Finished an implementation of KNN

README.md

@@ -4,6 +4,7 @@ GoLearn
 <img src="http://talks.golang.org/2013/advconc/gopherhat.jpg" width=125>
 
 A small start on a machine learning library in Go.
+[Doc](http://godoc.org/github.com/sjwhitworth/golearn).
 
 Install
 =======

base/domain.go

@@ -4,7 +4,7 @@
 package base
 
 import (
-	mat "github.com/skelterjohn/go.matrix"
+	mat64 "github.com/gonum/matrix/mat64"
 )
 
 // An object that can ingest some data and train on it.
@@ -25,5 +25,6 @@ type Model interface {
 
 // @todo: Implement BaseEstimator setters and getters.
 type BaseEstimator struct {
-	Data *mat.DenseMatrix
+	Estimator
+	Data *mat64.Dense
 }

examples/knnclassifier_iris.go

@@ -3,10 +3,10 @@ package main
 import (
 	"fmt"
 
+	mat64 "github.com/gonum/matrix/mat64"
 	data "github.com/sjwhitworth/golearn/data"
 	knn "github.com/sjwhitworth/golearn/knn"
 	util "github.com/sjwhitworth/golearn/utilities"
-	mat "github.com/skelterjohn/go.matrix"
 )
 
 func main() {
@@ -21,10 +21,10 @@ func main() {
 		randArray := util.RandomArray(3, 7)
 
 		//Initialises a vector with this array
-		random := mat.MakeDenseMatrix(randArray, 1, 3)
+		random := mat64.NewDense(1, 3, randArray)
 
 		//Calculates the Euclidean distance and returns the most popular label
-		labels, _ := cls.Predict(random, 3)
+		labels := cls.Predict(random, 3)
 		fmt.Println(labels)
 	}
 }

examples/knnregressor_random.go

@@ -3,10 +3,10 @@ package main
 import (
 	"fmt"
 
+	"github.com/gonum/matrix/mat64"
 	data "github.com/sjwhitworth/golearn/data"
 	knn "github.com/sjwhitworth/golearn/knn"
 	util "github.com/sjwhitworth/golearn/utilities"
-	mat "github.com/skelterjohn/go.matrix"
 )
 
 func main() {
@@ -22,10 +22,10 @@ func main() {
 		randArray := util.RandomArray(2, 100)
 
 		//Initialises a vector with this array
-		random := mat.MakeDenseMatrix(randArray, 1, 2)
+		random := mat64.NewDense(1, 2, randArray)
 
 		//Calculates the Euclidean distance and returns the most popular label
-		outcome, _ := cls.Predict(random, 3)
+		outcome := cls.Predict(random, 3)
 		fmt.Println(outcome)
 	}
 }

knn/knn.go

@@ -1,52 +1,77 @@
 /* Package KNN implements a K Nearest Neighbors object. It is capable of both classification
-   and regression. It accepts data in the form of a list of float64s, which are then reshaped
+   and regression. It accepts data in the form of a slice of float64s, which are then reshaped
    into a X by Y matrix. */
 
 package knn
 
 import (
+	"github.com/gonum/matrix/mat64"
 	base "github.com/sjwhitworth/golearn/base"
+	pairwiseMetrics "github.com/sjwhitworth/golearn/metrics/pairwise"
 	util "github.com/sjwhitworth/golearn/utilities"
-	mat "github.com/skelterjohn/go.matrix"
 )
 
-//A KNN Classifier. Consists of a data matrix, associated labels in the same order as the matrix, and a name.
+// A KNN Classifier. Consists of a data matrix, associated labels in the same order as the matrix, and a distance function.
+// The accepted distance functions at this time are 'euclidean' and 'manhattan'.
 type KNNClassifier struct {
 	base.BaseEstimator
 	Labels       []string
 	DistanceFunc string
 }
 
-//Mints a new classifier.
-func NewKnnClassifier(labels []string, numbers []float64, x int, y int, distfunc string) *KNNClassifier {
+// Returns a new classifier
+func NewKnnClassifier(labels []string, numbers []float64, rows int, cols int, distfunc string) *KNNClassifier {
+	if rows != len(labels) {
+		panic("Number of rows must equal number of labels")
+	}
+
 	KNN := KNNClassifier{}
-	KNN.Data = mat.MakeDenseMatrix(numbers, x, y)
+	KNN.Data = mat64.NewDense(rows, cols, numbers)
 	KNN.Labels = labels
 	KNN.DistanceFunc = distfunc
 	return &KNN
 }
 
