1
0
mirror of https://github.com/sjwhitworth/golearn.git synced 2025-04-26 13:49:14 +08:00

Removing Clutter

Partial Modularization of best split method. Shorten method by declaring variables in same line as well.

Also removing redundant functions, and adding into cart_utils.
This commit is contained in:
Ayush 2020-07-28 14:17:18 +05:30
parent ef751e62c4
commit 2d2af0a58f
4 changed files with 181 additions and 258 deletions

View File

@ -90,6 +90,16 @@ func entropy(y []int64, labels []int64) (float64, int64) {
return entropy, maxLabel
}
func calculateClassificationLoss(y []int64, labels []int64, criterion string) (float64, int64) {
if criterion == GINI {
return giniImpurity(y, labels)
} else if criterion == ENTROPY {
return entropy(y, labels)
} else {
panic("Invalid impurity function, choose from GINI or ENTROPY")
}
}
// Split the data into left node and right node based on feature and threshold
func classifierCreateSplit(data [][]float64, feature int64, y []int64, threshold float64) ([][]float64, [][]float64, []int64, []int64) {
var left [][]float64
@ -111,37 +121,6 @@ func classifierCreateSplit(data [][]float64, feature int64, y []int64, threshold
return left, right, lefty, righty
}
// Helper Function to check if data point is unique or not.
// We will use this to isolate unique values of a feature
func classifierStringInSlice(a float64, list []float64) bool {
for _, b := range list {
if b == a {
return true
}
}
return false
}
// Isolate only unique values. This way, we can try only unique splits and not redundant ones.
func classifierFindUnique(data []float64) []float64 {
var unique []float64
for i := range data {
if !classifierStringInSlice(data[i], unique) {
unique = append(unique, data[i])
}
}
return unique
}
// Isolate only the feature being considered for splitting. Reduces the complexity in managing splits.
func classifierGetFeature(data [][]float64, feature int64) []float64 {
var featureVals []float64
for i := range data {
featureVals = append(featureVals, data[i][feature])
}
return featureVals
}
// Function to Create New Decision Tree Classifier.
// It assigns all of the hyperparameters by user into the tree attributes.
func NewDecisionTreeClassifier(criterion string, maxDepth int64, labels []int64) *CARTDecisionTreeClassifier {
@ -153,19 +132,6 @@ func NewDecisionTreeClassifier(criterion string, maxDepth int64, labels []int64)
return &tree
}
// Make sure that split being considered has not been done before.
// Else we will unnecessarily try splits that won't improve Impurity.
func classifierValidate(triedSplits [][]float64, feature int64, threshold float64) bool {
for i := range triedSplits {
split := triedSplits[i]
featureTried, thresholdTried := split[0], split[1]
if int64(featureTried) == feature && thresholdTried == threshold {
return false
}
}
return true
}
// Reorder the data by feature being considered. Optimizes code by reducing the number of times we have to loop over data for splitting
func classifierReOrderData(featureVal []float64, data [][]float64, y []int64) ([][]float64, []int64) {
s := NewSlice(featureVal)
@ -202,7 +168,7 @@ func classifierUpdateSplit(left [][]float64, lefty []int64, right [][]float64, r
func (tree *CARTDecisionTreeClassifier) Fit(X base.FixedDataGrid) {
var emptyNode classifierNode
data := classifierConvertInstancesToProblemVec(X)
data := convertInstancesToProblemVec(X)
y := classifierConvertInstancesToLabelVec(X)
emptyNode = classifierBestSplit(*tree, data, y, tree.labels, emptyNode, tree.criterion, tree.maxDepth, 0)
@ -221,40 +187,29 @@ func classifierBestSplit(tree CARTDecisionTreeClassifier, data [][]float64, y []
}
numFeatures := len(data[0])
var bestGini float64
var origGini float64
var bestGini, origGini float64
// Calculate loss based on Criterion Specified by user
if criterion == GINI {
origGini, upperNode.LeftLabel = giniImpurity(y, labels)
} else if criterion == ENTROPY {
origGini, upperNode.LeftLabel = entropy(y, labels)
} else {
panic("Invalid impurity function, choose from GINI or ENTROPY")
}
origGini, upperNode.LeftLabel = calculateClassificationLoss(y, labels, criterion)
bestGini = origGini
bestLeft := data
bestRight := data
bestLefty := y
bestRighty := y
bestLeft, bestRight, bestLefty, bestRighty := data, data, y, y
numData := len(data)
bestLeftGini := bestGini
bestRightGini := bestGini
bestLeftGini, bestRightGini := bestGini, bestGini
upperNode.Use_not = true
var leftN classifierNode
var rightN classifierNode
var leftN, rightN classifierNode
// Iterate over all features
for i := 0; i < numFeatures; i++ {
featureVal := classifierGetFeature(data, int64(i))
unique := classifierFindUnique(featureVal)
featureVal := getFeature(data, int64(i))
unique := findUnique(featureVal)
sort.Float64s(unique)
numUnique := len(unique)
sortData, sortY := classifierReOrderData(featureVal, data, y)
@ -263,53 +218,43 @@ func classifierBestSplit(tree CARTDecisionTreeClassifier, data [][]float64, y []
var left, right [][]float64
var lefty, righty []int64
// Iterate over all possible thresholds for that feature
for j := range unique {
if j != (numUnique - 1) {
