// 初始化优化结构体
// 为结构体中的向量分配新的内存，向量的长度可能发生变化。
func (opt *lbfgsOptimizer) initStruct(labels, features int, isSparse bool) {
opt.labels = labels
opt.x = make([]*util.Matrix, *lbfgs_history_size)
opt.g = make([]*util.Matrix, *lbfgs_history_size)
opt.s = make([]*util.Matrix, *lbfgs_history_size)
opt.y = make([]*util.Matrix, *lbfgs_history_size)
opt.ro = util.NewVector(*lbfgs_history_size)
opt.alpha = util.NewVector(*lbfgs_history_size)
opt.beta = util.NewVector(*lbfgs_history_size)
if !isSparse {
opt.q = util.NewMatrix(labels, features)
opt.z = util.NewMatrix(labels, features)
for i := 0; i < *lbfgs_history_size; i++ {
opt.x[i] = util.NewMatrix(labels, features)
opt.g[i] = util.NewMatrix(labels, features)
opt.s[i] = util.NewMatrix(labels, features)
opt.y[i] = util.NewMatrix(labels, features)
}
} else {
opt.q = util.NewSparseMatrix(labels)
opt.z = util.NewSparseMatrix(labels)
for i := 0; i < *lbfgs_history_size; i++ {
opt.x[i] = util.NewSparseMatrix(labels)
opt.g[i] = util.NewSparseMatrix(labels)
opt.s[i] = util.NewSparseMatrix(labels)
opt.y[i] = util.NewSparseMatrix(labels)
}
}
}
// 输入和输出都有 bias 项
func (rbm *RBM) SampleHidden(v *util.Vector, n int, binary bool) *util.Vector {
rbm.lock.RLock()
defer rbm.lock.RUnlock()
hiddenDim := rbm.options.NumHiddenUnits + 1
visibleDim := rbm.lock.weights.NumValues()
hiddenUnits := util.NewVector(hiddenDim)
visibleUnits := util.NewVector(visibleDim)
hiddenUnits.Set(0, 1.0)
visibleUnits.Set(0, 1.0)
for j := 1; j < visibleDim; j++ {
visibleUnits.Set(j, v.Get(j))
}
// 更新 hidden units
for i := 1; i < hiddenDim; i++ {
prob := rbm.logistic(visibleUnits, i, true)
if binary {
hiddenUnits.Set(i, rbm.bernoulli(prob))
} else {
hiddenUnits.Set(i, prob)
}
}
// reconstruct n-1 次
for nCD := 0; nCD < n; nCD++ {
for j := 1; j < visibleDim; j++ {
var prob float64
prob = rbm.logistic(hiddenUnits, j, false)
visibleUnits.Set(j, prob)
}
for i := 1; i < hiddenDim; i++ {
prob := rbm.logistic(visibleUnits, i, true)
if binary {
hiddenUnits.Set(i, rbm.bernoulli(prob))
} else {
hiddenUnits.Set(i, prob)
}
}
}
return hiddenUnits
}
func main() {
flag.Parse()
runtime.GOMAXPROCS(runtime.NumCPU())
// 载入训练集
set := contrib.LoadLibSVMDataset(*libsvm_file, false)
// 创建训练器
machine := rbm.LoadRBM(*model)
visibleDim := set.GetOptions().FeatureDimension
hiddenDim := machine.GetOptions().NumHiddenUnits + 1
iter := set.CreateIterator()
iter.Start()
for !iter.End() {
instance := iter.GetInstance()
v := util.NewVector(visibleDim)
content := fmt.Sprintf("%s", instance.Output.LabelString)
for i := 0; i < visibleDim; i++ {
value := instance.Features.Get(i)
v.Set(i, value)
if value != 0.0 && *append {
content = fmt.Sprintf("%s %d:%d", content, i+1, int(value))
}
}
h := machine.SampleHidden(v, *numCD, *useBinary)
for i := 1; i < hiddenDim; i++ {
value := h.Get(i)
if value != 0.0 {
if *append {
if *useBinary {
content = fmt.Sprintf("%s %d:%d", content, visibleDim+i-1, int(value))
} else {
content = fmt.Sprintf("%s %d:%.3f", content, visibleDim+i-1, value)
}
} else {
if *useBinary {
content = fmt.Sprintf("%s %d:%d", content, i, int(value))
} else {
content = fmt.Sprintf("%s %d:%.3f", content, i, value)
}
}
}
}
fmt.Printf("%s\n", content)
iter.Next()
}
}
