func (c *FactorizeMachine) Init(params map[string]string) {
c.w = core.NewVector()
factor_number, _ := strconv.ParseInt(params["factors"], 10, 64)
c.params.FactorNumber = int(factor_number)
c.params.LearningRate, _ = strconv.ParseFloat(params["learning-rate"], 64)
c.params.Regularization, _ = strconv.ParseFloat(params["regularization"], 64)
for i := 0; i < c.params.FactorNumber; i++ {
c.v = append(c.v, core.NewVector())
}
}
func (algo *GaussianProcess) ExtractTargetValuesAsVector(samples []*core.RealSample) *core.Vector {
targets := core.NewVector()
for i := 0; i < len(samples); i++ {
targets.SetValue(int64(i), samples[i].Value)
}
return targets
}
func CovVector(X []*core.RealSample, y *core.RealSample, cov_func CovFunc) *core.Vector {
l := int64(len(X))
ret := core.NewVector()
for i := int64(0); i < l; i++ {
ret.SetValue(i, cov_func(X[i].GetFeatureVector(), y.GetFeatureVector()))
}
return ret
}
func RandomInitVector(dim int64) *core.Vector {
v := core.NewVector()
var i int64
for i = 0; i < dim; i++ {
v.Data[i] = (rand.Float64() - 0.5) / math.Sqrt(float64(dim))
}
return v
}
func (dt *RegressionTree) FindBestSplit(samples []*core.MapBasedSample, node *TreeNode, select_features map[int64]bool) {
feature_weight_labels := make(map[int64]*core.FeatureGoalDistribution)
sum_total := 0.0
sum_total2 := 0.0
count_total := 0.0
for _, k := range node.samples {
sum_total += samples[k].Prediction
sum_total2 += samples[k].Prediction * samples[k].Prediction
count_total += 1.0
}
feature_sum_right := core.NewVector()
feature_sum_right2 := core.NewVector()
feature_count_right := core.NewVector()
for _, k := range node.samples {
for fid, fvalue := range samples[k].Features {
feature_count_right.AddValue(fid, 1.0)
feature_sum_right.AddValue(fid, samples[k].Prediction)
feature_sum_right2.AddValue(fid, samples[k].Prediction*samples[k].Prediction)
_, ok := feature_weight_labels[fid]
if !ok {
feature_weight_labels[fid] = core.NewFeatureGoalDistribution()
}
feature_weight_labels[fid].AddWeightGoal(fvalue, samples[k].Prediction)
}
}
min_vari := 1e20
node.feature_split = core.Feature{Id: -1, Value: 0}
for fid, distribution := range feature_weight_labels {
sort.Sort(distribution)
split, vari := distribution.BestSplitByVariance(sum_total-feature_sum_right.GetValue(fid),
sum_total2-feature_sum_right2.GetValue(fid),
count_total-feature_count_right.GetValue(fid),
feature_sum_right.GetValue(fid),
feature_sum_right2.GetValue(fid),
feature_count_right.GetValue(fid))
if min_vari > vari {
min_vari = vari
node.feature_split.Id = fid
node.feature_split.Value = split
}
}
}
func (algo *GaussianProcess) Init(params map[string]string) {
dim, _ := strconv.ParseInt(params["dim"], 10, 64)
algo.Params = GaussianProcessParameters{}
algo.Params.Dim = dim // Pass in dim as a param.. and require feature space to be continous.