 // Returns a classification for the vector, based on a vector input, using the KNN algorithm.
-// @todo: Lots of room to improve this. V messy.
-func (KNN *KNNClassifier) Predict(vector *mat.DenseMatrix, K int) (string, []int) {
+// See http://en.wikipedia.org/wiki/K-nearest_neighbors_algorithm.
+func (KNN *KNNClassifier) Predict(vector []float64, K int) string {
 
-	rows := KNN.Data.Rows()
+	convertedVector := util.FloatsToMatrix(vector)
+	// Get the number of rows
+	rows, _ := KNN.Data.Dims()
 	rownumbers := make(map[int]float64)
 	labels := make([]string, 0)
 	maxmap := make(map[string]int)
 
+	// Check what distance function we are using
+	switch KNN.DistanceFunc {
+	case "euclidean":
+		{
+			euclidean := pairwiseMetrics.NewEuclidean()
 			for i := 0; i < rows; i++ {
-		row := KNN.Data.GetRowVector(i)
-
-		//Will put code in to check errs later
-		eucdistance, _ := util.ComputeDistance(KNN.DistanceFunc, row, vector)
-		rownumbers[i] = eucdistance
+				row := KNN.Data.RowView(i)
+				rowMat := util.FloatsToMatrix(row)
+				distance := euclidean.Distance(rowMat, convertedVector)
+				rownumbers[i] = distance
+			}
+		}
+	case "manhattan":
+		{
+			manhattan := pairwiseMetrics.NewEuclidean()
+			for i := 0; i < rows; i++ {
+				row := KNN.Data.RowView(i)
+				rowMat := util.FloatsToMatrix(row)
+				distance := manhattan.Distance(rowMat, convertedVector)
+				rownumbers[i] = distance
+			}
+		}
 	}
 
 	sorted := util.SortIntMap(rownumbers)
 	values := sorted[:K]
 
 	for _, elem := range values {
+		// It's when we access this map
 		labels = append(labels, KNN.Labels[elem])
 
 		if _, ok := maxmap[KNN.Labels[elem]]; ok {
@@ -59,5 +84,5 @@ func (KNN *KNNClassifier) Predict(vector *mat.DenseMatrix, K int) (string, []int
 	sortedlabels := util.SortStringMap(maxmap)
 	label := sortedlabels[0]
 
-	return label, values
+	return label
 }

knn/knn_test.go

@@ -0,0 +1,33 @@
+package knn
+
+import (
+	"testing"
+
+	. "github.com/smartystreets/goconvey/convey"
+)
+
+func TestKnnClassifier(t *testing.T) {
+	Convey("Given labels, a classifier and data", t, func() {
+		labels := []string{"blue", "blue", "red", "red"}
+		data := []float64{1, 1, 1, 1, 1, 1, 3, 3, 3, 6, 6, 6}
+		cls := NewKnnClassifier(labels, data, 4, 3, "euclidean")
+
+		Convey("When predicting the label for our first vector", func() {
+			// The vector we're going to predict
+			vector := []float64{1.2, 1.2, 1.5}
+			result := cls.Predict(vector, 2)
+			Convey("The result should be 'blue", func() {
+				So(result, ShouldEqual, "blue")
+			})
+		})
+
+		Convey("When predicting the label for our first vector", func() {
+			// The vector we're going to predict
+			vector2 := []float64{5, 5, 5}
+			result2 := cls.Predict(vector2, 2)
+			Convey("The result should be 'red", func() {
+				So(result2, ShouldEqual, "red")
+			})
+		})
+	})
+}

knn/knnregressor.go

@@ -3,48 +3,70 @@
 package knn
 
 import (
+	"github.com/gonum/matrix/mat64"
+	"github.com/sjwhitworth/golearn/base"
+	pairwiseMetrics "github.com/sjwhitworth/golearn/metrics/pairwise"
 	util "github.com/sjwhitworth/golearn/utilities"
-	mat "github.com/skelterjohn/go.matrix"
 )
 
 //A KNN Regressor. Consists of a data matrix, associated result variables in the same order as the matrix, and a name.
 type KNNRegressor struct {
-	Data         *mat.DenseMatrix
-	Labels       []float64
+	base.BaseEstimator
+	Values       []float64
 	DistanceFunc string
 }
 