threshold := (unique[j] + unique[j+1]) / 2
// Ensure that same split has not been made before
if classifierValidate(tree.triedSplits, int64(i), threshold) {
// We need to split data from fresh when considering new feature for the first time.
// Otherwise, we need to update the split by moving data points from left to right.
if firstTime {
left, right, lefty, righty = classifierCreateSplit(sortData, int64(i), sortY, threshold)
firstTime = false
} else {
left, lefty, right, righty = classifierUpdateSplit(left, lefty, right, righty, int64(i), threshold)
}
for j := 0; j < len(unique)-1; j++ {
var leftGini float64
var rightGini float64
var leftLabels int64
var rightLabels int64
if criterion == GINI {
leftGini, leftLabels = giniImpurity(lefty, labels)
rightGini, rightLabels = giniImpurity(righty, labels)
} else if criterion == ENTROPY {
leftGini, leftLabels = entropy(lefty, labels)
rightGini, rightLabels = entropy(righty, labels)
}
// Calculate weighted gini impurity of child nodes
subGini := (leftGini * float64(len(left)) / float64(numData)) + (rightGini * float64(len(right)) / float64(numData))
// If we find a split that reduces impurity
if subGini < bestGini {
bestGini = subGini
bestLeft = left
bestRight = right
bestLefty = lefty
bestRighty = righty
upperNode.Threshold = threshold
upperNode.Feature = int64(i)
upperNode.LeftLabel = leftLabels
upperNode.RightLabel = rightLabels
bestLeftGini = leftGini
bestRightGini = rightGini
}
threshold := (unique[j] + unique[j+1]) / 2
// Ensure that same split has not been made before
if validate(tree.triedSplits, int64(i), threshold) {
// We need to split data from fresh when considering new feature for the first time.
// Otherwise, we need to update the split by moving data points from left to right.
if firstTime {
left, right, lefty, righty = classifierCreateSplit(sortData, int64(i), sortY, threshold)
firstTime = false
} else {
left, lefty, right, righty = classifierUpdateSplit(left, lefty, right, righty, int64(i), threshold)
}
var leftGini, rightGini float64
var leftLabels, rightLabels int64
leftGini, leftLabels = calculateClassificationLoss(lefty, labels, criterion)
rightGini, rightLabels = calculateClassificationLoss(righty, labels, criterion)
// Calculate weighted gini impurity of child nodes
subGini := (leftGini * float64(len(left)) / float64(numData)) + (rightGini * float64(len(right)) / float64(numData))
// If we find a split that reduces impurity
if subGini < bestGini {
bestGini = subGini
bestLeft, bestRight = left, right
bestLefty, bestRighty = lefty, righty
upperNode.Threshold, upperNode.Feature = threshold, int64(i)
upperNode.LeftLabel, upperNode.RightLabel = leftLabels, rightLabels
bestLeftGini, bestRightGini = leftGini, rightGini
}
}
}
}
@ -366,10 +311,8 @@ func classifierPrintTreeFromNode(tree classifierNode, spacing string) string {
returnString += spacing + "---> True" + "\n"
returnString += " " + spacing + "PREDICT "
returnString += strconv.FormatInt(tree.LeftLabel, 10) + "\n"
}
if tree.Right == nil {
returnString += spacing + "---> False" + "\n"
returnString += " " + spacing + "PREDICT "
returnString += strconv.FormatInt(tree.RightLabel, 10) + "\n"
@ -409,7 +352,7 @@ func classifierPredictSingle(tree classifierNode, instance []float64) int64 {
// Given test data, return predictions for every datapoint. calls classifierPredictFromNode
func (tree *CARTDecisionTreeClassifier) Predict(X_test base.FixedDataGrid) []int64 {
root := *tree.RootNode
test := classifierConvertInstancesToProblemVec(X_test)
test := convertInstancesToProblemVec(X_test)
return classifierPredictFromNode(root, test)
}
@ -429,7 +372,7 @@ func classifierPredictFromNode(tree classifierNode, test [][]float64) []int64 {
// Calls classifierEvaluateFromNode
func (tree *CARTDecisionTreeClassifier) Evaluate(test base.FixedDataGrid) float64 {
rootNode := *tree.RootNode
xTest := classifierConvertInstancesToProblemVec(test)
xTest := convertInstancesToProblemVec(test)
yTest := classifierConvertInstancesToLabelVec(test)
return classifierEvaluateFromNode(rootNode, xTest, yTest)
}
@ -447,31 +390,6 @@ func classifierEvaluateFromNode(tree classifierNode, xTest [][]float64, yTest []
return accuracy
}
// Helper function to convert base.FixedDataGrid into required format. Called in Fit, Predict
func classifierConvertInstancesToProblemVec(X base.FixedDataGrid) [][]float64 {
// Allocate problem array
_, rows := X.Size()
problemVec := make([][]float64, rows)
// Retrieve numeric non-class Attributes
numericAttrs := base.NonClassFloatAttributes(X)
numericAttrSpecs := base.ResolveAttributes(X, numericAttrs)
// Convert each row
X.MapOverRows(numericAttrSpecs, func(row [][]byte, rowNo int) (bool, error) {
// Allocate a new row
probRow := make([]float64, len(numericAttrSpecs))
// Read out the row
for i, _ := range numericAttrSpecs {
probRow[i] = base.UnpackBytesToFloat(row[i])
}
// Add the row
problemVec[rowNo] = probRow
return true, nil
})
return problemVec
}
// Helper function to convert base.FixedDataGrid into required format. Called in Fit, Predict
func classifierConvertInstancesToLabelVec(X base.FixedDataGrid) []int64 {
// Get the class Attributes