func (classifier *MaxEntClassifier) Predict(instance *data.Instance) data.InstanceOutput {
output := data.InstanceOutput{}
// 当使用NamedFeatures时转化为Features
if instance.NamedFeatures != nil {
if classifier.FeatureDictionary == nil {
return output
}
instance.Features = util.NewSparseVector()
// 第0个feature始终是1
instance.Features.Set(0, 1.0)
for k, v := range instance.NamedFeatures {
id := classifier.FeatureDictionary.TranslateIdFromName(k)
instance.Features.Set(id, v)
}
}
output.LabelDistribution = util.NewVector(classifier.NumLabels)
output.LabelDistribution.Set(0, 1.0)
z := float64(1)
mostPossibleLabel := 0
mostPossibleLabelWeight := float64(1)
for iLabel := 1; iLabel < classifier.NumLabels; iLabel++ {
sum := float64(0)
for _, k := range classifier.Weights.GetValues(iLabel - 1).Keys() {
sum += classifier.Weights.Get(iLabel-1, k) * instance.Features.Get(k)
}
exp := math.Exp(sum)
if exp > mostPossibleLabelWeight {
mostPossibleLabel = iLabel
mostPossibleLabelWeight = exp
}
z += exp
output.LabelDistribution.Set(iLabel, exp)
}
output.LabelDistribution.Scale(1 / z)
output.Label = mostPossibleLabel
if classifier.LabelDictionary != nil {
output.LabelString = classifier.LabelDictionary.GetNameFromId(output.Label)
}
return output
}
func TestComputeInstanceDerivative(t *testing.T) {
weights := util.NewMatrix(2, 3)
weights.GetValues(0).SetValues([]float64{1, 2, 3})
weights.GetValues(1).SetValues([]float64{3, 4, 5})
de := util.NewMatrix(2, 3)
instance := data.Instance{}
instance.Features = util.NewVector(3)
instance.Features.SetValues([]float64{1, 0.3, 0.4})
instance.Output = &data.InstanceOutput{Label: 0}
MaxEntComputeInstanceDerivative(weights, &instance, de)
util.ExpectNear(t, 0.0322, de.Get(0, 0), 0.0001)
util.ExpectNear(t, 0.0096, de.Get(0, 1), 0.0001)
util.ExpectNear(t, 0.0128, de.Get(0, 2), 0.0001)
util.ExpectNear(t, 0.9658, de.Get(1, 0), 0.0001)
util.ExpectNear(t, 0.2897, de.Get(1, 1), 0.0001)
util.ExpectNear(t, 0.3863, de.Get(1, 2), 0.0001)
instance.Features.SetValues([]float64{1, 0.6, 0.7})
instance.Output.Label = 1
MaxEntComputeInstanceDerivative(weights, &instance, de)
util.ExpectNear(t, -0.9900, de.Get(0, 0), 0.0001)
util.ExpectNear(t, -0.5940, de.Get(0, 1), 0.0001)
util.ExpectNear(t, -0.6930, de.Get(0, 2), 0.0001)
util.ExpectNear(t, 0.9899, de.Get(1, 0), 0.0001)
util.ExpectNear(t, 0.5939, de.Get(1, 1), 0.0001)
util.ExpectNear(t, 0.6929, de.Get(1, 2), 0.0001)
instance.Features.SetValues([]float64{1, 0.4, 0.2})
instance.Output.Label = 2
MaxEntComputeInstanceDerivative(weights, &instance, de)
util.ExpectNear(t, 0.0390, de.Get(0, 0), 0.0001)
util.ExpectNear(t, 0.0156, de.Get(0, 1), 0.0001)
util.ExpectNear(t, 0.0078, de.Get(0, 2), 0.0001)
util.ExpectNear(t, -0.0425, de.Get(1, 0), 0.0001)
util.ExpectNear(t, -0.0170, de.Get(1, 1), 0.0001)
util.ExpectNear(t, -0.0085, de.Get(1, 2), 0.0001)
}
func TestSkipIterator(t *testing.T) {
set := NewInmemDataset()
instance1 := new(Instance)
instance1.Features = util.NewVector(3)
instance1.Features.SetValues([]float64{1, 2, 3})
util.Expect(t, "true", set.AddInstance(instance1))
instance2 := new(Instance)
instance2.Features = util.NewVector(3)
instance2.Features.SetValues([]float64{3, 4, 5})
util.Expect(t, "true", set.AddInstance(instance2))