-//Mints a new classifier.
-func NewKnnRegressor(labels []float64, numbers []float64, x int, y int, distfunc string) *KNNRegressor {
+// Mints a new classifier.
+func NewKnnRegressor(values []float64, numbers []float64, x int, y int, distfunc string) *KNNRegressor {
 	KNN := KNNRegressor{}
-	KNN.Data = mat.MakeDenseMatrix(numbers, x, y)
-	KNN.Labels = labels
+	KNN.Data = mat64.NewDense(x, y, numbers)
+	KNN.Values = values
+	KNN.DistanceFunc = distfunc
 	return &KNN
 }
 
 //Returns an average of the K nearest labels/variables, based on a vector input.
-func (KNN *KNNRegressor) Predict(vector *mat.DenseMatrix, K int) (float64, []int) {
+func (KNN *KNNRegressor) Predict(vector *mat64.Dense, K int) float64 {
 
-	rows := KNN.Data.Rows()
+	// Get the number of rows
+	rows, _ := KNN.Data.Dims()
 	rownumbers := make(map[int]float64)
-	labels := make([]float64, 1)
-	sum := 0.0
+	labels := make([]float64, 0)
 
+	// Check what distance function we are using
+	switch KNN.DistanceFunc {
+	case "euclidean":
+		{
+			euclidean := pairwiseMetrics.NewEuclidean()
 			for i := 0; i < rows; i++ {
-		row := KNN.Data.GetRowVector(i)
-		eucdistance, _ := util.ComputeDistance(KNN.DistanceFunc, row, vector)
-		rownumbers[i] = eucdistance
+				row := KNN.Data.RowView(i)
+				rowMat := util.FloatsToMatrix(row)
+				distance := euclidean.Distance(rowMat, vector)
+				rownumbers[i] = distance
+			}
+		}
+	case "manhattan":
+		{
+			manhattan := pairwiseMetrics.NewEuclidean()
+			for i := 0; i < rows; i++ {
+				row := KNN.Data.RowView(i)
+				rowMat := util.FloatsToMatrix(row)
+				distance := manhattan.Distance(rowMat, vector)
+				rownumbers[i] = distance
+			}
+		}
 	}
 
 	sorted := util.SortIntMap(rownumbers)
 	values := sorted[:K]
 
+	var sum float64
 	for _, elem := range values {
-		value := KNN.Labels[elem]
+		value := KNN.Values[elem]
 		labels = append(labels, value)
 		sum += value
 	}
 
 	average := sum / float64(K)
-	return average, values
+	return average
 }

knn/testknn.go

@@ -1 +0,0 @@
-package knn

lm/linear_regression.go

@@ -1,7 +1,12 @@
 package lm
 
+<<<<<<< HEAD
 import base "golearn/base"
 
 type LinearModel struct {
 	base.BaseEstimator
+=======
+type LinearModel struct {
+	// base.BaseRegressor
+>>>>>>> 798751c839ea79f6dff13e790f80a4d5bcbfea68
 }

metrics/pairwise/euclidean.go

@@ -0,0 +1,31 @@
+package pairwise
+
+import (
+	"math"
+
+	"github.com/gonum/matrix/mat64"
+)
+
+type Euclidean struct{}
+
+func NewEuclidean() *Euclidean {
+	return &Euclidean{}
+}
+
+// Compute usual inner product in the sense of euclidean.
+func (self *Euclidean) InnerProduct(vectorX *mat64.Dense, vectorY *mat64.Dense) float64 {
+	result := vectorX.Dot(vectorY)
+
+	return result
+}
+
+// Compute usual distance in the sense of euclidean.
+// Also known as L2 distance.
+func (self *Euclidean) Distance(vectorX *mat64.Dense, vectorY *mat64.Dense) float64 {
+	subVector := mat64.NewDense(0, 0, nil)
+	subVector.Sub(vectorX, vectorY)
+
+	result := self.InnerProduct(subVector, subVector)
+
+	return math.Sqrt(result)
+}

metrics/pairwise/euclidean_test.go

@@ -0,0 +1,36 @@
+package pairwise
+
+import (
+	"testing"
+
+	"github.com/gonum/matrix/mat64"
+	. "github.com/smartystreets/goconvey/convey"
+)
+
+func TestEuclidean(t *testing.T) {
+	var vectorX, vectorY *mat64.Dense
+	euclidean := NewEuclidean()
+
+	Convey("Given two vectors", t, func() {
+		vectorX = mat64.NewDense(3, 1, []float64{1, 2, 3})
+		vectorY = mat64.NewDense(3, 1, []float64{2, 4, 5})
+
+		Convey("When doing inner product", func() {
+			result := euclidean.InnerProduct(vectorX, vectorY)
+
+			Convey("The result should be 25", func() {
+				So(result, ShouldEqual, 25)
+			})
+		})
+
+		Convey("When calculating distance", func() {
+			result := euclidean.Distance(vectorX, vectorY)
+
+			Convey("The result should be 3", func() {
+				So(result, ShouldEqual, 3)
+			})
+
+		})
+
+	})
+}

