := yTrue.Dims()
-	return mat.Sum(sum) / float64(w*h)
+	return -(mat.Sum(sum) / float64(w*h))
 }
 
 func (regr *LogisticRegression) forward(X mat.Matrix) mat.Matrix {
@@ -76,8 +76,8 @@ func (regr *LogisticRegression) backprop(x, y mat.Matrix) float64 {
 	return loss
 }
 
-func (regr *LogisticRegression) Fit(X, Y mat.Matrix, epochs int, losses *[]float64) {
-	for i := 0; i < epochs; i++ {
+func (regr *LogisticRegression) Fit(X, Y mat.Matrix, losses *[]float64) {
+	for i := 0; i < regr.Epochs; i++ {
 		regr.backprop(X, Y)
 		if losses != nil {
 			*losses = append(*losses, regr.Loss(Y, regr.forward(X)))

regr/src/LogisticRegression_test.go

@@ -6,16 +6,55 @@ import (
 )
 
 func TestLogisticRegression(t *testing.T) {
-	X := [][]float64{{.1, .1, .1}, {.2, .2, .2}, {.1, .1, .1}, {.2, .2, .2}}
-	Y := [][]float64{{0}, {1}, {0}, {1}}
+	X := [][]float64{
+		{.1, .1, .1},
+		{.2, .2, .2},
+		{.1, .1, .1},
+		{.2, .2, .2},
+	}
+	//Y := [][]float64{{1}, {0}, {1}, {0}}
+	Y := [][]float64{
+		{0},
+		{1},
+		{0},
+		{1},
+	}
 	XDense := Array2DToDense(X)
 	YDense := Array2DToDense(Y)
 	epochs := 1000
 	regr := &LogisticRegression{
-		LearningRate: .1,
+		Epochs:       epochs,
+		LearningRate: .01,
 	}
-	regr.Fit(XDense, YDense, epochs, nil)
+	regr.Fit(XDense, YDense, nil)
 	fmt.Println(regr.Weights, regr.Bias)
 	yPred := regr.Predict(XDense)
 	fmt.Println(YDense, yPred, regr.Loss(YDense, yPred))
 }
+
+func TestMCLogisticRegression(t *testing.T) {
+	X := [][]float64{
+		{.1, .1, .1},
+		{.2, .2, .2},
+		{.1, .1, .1},
+		{.2, .2, .2},
+	}
+	//Y := [][]float64{{1}, {0}, {1}, {0}}
+	Y := [][]float64{
+		{0, 1},
+		{1, 0},
+		{0, 1},
+		{1, 0},
+	}
+	XDense := Array2DToDense(X)
+	YDense := Array2DToDense(Y)
+	epochs := 1000
+	regr := &MCLogisticRegression{
+		Epochs:       epochs,
+		LearningRate: .01,
+	}
+	regr.Fit(XDense, YDense)
+	// fmt.Println(regr.Weights, regr.Bias)
+	yPred := regr.Predict(XDense)
+	fmt.Println(YDense, yPred) //, regr.Loss(YDense, yPred))
+}

regr/src/MCLogisticRegression.go

@@ -0,0 +1,44 @@
+package src
+
+import (
+	"sync"
+
+	"gonum.org/v1/gonum/mat"
+)
+
+type MCLogisticRegression struct {
+	Epochs       int
+	LearningRate float64
+	Models       []*LogisticRegression
+	cols         int
+	rows         int
+}
+
+func (regr *MCLogisticRegression) Fit(x, y mat.Matrix) {
+	regr.rows, regr.cols = y.Dims()
+	regr.Models = make([]*LogisticRegression, regr.cols)
+	for j := 0; j < regr.cols; j++ {
+		regr.Models[j] = &LogisticRegression{
+			Epochs:       regr.Epochs,
+			LearningRate: regr.LearningRate,
+		}
+		yj := mat.Col(nil, j, y)
+		Y := mat.NewDense(len(yj), 1, yj)
+		regr.Models[j].Fit(x, Y, nil)
+	}
+}
+
+func (regr *MCLogisticRegression) Predict(x mat.Matrix) mat.Matrix {
+	probs := mat.NewDense(regr.rows, regr.cols, nil)
+	wg := sync.WaitGroup{}
+	wg.Add(regr.cols)
+	for j := 0; j < regr.cols; j++ {
+		go func(j int) {
+			pred := regr.Models[j].Predict(x).(*mat.Dense)
+			probs.SetCol(j, mat.Col(nil, 0, pred))
+			wg.Done()
+		}(j)
+	}
+	wg.Wait()
+	return probs
+}