instance3 := new(Instance)
instance3.Features = util.NewVector(3)
instance3.Features.SetValues([]float64{7, 8, 9})
util.Expect(t, "true", set.AddInstance(instance3))
instance4 := new(Instance)
instance4.Features = util.NewVector(3)
instance4.Features.SetValues([]float64{30, 40, 50})
util.Expect(t, "true", set.AddInstance(instance4))
instance5 := new(Instance)
instance5.Features = util.NewVector(3)
instance5.Features.SetValues([]float64{70, 80, 90})
util.Expect(t, "true", set.AddInstance(instance5))
instance6 := new(Instance)
instance6.Features = util.NewVector(3)
instance6.Features.SetValues([]float64{31, 41, 51})
util.Expect(t, "true", set.AddInstance(instance6))
set.Finalize()
iter := NewSkipIterator(set, []int{0, 4})
iter.Start()
util.Expect(t, "false", iter.End())
util.Expect(t, "1", iter.GetInstance().Features.Get(0))
util.Expect(t, "2", iter.GetInstance().Features.Get(1))
util.Expect(t, "3", iter.GetInstance().Features.Get(2))
iter.Next()
util.Expect(t, "false", iter.End())
util.Expect(t, "70", iter.GetInstance().Features.Get(0))
util.Expect(t, "80", iter.GetInstance().Features.Get(1))
util.Expect(t, "90", iter.GetInstance().Features.Get(2))
iter.Next()
util.Expect(t, "true", iter.End())
iter = NewSkipIterator(set, []int{3, 1})
iter.Start()
util.Expect(t, "false", iter.End())
util.Expect(t, "30", iter.GetInstance().Features.Get(0))
util.Expect(t, "40", iter.GetInstance().Features.Get(1))
util.Expect(t, "50", iter.GetInstance().Features.Get(2))
iter.Next()
util.Expect(t, "true", iter.End())
iter = NewSkipIterator(set, []int{1, 2, 1})
iter.Start()
util.Expect(t, "false", iter.End())
util.Expect(t, "3", iter.GetInstance().Features.Get(0))
util.Expect(t, "4", iter.GetInstance().Features.Get(1))
util.Expect(t, "5", iter.GetInstance().Features.Get(2))
iter.Next()
util.Expect(t, "false", iter.End())
util.Expect(t, "30", iter.GetInstance().Features.Get(0))
util.Expect(t, "40", iter.GetInstance().Features.Get(1))
util.Expect(t, "50", iter.GetInstance().Features.Get(2))
iter.Next()
util.Expect(t, "false", iter.End())
util.Expect(t, "31", iter.GetInstance().Features.Get(0))
util.Expect(t, "41", iter.GetInstance().Features.Get(1))
util.Expect(t, "51", iter.GetInstance().Features.Get(2))
iter.Next()
util.Expect(t, "true", iter.End())
}
func TestSkipDataset(t *testing.T) {
set := NewInmemDataset()
instance1 := new(Instance)
instance1.Features = util.NewVector(3)
instance1.Features.SetValues([]float64{1, 2, 3})
util.Expect(t, "true", set.AddInstance(instance1))
instance2 := new(Instance)
instance2.Features = util.NewVector(3)
instance2.Features.SetValues([]float64{3, 4, 5})
util.Expect(t, "true", set.AddInstance(instance2))
instance3 := new(Instance)
instance3.Features = util.NewVector(3)
instance3.Features.SetValues([]float64{7, 8, 9})
util.Expect(t, "true", set.AddInstance(instance3))
instance4 := new(Instance)
instance4.Features = util.NewVector(3)
instance4.Features.SetValues([]float64{30, 40, 50})
util.Expect(t, "true", set.AddInstance(instance4))
instance5 := new(Instance)
instance5.Features = util.NewVector(3)
instance5.Features.SetValues([]float64{70, 80, 90})
util.Expect(t, "true", set.AddInstance(instance5))
instance6 := new(Instance)