metrics/pairwise/manhattan.go

@@ -0,0 +1,40 @@
+package pairwise
+
+import (
+	"math"
+
+	"github.com/gonum/matrix/mat64"
+)
+
+type Manhattan struct{}
+
+func NewManhattan() *Manhattan {
+	return &Manhattan{}
+}
+
+// Manhattan distance, also known as L1 distance.
+// Compute sum of absolute values of elements.
+func (self *Manhattan) Distance(vectorX *mat64.Dense, vectorY *mat64.Dense) float64 {
+	var length int
+	subVector := mat64.NewDense(0, 0, nil)
+	subVector.Sub(vectorX, vectorY)
+
+	r, c := subVector.Dims()
+
+	if r == 1 {
+		// Force transpose to column vector
+		subVector.TCopy(subVector)
+		length = c
+	} else if c == 1 {
+		length = r
+	} else {
+		panic(mat64.ErrShape)
+	}
+
+	result := .0
+	for i := 0; i < length; i++ {
+		result += math.Abs(subVector.At(i, 0))
+	}
+
+	return result
+}

metrics/pairwise/manhattan_test.go

@@ -0,0 +1,42 @@
+package pairwise
+
+import (
+	"testing"
+
+	"github.com/gonum/matrix/mat64"
+	. "github.com/smartystreets/goconvey/convey"
+)
+
+func TestManhattan(t *testing.T) {
+	var vectorX, vectorY *mat64.Dense
+	manhattan := NewManhattan()
+
+	Convey("Given two vectors", t, func() {
+		vectorX = mat64.NewDense(3, 1, []float64{2, 2, 3})
+		vectorY = mat64.NewDense(3, 1, []float64{1, 4, 5})
+
+		Convey("When calculating distance with column vectors", func() {
+			result := manhattan.Distance(vectorX, vectorY)
+
+			Convey("The result should be 5", func() {
+				So(result, ShouldEqual, 5)
+			})
+		})
+
+		Convey("When calculating distance with row vectors", func() {
+			vectorX.TCopy(vectorX)
+			vectorY.TCopy(vectorY)
+			result := manhattan.Distance(vectorX, vectorY)
+
+			Convey("The result should be 5", func() {
+				So(result, ShouldEqual, 5)
+			})
+		})
+
+		Convey("When calculating distance with row and column vectors", func() {
+			vectorX.TCopy(vectorX)
+			So(func() { manhattan.Distance(vectorX, vectorY) }, ShouldPanicWith, mat64.ErrShape)
+		})
+
+	})
+}

metrics/pairwise/pairwise.go

@@ -0,0 +1,2 @@
+// Package pairwise implements utilities to evaluate pairwise distances or inner product (via kernel).
+package pairwise

metrics/pairwise/poly_kernel.go

@@ -0,0 +1,34 @@
+package pairwise
+
+import (
+	"math"
+
+	"github.com/gonum/matrix/mat64"
+)
+
+type PolyKernel struct {
+	degree int
+}
+
+// Return a d-degree polynomial kernel
+func NewPolyKernel(degree int) *PolyKernel {
+	return &PolyKernel{degree: degree}
+}
+
+// Compute inner product through kernel trick
+// K(x, y) = (x^T y + 1)^d
+func (self *PolyKernel) InnerProduct(vectorX *mat64.Dense, vectorY *mat64.Dense) float64 {
+	result := vectorX.Dot(vectorY)
+	result = math.Pow(result+1, float64(self.degree))
+
+	return result
+}
+
+// Compute distance under the polynomial kernel, maybe no need.
+func (self *PolyKernel) Distance(vectorX *mat64.Dense, vectorY *mat64.Dense) float64 {
+	subVector := mat64.NewDense(0, 0, nil)
+	subVector.Sub(vectorX, vectorY)
+	result := self.InnerProduct(subVector, subVector)
+
+	return math.Sqrt(result)
+}