View File

@ -81,6 +81,16 @@ func mseImpurity(y []float64) (float64, float64) {
return meanSquaredError(y, yHat), yHat
}
func calculateRegressionLoss(y []float64, criterion string) (float64, float64) {
if criterion == MAE {
return maeImpurity(y)
} else if criterion == MSE {
return mseImpurity(y)
} else {
panic("Invalid impurity function, choose from MAE or MSE")
}
}
// Split the data into left and right based on trehsold and feature.
func regressorCreateSplit(data [][]float64, feature int64, y []float64, threshold float64) ([][]float64, [][]float64, []float64, []float64) {
var left [][]float64
@ -102,39 +112,6 @@ func regressorCreateSplit(data [][]float64, feature int64, y []float64, threshol
return left, right, lefty, righty
}
// Helper function for finding unique values.
// Used for isolating unique values in a feature.
func regressorStringInSlice(a float64, list []float64) bool {
for _, b := range list {
if b == a {
return true
}
}
return false
}
// Isolate only unique values.
// This way we can only try unique splits.
func regressorFindUnique(data []float64) []float64 {
var unique []float64
for i := range data {
if !regressorStringInSlice(data[i], unique) {
unique = append(unique, data[i])
}
}
return unique
}
// Extract out a single feature from data.
// Reduces complexity in managing splits and sorting
func regressorGetFeature(data [][]float64, feature int64) []float64 {
var featureVals []float64
for i := range data {
featureVals = append(featureVals, data[i][feature])
}
return featureVals
}
// Interface for creating new Decision Tree Regressor
func NewDecisionTreeRegressor(criterion string, maxDepth int64) *CARTDecisionTreeRegressor {
var tree CARTDecisionTreeRegressor
@ -143,19 +120,6 @@ func NewDecisionTreeRegressor(criterion string, maxDepth int64) *CARTDecisionTre
return &tree
}
// Validate that the split being tested has not been done before.
// This prevents redundant splits from hapenning.
func regressorValidate(triedSplits [][]float64, feature int64, threshold float64) bool {
for i := range triedSplits {
split := triedSplits[i]
featureTried, thresholdTried := split[0], split[1]
if int64(featureTried) == feature && thresholdTried == threshold {
return false
}
}
return true
}
// Re order data based on a feature for optimizing code
// Helps in updating splits without reiterating entire dataset
func regressorReOrderData(featureVal []float64, data [][]float64, y []float64) ([][]float64, []float64) {
@ -204,6 +168,7 @@ func (tree *CARTDecisionTreeRegressor) Fit(X base.FixedDataGrid) {
// Recursive function - stops if maxDepth is reached or nodes are pure
func regressorBestSplit(tree CARTDecisionTreeRegressor, data [][]float64, y []float64, upperNode regressorNode, criterion string, maxDepth int64, depth int64) regressorNode {
// Ensure that we have not reached maxDepth. maxDepth =-1 means split until nodes are pure
depth++
if depth > maxDepth && maxDepth != -1 {
@ -211,39 +176,27 @@ func regressorBestSplit(tree CARTDecisionTreeRegressor, data [][]float64, y []fl
}
numFeatures := len(data[0])
var bestLoss float64
var origLoss float64
var bestLoss, origLoss float64
if criterion == MAE {
origLoss, upperNode.LeftPred = maeImpurity(y)
} else if criterion == MSE {
origLoss, upperNode.LeftPred = mseImpurity(y)
} else {
panic("Invalid impurity function, choose from MAE or MSE")
}
origLoss, upperNode.LeftPred = calculateRegressionLoss(y, criterion)
bestLoss = origLoss
bestLeft := data
bestRight := data
bestLefty := y
bestRighty := y
bestLeft, bestRight, bestLefty, bestRighty := data, data, y, y
numData := len(data)
bestLeftLoss := bestLoss
bestRightLoss := bestLoss
bestLeftLoss, bestRightLoss := bestLoss, bestLoss
upperNode.Use_not = true
var leftN regressorNode
var rightN regressorNode