instance6.Features = util.NewVector(3)
instance6.Features.SetValues([]float64{31, 41, 51})
util.Expect(t, "true", set.AddInstance(instance6))
set.Finalize()
buckets := []SkipBucket{}
buckets = append(buckets, SkipBucket{true, 0})
buckets = append(buckets, SkipBucket{false, 1})
buckets = append(buckets, SkipBucket{true, 3})
ss := NewSkipDataset(set, buckets)
util.Expect(t, "2", ss.NumInstances())
iter := ss.CreateIterator()
iter.Start()
util.Expect(t, "false", iter.End())
util.Expect(t, "1", iter.GetInstance().Features.Get(0))
util.Expect(t, "2", iter.GetInstance().Features.Get(1))
util.Expect(t, "3", iter.GetInstance().Features.Get(2))
iter.Next()
util.Expect(t, "false", iter.End())
util.Expect(t, "70", iter.GetInstance().Features.Get(0))
util.Expect(t, "80", iter.GetInstance().Features.Get(1))
util.Expect(t, "90", iter.GetInstance().Features.Get(2))
iter.Next()
util.Expect(t, "true", iter.End())
buckets = []SkipBucket{}
buckets = append(buckets, SkipBucket{true, 3})
buckets = append(buckets, SkipBucket{false, 1})
ss = NewSkipDataset(set, buckets)
util.Expect(t, "1", ss.NumInstances())
iter = ss.CreateIterator()
iter.Start()
util.Expect(t, "false", iter.End())
util.Expect(t, "30", iter.GetInstance().Features.Get(0))
util.Expect(t, "40", iter.GetInstance().Features.Get(1))
util.Expect(t, "50", iter.GetInstance().Features.Get(2))
iter.Next()
util.Expect(t, "true", iter.End())
buckets = []SkipBucket{}
buckets = append(buckets, SkipBucket{true, 1})
buckets = append(buckets, SkipBucket{false, 1})
buckets = append(buckets, SkipBucket{true, 1})
buckets = append(buckets, SkipBucket{false, 1})
ss = NewSkipDataset(set, buckets)
util.Expect(t, "3", ss.NumInstances())
iter = ss.CreateIterator()
iter.Start()
util.Expect(t, "false", iter.End())
util.Expect(t, "3", iter.GetInstance().Features.Get(0))
util.Expect(t, "4", iter.GetInstance().Features.Get(1))
util.Expect(t, "5", iter.GetInstance().Features.Get(2))
iter.Next()
util.Expect(t, "false", iter.End())
util.Expect(t, "30", iter.GetInstance().Features.Get(0))
util.Expect(t, "40", iter.GetInstance().Features.Get(1))
util.Expect(t, "50", iter.GetInstance().Features.Get(2))
iter.Next()
util.Expect(t, "false", iter.End())
util.Expect(t, "31", iter.GetInstance().Features.Get(0))
util.Expect(t, "41", iter.GetInstance().Features.Get(1))
util.Expect(t, "51", iter.GetInstance().Features.Get(2))
iter.Next()
util.Expect(t, "true", iter.End())
}
func LoadLibSVMDataset(path string, usingSparseRepresentation bool) data.Dataset {
log.Print("载入libsvm格式文件", path)
content, err := ioutil.ReadFile(path)
if err != nil {
log.Fatalf("无法打开文件\"%v\"，错误提示：%v\n", path, err)
}
lines := strings.Split(string(content), "\n")
minFeature := 10000
maxFeature := 0
labels := make(map[string]int)
labelIndex := 0
for _, l := range lines {
if l == "" {
continue
}
fields := strings.Split(l, " ")
_, ok := labels[fields[0]]
if !ok {
labels[fields[0]] = labelIndex
labelIndex++
}
for i := 1; i < len(fields); i++ {
if fields[i] == "" {
continue
}
fs := strings.Split(fields[i], ":")
fid, _ := strconv.Atoi(fs[0])
if fid > maxFeature {
maxFeature = fid
}
if fid < minFeature {
minFeature = fid
}
}
}
if minFeature == 0 || maxFeature < 2 {
