func SaveLibSVMDataset(path string, set data.Dataset) {
log.Print("保存数据集到libsvm格式文件", path)
f, err := os.Create(path)
defer f.Close()
if err != nil {
log.Fatalf("无法打开文件\"%v\"，错误提示：%v\n", path, err)
}
w := bufio.NewWriter(f)
defer w.Flush()
iter := set.CreateIterator()
iter.Start()
for !iter.End() {
instance := iter.GetInstance()
if instance.Output.LabelString == "" {
fmt.Fprintf(w, "%d ", instance.Output.Label)
} else {
fmt.Fprintf(w, "%s ", instance.Output.LabelString)
}
for _, k := range instance.Features.Keys() {
// 跳过第0个特征，因为它始终是1
if k == 0 {
continue
}
if instance.Features.Get(k) != 0 {
// libsvm格式的特征从1开始
fmt.Fprintf(w, "%d:%s ", k, strconv.FormatFloat(instance.Features.Get(k), 'f', -1, 64))
}
}
fmt.Fprint(w, "\n")
iter.Next()
}
}
func (rbm *RBM) feeder(set data.Dataset, ch chan *data.Instance) {
iter := set.CreateIterator()
iter.Start()
for it := 0; it < set.NumInstances(); it++ {
instance := iter.GetInstance()
ch <- instance
iter.Next()
}
}
// 输出的度量名字为 "confusion:M/N" 其中M为真实标注，N为预测标注
func (e *ConfusionMatrixEvaluator) Evaluate(m supervised.Model, set data.Dataset) (result Evaluation) {
result.Metrics = make(map[string]float64)
iter := set.CreateIterator()
iter.Start()
for !iter.End() {
instance := iter.GetInstance()
out := m.Predict(instance)
name := fmt.Sprintf("confusion:%d/%d", instance.Output.Label, out.Label)
result.Metrics[name]++
iter.Next()
}
return
}
func (e *PREvaluator) Evaluate(m supervised.Model, set data.Dataset) (result Evaluation) {
tp := 0 // true-positive
tn := 0 // true-negative
fp := 0 // false-positive
fn := 0 // false-negative
iter := set.CreateIterator()
iter.Start()
for !iter.End() {
instance := iter.GetInstance()
if instance.Output.Label > 2 {
log.Fatal("调用PREvaluator但不是二分类问题")
}
out := m.Predict(instance)
if out.Label == 0 {
if instance.Output.Label == 0 {
tn++
} else {
fn++
}
} else {
if instance.Output.Label == 0 {
fp++
} else {
tp++
}
}
iter.Next()
}
result.Metrics = make(map[string]float64)
result.Metrics["precision"] = float64(tp) / float64(tp+fp)
result.Metrics["recall"] = float64(tp) / float64(tp+fn)
result.Metrics["tp"] = float64(tp)
result.Metrics["fp"] = float64(fp)
result.Metrics["tn"] = float64(tn)
result.Metrics["fn"] = float64(fn)
result.Metrics["fscore"] =
2 * result.Metrics["precision"] * result.Metrics["recall"] / (result.Metrics["precision"] + result.Metrics["recall"])
return
}
func (e *AccuracyEvaluator) Evaluate(m supervised.Model, set data.Dataset) (result Evaluation) {
correctPrediction := 0
totalPrediction := 0
iter := set.CreateIterator()
iter.Start()
for !iter.End() {
instance := iter.GetInstance()
out := m.Predict(instance)
if instance.Output.Label == out.Label {
correctPrediction++
}
totalPrediction++
iter.Next()
}
result.Metrics = make(map[string]float64)
result.Metrics["accuracy"] = float64(correctPrediction) / float64(totalPrediction)
return
}
func (opt *gdOptimizer) OptimizeWeights(
weights *util.Matrix, derivative_func ComputeInstanceDerivativeFunc, set data.Dataset) {
// 偏导数向量
derivative := weights.Populate()
// 学习率计算器
learningRate := NewLearningRate(opt.options)
// 优化循环
iterator := set.CreateIterator()
step := 0
var learning_rate float64
convergingSteps := 0
oldWeights := weights.Populate()
weightsDelta := weights.Populate()
instanceDerivative := weights.Populate()
log.Print("开始梯度递降优化")
for {
if opt.options.MaxIterations > 0 && step >= opt.options.MaxIterations {
break
}
step++
// 每次遍历样本前对偏导数向量清零
derivative.Clear()
// 遍历所有样本，计算偏导数向量并累加
iterator.Start()
instancesProcessed := 0
for !iterator.End() {
instance := iterator.GetInstance()
derivative_func(weights, instance, instanceDerivative)
derivative.Increment(instanceDerivative, 1.0/float64(set.NumInstances()))
iterator.Next()
instancesProcessed++
if opt.options.GDBatchSize > 0 && instancesProcessed >= opt.options.GDBatchSize {
// 添加正则化项
derivative.Increment(ComputeRegularization(weights, opt.options),
float64(instancesProcessed)/(float64(set.NumInstances())*float64(set.NumInstances())))
// 计算特征权重的增量
delta := opt.GetDeltaX(weights, derivative)
// 根据学习率更新权重
learning_rate = learningRate.ComputeLearningRate(delta)
weights.Increment(delta, learning_rate)
// 重置
derivative.Clear()
instancesProcessed = 0
}
}
if instancesProcessed > 0 {
// 处理剩余的样本
derivative.Increment(ComputeRegularization(weights, opt.options),
float64(instancesProcessed)/(float64(set.NumInstances())*float64(set.NumInstances())))
delta := opt.GetDeltaX(weights, derivative)
learning_rate = learningRate.ComputeLearningRate(delta)
weights.Increment(delta, learning_rate)
}
weightsDelta.WeightedSum(weights, oldWeights, 1, -1)
oldWeights.DeepCopy(weights)
weightsNorm := weights.Norm()
weightsDeltaNorm := weightsDelta.Norm()
log.Printf("#%d |w|=%1.3g |dw|/|w|=%1.3g lr=%1.3g", step, weightsNorm, weightsDeltaNorm/weightsNorm, learning_rate)
// 判断是否溢出
if math.IsNaN(weightsNorm) {
log.Fatal("优化失败：不收敛")
}
// 判断是否收敛
if weightsDelta.Norm()/weights.Norm() < opt.options.ConvergingDeltaWeight {
convergingSteps++
if convergingSteps > opt.options.ConvergingSteps {
log.Printf("收敛")
break
}
}
}
}