var leftN, rightN regressorNode
// Iterate over all features
for i := 0; i < numFeatures; i++ {
featureVal := regressorGetFeature(data, int64(i))
unique := regressorFindUnique(featureVal)
featureVal := getFeature(data, int64(i))
unique := findUnique(featureVal)
sort.Float64s(unique)
numUnique := len(unique)
sortData, sortY := regressorReOrderData(featureVal, data, y)
@ -252,49 +205,36 @@ func regressorBestSplit(tree CARTDecisionTreeRegressor, data [][]float64, y []fl
var left, right [][]float64
var lefty, righty []float64
for j := range unique {
if j != (numUnique - 1) {
threshold := (unique[j] + unique[j+1]) / 2
if regressorValidate(tree.triedSplits, int64(i), threshold) {
if firstTime {
left, right, lefty, righty = regressorCreateSplit(sortData, int64(i), sortY, threshold)
firstTime = false
} else {
left, lefty, right, righty = regressorUpdateSplit(left, lefty, right, righty, int64(i), threshold)
}
var leftLoss float64
var rightLoss float64
var leftPred float64
var rightPred float64
if criterion == MAE {
leftLoss, leftPred = maeImpurity(lefty)
rightLoss, rightPred = maeImpurity(righty)
} else if criterion == MSE {
leftLoss, leftPred = mseImpurity(lefty)
rightLoss, rightPred = mseImpurity(righty)
}
subLoss := (leftLoss * float64(len(left)) / float64(numData)) + (rightLoss * float64(len(right)) / float64(numData))
if subLoss < bestLoss {
bestLoss = subLoss
bestLeft = left
bestRight = right
bestLefty = lefty
bestRighty = righty
upperNode.Threshold = threshold
upperNode.Feature = int64(i)
upperNode.LeftPred = leftPred
upperNode.RightPred = rightPred
bestLeftLoss = leftLoss
bestRightLoss = rightLoss
}
for j := 0; j < len(unique)-1; j++ {
threshold := (unique[j] + unique[j+1]) / 2
if validate(tree.triedSplits, int64(i), threshold) {
if firstTime {
left, right, lefty, righty = regressorCreateSplit(sortData, int64(i), sortY, threshold)
firstTime = false
} else {
left, lefty, right, righty = regressorUpdateSplit(left, lefty, right, righty, int64(i), threshold)
}
var leftLoss, rightLoss float64
var leftPred, rightPred float64
leftLoss, leftPred = calculateRegressionLoss(lefty, criterion)
rightLoss, rightPred = calculateRegressionLoss(righty, criterion)
subLoss := (leftLoss * float64(len(left)) / float64(numData)) + (rightLoss * float64(len(right)) / float64(numData))
if subLoss < bestLoss {
bestLoss = subLoss
bestLeft, bestRight = left, right
bestLefty, bestRighty = lefty, righty
upperNode.Threshold, upperNode.Feature = threshold, int64(i)
upperNode.LeftPred, upperNode.RightPred = leftPred, rightPred
bestLeftLoss, bestRightLoss = leftLoss, rightLoss
}
}
}
}
@ -312,19 +252,16 @@ func regressorBestSplit(tree CARTDecisionTreeRegressor, data [][]float64, y []fl
if leftN.Use_not == true {
upperNode.Left = &leftN
}
}
if bestRightLoss > 0 {
tree.triedSplits = append(tree.triedSplits, []float64{float64(upperNode.Feature), upperNode.Threshold})
rightN = regressorBestSplit(tree, bestRight, bestRighty, rightN, criterion, maxDepth, depth)
if rightN.Use_not == true {
upperNode.Right = &rightN
}
}
}
return upperNode
}
@ -349,20 +286,17 @@ func regressorPrintTreeFromNode(tree regressorNode, spacing string) string {
returnString += fmt.Sprintf("%.3f", tree.LeftPred) + "\n"
}
if tree.Right == nil {
returnString += spacing + "---> False" + "\n"
returnString += " " + spacing + "PREDICT "
returnString += fmt.Sprintf("%.3f", tree.RightPred) + "\n"
}
if tree.Left != nil {
// fmt.Println(spacing + "---> True")
returnString += spacing + "---> True" + "\n"
returnString += regressorPrintTreeFromNode(*tree.Left, spacing+" ")
}
if tree.Right != nil {
// fmt.Println(spacing + "---> False")
returnString += spacing + "---> False" + "\n"
returnString += regressorPrintTreeFromNode(*tree.Right, spacing+" ")
}