log.Fatal("文件输入格式不合法")
}
log.Printf("feature 数目 %d", maxFeature)
log.Printf("label 数目 %d", len(labels))
set := data.NewInmemDataset()
for _, l := range lines {
if l == "" {
continue
}
fields := strings.Split(l, " ")
instance := new(data.Instance)
instance.Output = &data.InstanceOutput{
Label: labels[fields[0]],
LabelString: fields[0],
}
if usingSparseRepresentation {
instance.NamedFeatures = make(map[string]float64)
} else {
instance.Features = util.NewVector(maxFeature + 1)
}
// 常数项
if !usingSparseRepresentation {
instance.Features.Set(0, 1)
}
for i := 1; i < len(fields); i++ {
if fields[i] == "" {
continue
}
fs := strings.Split(fields[i], ":")
fid, _ := strconv.Atoi(fs[0])
value, _ := strconv.ParseFloat(fs[1], 64)
if usingSparseRepresentation {
instance.NamedFeatures[fs[0]] = value
} else {
instance.Features.Set(fid, value)
}
}
set.AddInstance(instance)
}
set.Finalize()
log.Print("载入数据样本数目 ", set.NumInstances())
return set
}
func (rbm *RBM) derivativeWorker(ch chan *data.Instance, out chan *util.Matrix,
visibleDim int, hiddenDim int) {
// 可见单元
visibleUnits := util.NewVector(visibleDim)
visibleUnits.Set(0, 1.0)
// 不可见单元
hiddenUnitsProb := util.NewVector(hiddenDim)
hiddenUnitsBinary := util.NewVector(hiddenDim)
hiddenUnitsProb.Set(0, 1.0)
hiddenUnitsBinary.Set(0, 1.0)
for {
instance := <-ch
derivative := util.NewMatrix(hiddenDim, visibleDim)
// 设置 visible units 的初始值
for j := 1; j < visibleDim; j++ {
visibleUnits.Set(j, instance.Features.Get(j))
}
rbm.lock.RLock()
// 更新 hidden units
for i := 1; i < hiddenDim; i++ {
prob := rbm.logistic(visibleUnits, i, true)
hiddenUnitsProb.Set(i, prob)
if rbm.options.UseBinaryHiddenUnits {
hiddenUnitsBinary.Set(i, rbm.bernoulli(prob))
}
}
// 计算 positive statistics
for i := 0; i < hiddenDim; i++ {
for j := 0; j < visibleDim; j++ {
derivative.Set(i, j, visibleUnits.Get(j)*hiddenUnitsProb.Get(i))
}
}
// 计算CD_n
for nCD := 0; nCD < rbm.options.NumCD; nCD++ {
for j := 1; j < visibleDim; j++ {
var prob float64
if rbm.options.UseBinaryHiddenUnits {
prob = rbm.logistic(hiddenUnitsBinary, j, false)
} else {
prob = rbm.logistic(hiddenUnitsProb, j, false)
}
visibleUnits.Set(j, prob)
}
for i := 1; i < hiddenDim; i++ {
prob := rbm.logistic(visibleUnits, i, true)
hiddenUnitsProb.Set(i, prob)
if rbm.options.UseBinaryHiddenUnits {
hiddenUnitsBinary.Set(i, rbm.bernoulli(prob))
}
}
}
rbm.lock.RUnlock()
// 计算 negative statistics
for i := 0; i < hiddenDim; i++ {
for j := 0; j < visibleDim; j++ {
old := derivative.Get(i, j)
derivative.Set(i, j, old-visibleUnits.Get(j)*hiddenUnitsProb.Get(i))
}
}
out <- derivative
}
}
func TestTrain(t *testing.T) {
set := data.NewInmemDataset()
instance1 := new(data.Instance)
instance1.Features = util.NewVector(4)
instance1.Features.SetValues([]float64{1, 1, 1, 3})
instance1.Output = &data.InstanceOutput{Label: 0}
set.AddInstance(instance1)
instance2 := new(data.Instance)
instance2.Features = util.NewVector(4)
instance2.Features.SetValues([]float64{1, 3, 1, 5})
instance2.Output = &data.InstanceOutput{Label: 0}
set.AddInstance(instance2)
instance3 := new(data.Instance)
instance3.Features = util.NewVector(4)
instance3.Features.SetValues([]float64{1, 3, 4, 7})
instance3.Output = &data.InstanceOutput{Label: 1}