metrics/pairwise/poly_kernel_test.go

@@ -0,0 +1,36 @@
+package pairwise
+
+import (
+	"testing"
+
+	"github.com/gonum/matrix/mat64"
+	. "github.com/smartystreets/goconvey/convey"
+)
+
+func TestPolyKernel(t *testing.T) {
+	var vectorX, vectorY *mat64.Dense
+	polyKernel := NewPolyKernel(3)
+
+	Convey("Given two vectors", t, func() {
+		vectorX = mat64.NewDense(3, 1, []float64{1, 2, 3})
+		vectorY = mat64.NewDense(3, 1, []float64{2, 4, 5})
+
+		Convey("When doing inner product", func() {
+			result := polyKernel.InnerProduct(vectorX, vectorY)
+
+			Convey("The result should be 17576", func() {
+				So(result, ShouldEqual, 17576)
+			})
+		})
+
+		Convey("When calculating distance", func() {
+			result := polyKernel.Distance(vectorX, vectorY)
+
+			Convey("The result should be 31.622776601683793", func() {
+				So(result, ShouldEqual, 31.622776601683793)
+			})
+
+		})
+
+	})
+}

metrics/pairwise/rbf_kernel.go

@@ -0,0 +1,27 @@
+package pairwise
+
+import (
+	"math"
+
+	"github.com/gonum/matrix/mat64"
+)
+
+type RBFKernel struct {
+	gamma float64
+}
+
+// Radial Basis Function Kernel
+func NewRBFKernel(gamma float64) *RBFKernel {
+	return &RBFKernel{gamma: gamma}
+}
+
+// Compute inner product through kernel trick
+// K(x, y) = exp(-gamma * ||x - y||^2)
+func (self *RBFKernel) InnerProduct(vectorX *mat64.Dense, vectorY *mat64.Dense) float64 {
+	euclidean := NewEuclidean()
+	distance := euclidean.Distance(vectorX, vectorY)
+
+	result := math.Exp(-self.gamma * math.Pow(distance, 2))
+
+	return result
+}

metrics/pairwise/rbf_kernel_test.go

@@ -0,0 +1,28 @@
+package pairwise
+
+import (
+	"testing"
+
+	"github.com/gonum/matrix/mat64"
+	. "github.com/smartystreets/goconvey/convey"
+)
+
+func TestRBFKernel(t *testing.T) {
+	var vectorX, vectorY *mat64.Dense
+	rbfKernel := NewRBFKernel(0.1)
+
+	Convey("Given two vectors", t, func() {
+		vectorX = mat64.NewDense(3, 1, []float64{1, 2, 3})
+		vectorY = mat64.NewDense(3, 1, []float64{2, 4, 5})
+
+		Convey("When doing inner product", func() {
+			result := rbfKernel.InnerProduct(vectorX, vectorY)
+
+			Convey("The result should be 0.4065696597405991", func() {
+				So(result, ShouldEqual, 0.4065696597405991)
+
+			})
+		})
+
+	})
+}

utilities/distance.go

@@ -1,36 +0,0 @@
-package utilities
-
-import (
-	"fmt"
-	"math"
-
-	mat "github.com/skelterjohn/go.matrix"
-)
-
-// Computes the 'distance' between two vectors, where the distance is one of the following methods -
-// euclidean (more to come)
-func ComputeDistance(metric string, vector *mat.DenseMatrix, testrow *mat.DenseMatrix) (float64, error) {
-	var sum float64
-
-	switch metric {
-	case "euclidean":
-		{
-			difference, err := testrow.MinusDense(vector)
-			flat := difference.Array()
-
-			if err != nil {
-				fmt.Println(err)
-			}
-
-			for _, i := range flat {
-				squared := math.Pow(i, 2)
-				sum += squared
-			}
-
-			eucdistance := math.Sqrt(sum)
-			return eucdistance, nil
-		}
-	default:
-		return 0.0, fmt.Errorf("ValueError: %s is not an implemented distance method", metric)
-	}
-}

utilities/utilities.go

@@ -5,6 +5,8 @@ import (
 	rand "math/rand"
 	"sort"
 	"strconv"
+
+	mat64 "github.com/gonum/matrix/mat64"
 )
 
 type sortedIntMap struct {
@@ -88,3 +90,7 @@ func ConvertLabelsToFloat(labels []string) []float64 {
 	}
 	return floats
 }
+
+func FloatsToMatrix(floats []float64) *mat64.Dense {
+	return mat64.NewDense(1, len(floats), floats)
+}