View File

@ -38,10 +38,10 @@ func TestRegressor(t *testing.T) {
So(len(righty), ShouldEqual, 2)
// Is isolating unique values working properly
So(len(classifierFindUnique([]float64{10, 1, 1})), ShouldEqual, 2)
So(len(findUnique([]float64{10, 1, 1})), ShouldEqual, 2)
// is data reordered correctly
orderedData, orderedY := classifierReOrderData(classifierGetFeature(classifierData, 1), classifierData, classifiery)
orderedData, orderedY := classifierReOrderData(getFeature(classifierData, 1), classifierData, classifiery)
fmt.Println(orderedData)
fmt.Println(orderedY)
So(orderedData[1][1], ShouldEqual, 3.0)
@ -85,11 +85,8 @@ func TestRegressor(t *testing.T) {
So(len(rightData), ShouldEqual, 2)
So(len(righty), ShouldEqual, 2)
// Is isolating unique values working properly
So(len(regressorFindUnique([]float64{10, 1, 1})), ShouldEqual, 2)
// is data reordered correctly
regressorOrderedData, regressorOrderedY := regressorReOrderData(regressorGetFeature(data, 1), data, y)
regressorOrderedData, regressorOrderedY := regressorReOrderData(getFeature(data, 1), data, y)
So(regressorOrderedData[1][1], ShouldEqual, 3.0)
So(regressorOrderedY[0], ShouldEqual, 2)

74
trees/cart_utils.go Normal file
View File

@ -0,0 +1,74 @@
package trees
import (
"github.com/sjwhitworth/golearn/base"
)
// Helper Function to check if data point is unique or not.
// We will use this to isolate unique values of a feature
func stringInSlice(a float64, list []float64) bool {
for _, b := range list {
if b == a {
return true
}
}
return false
}
// Isolate only unique values. This way, we can try only unique splits and not redundant ones.
func findUnique(data []float64) []float64 {
var unique []float64
for i := range data {
if !stringInSlice(data[i], unique) {
unique = append(unique, data[i])
}
}
return unique
}
// Isolate only the feature being considered for splitting. Reduces the complexity in managing splits.
func getFeature(data [][]float64, feature int64) []float64 {
var featureVals []float64
for i := range data {
featureVals = append(featureVals, data[i][feature])
}
return featureVals
}
// Make sure that split being considered has not been done before.
// Else we will unnecessarily try splits that won't improve Impurity.
func validate(triedSplits [][]float64, feature int64, threshold float64) bool {
for i := range triedSplits {
split := triedSplits[i]
featureTried, thresholdTried := split[0], split[1]
if int64(featureTried) == feature && thresholdTried == threshold {
return false
}
}
return true
}
// Helper function to convert base.FixedDataGrid into required format. Called in Fit, Predict
func convertInstancesToProblemVec(X base.FixedDataGrid) [][]float64 {
// Allocate problem array
_, rows := X.Size()
problemVec := make([][]float64, rows)
// Retrieve numeric non-class Attributes
numericAttrs := base.NonClassFloatAttributes(X)
numericAttrSpecs := base.ResolveAttributes(X, numericAttrs)
// Convert each row
X.MapOverRows(numericAttrSpecs, func(row [][]byte, rowNo int) (bool, error) {
// Allocate a new row
probRow := make([]float64, len(numericAttrSpecs))
// Read out the row
for i, _ := range numericAttrSpecs {
probRow[i] = base.UnpackBytesToFloat(row[i])
}
// Add the row
problemVec[rowNo] = probRow
return true, nil
})
return problemVec
}