set.AddInstance(instance3)
instance4 := new(data.Instance)
instance4.Features = util.NewVector(4)
instance4.Features.SetValues([]float64{1, 2, 8, 6})
instance4.Output = &data.InstanceOutput{Label: 1}
set.AddInstance(instance4)
set.Finalize()
gdTrainerOptions := TrainerOptions{
Optimizer: optimizer.OptimizerOptions{
OptimizerName: "gd",
RegularizationScheme: 2,
RegularizationFactor: 1,
LearningRate: 0.1,
ConvergingDeltaWeight: 1e-6,
ConvergingSteps: 3,
MaxIterations: 0,
GDBatchSize: 0, // full-bath
},
}
gdTrainer := NewMaxEntClassifierTrainer(gdTrainerOptions)
lbfgsTrainerOptions := TrainerOptions{
Optimizer: optimizer.OptimizerOptions{
OptimizerName: "lbfgs",
RegularizationScheme: 2,
RegularizationFactor: 1,
LearningRate: 1,
ConvergingDeltaWeight: 1e-6,
ConvergingSteps: 3,
MaxIterations: 0,
},
}
lbfgsTrainer := NewMaxEntClassifierTrainer(lbfgsTrainerOptions)
lbfgsTrainer.Train(set)
gdTrainer.Train(set).Write("test.mlf")
model := LoadModel("test.mlf")
util.Expect(t, "0", model.Predict(instance1).Label)
util.Expect(t, "0", model.Predict(instance2).Label)
util.Expect(t, "1", model.Predict(instance3).Label)
util.Expect(t, "1", model.Predict(instance4).Label)
}
func TestInMemDataset(t *testing.T) {
set := NewInmemDataset()
instance1 := new(Instance)
instance1.Features = util.NewVector(3)
instance1.Features.SetValues([]float64{1, 2, 3})
instance1.Output = &InstanceOutput{Label: 1}
util.Expect(t, "true", set.AddInstance(instance1))
instance2 := new(Instance)
instance2.Features = util.NewVector(3)
instance2.Features.SetValues([]float64{3, 4, 5})
instance2.Output = &InstanceOutput{Label: 2}
util.Expect(t, "true", set.AddInstance(instance2))
instance3 := new(Instance)
instance3.Features = util.NewSparseVector()
instance3.Features.SetValues([]float64{3, 4, 5})
instance3.Output = &InstanceOutput{Label: 0}
util.Expect(t, "false", set.AddInstance(instance3))
instance4 := new(Instance)
instance4.Features = util.NewVector(4)
instance4.Features.SetValues([]float64{3, 4, 5, 6})
instance4.Output = &InstanceOutput{Label: 4}
util.Expect(t, "false", set.AddInstance(instance4))
instance5 := new(Instance)
instance5.Features = util.NewVector(3)
instance5.Features.SetValues([]float64{3, 5, 5})
util.Expect(t, "false", set.AddInstance(instance5))
set.Finalize()
// 检查数据集选项
util.Expect(t, "false", set.GetOptions().FeatureIsSparse)
util.Expect(t, "3", set.GetOptions().FeatureDimension)
util.Expect(t, "true", set.GetOptions().IsSupervisedLearning)
util.Expect(t, "3", set.GetOptions().NumLabels)
util.Expect(t, "2", set.NumInstances())
iter := set.CreateIterator()
iter.Start()
util.Expect(t, "false", iter.End())
util.Expect(t, "1", iter.GetInstance().Features.Get(0))
util.Expect(t, "2", iter.GetInstance().Features.Get(1))
util.Expect(t, "3", iter.GetInstance().Features.Get(2))
iter.Next()
util.Expect(t, "false", iter.End())
util.Expect(t, "3", iter.GetInstance().Features.Get(0))
util.Expect(t, "4", iter.GetInstance().Features.Get(1))
util.Expect(t, "5", iter.GetInstance().Features.Get(2))
iter.Next()
util.Expect(t, "true", iter.End())
iter = set.CreateIterator()
iter.Start()
iter.Skip(2)
util.Expect(t, "true", iter.End())
util.Expect(t, "2", set.NumInstances())
}