algo.Params.Theta = 1e-7 // Used by approximate inversion as the diagonal noise
radius := 0.1
camp := 40.0
cf := CovSEARD{}
radiuses := core.NewVector()
for i := int64(1); i <= dim; i++ {
radiuses.SetValue(i, radius)
}
cf.Init(radiuses, camp)
algo.CovarianceFunc = cf.Cov
}
func (algo *NeuralNetwork) PredictMultiClass(sample *core.Sample) *core.ArrayVector {
y := core.NewVector()
z := core.NewArrayVector()
for i := int64(0); i < algo.Params.Hidden; i++ {
sum := float64(0)
for _, f := range sample.Features {
sum += f.Value * algo.Model.L1.Data[i].GetValue(f.Id)
}
y.Data[i] = util.Sigmoid(sum)
}
y.Data[algo.Params.Hidden] = 1
for i := 0; i <= int(algo.MaxLabel); i++ {
sum := float64(0)
for j := int64(0); j <= algo.Params.Hidden; j++ {
sum += y.GetValue(j) * algo.Model.L2.GetValue(j, int64(i))
}
z.SetValue(i, sum)
}
z = z.SoftMaxNorm()
return z
}
func (lr *LROWLQN) updateValueGrad(pos *core.Vector, dataset *core.DataSet) {
var totalLoss float64 = 0.0
var grad *core.Vector = core.NewVector()
for _, sample := range dataset.Samples {
var score float64 = lr.getScore(pos, sample)
var signScore float64 = score
if sample.Label == 0 {
signScore = -score
}
var prob float64
var lnProb float64
if signScore < -30 {
prob = 0
lnProb = signScore
} else if signScore > 30 {
prob = 1
lnProb = 0
} else {
prob = 1.0 / (1.0 + math.Exp(-signScore))
lnProb = math.Log(prob)
}
var scale float64
if sample.Label == 0 {
scale = (1 - prob)
} else {
scale = -(1 - prob)
}
totalLoss += -lnProb
for _, fea := range sample.Features {
grad.AddValue(fea.Id, scale*fea.Value)
}
}
lr.lastPos = pos.Copy()
lr.lastCost = totalLoss
lr.lastGrad = grad
}
func (c *SVM) Init(params map[string]string) {
c.C, _ = strconv.ParseFloat(params["c"], 64)
c.e, _ = strconv.ParseFloat(params["e"], 64)
c.w = core.NewVector()
}
func (lr *LROWLQN) Init(params map[string]string) {
lr.Model = core.NewVector()
lr.Params.Regularization, _ = strconv.ParseFloat(params["regularization"], 64)
}
func (lr *LROWLQN) Train(dataset *core.DataSet) {
lr.dataSet = dataset
minimizer := NewOWLQNMinimizer(lr.Params.Regularization)
lr.Model = minimizer.Minimize(lr, core.NewVector())
}
func (algo *NeuralNetwork) Train(dataset *core.DataSet) {
algo.Model = TwoLayerWeights{}
algo.Model.L1 = core.NewMatrix()
algo.Model.L2 = core.NewMatrix()
for i := int64(0); i < algo.Params.Hidden; i++ {
algo.Model.L1.Data[i] = core.NewVector()
}
initalized := make(map[int64]int)
max_label := 0
for _, sample := range dataset.Samples {
if max_label < sample.Label {
max_label = sample.Label
}
for _, f := range sample.Features {
_, ok := initalized[f.Id]
if !ok {
for i := int64(0); i < algo.Params.Hidden; i++ {
algo.Model.L1.SetValue(i, f.Id, (rand.Float64()-0.5)/math.Sqrt(float64(algo.Params.Hidden)))
}
initalized[f.Id] = 1
}
}
}
algo.MaxLabel = int64(max_label)
for i := int64(0); i <= algo.Params.Hidden; i++ {
for j := int64(0); j <= algo.MaxLabel; j++ {
algo.Model.L2.SetValue(i, j, (rand.NormFloat64() / math.Sqrt(float64(algo.MaxLabel)+1.0)))
}
}
for step := 0; step < algo.Params.Steps; step++ {
if algo.Params.Verbose <= 0 {
fmt.Printf(".")
}
total := len(dataset.Samples)
counter := 0
for _, sample := range dataset.Samples {
y := core.NewVector()
z := core.NewVector()
e := core.NewVector()
delta_hidden := core.NewVector()
for i := int64(0); i < algo.Params.Hidden; i++ {
sum := float64(0)
wi := algo.Model.L1.Data[i]
for _, f := range sample.Features {
sum += f.Value * wi.GetValue(f.Id)
}
y.Data[i] = util.Sigmoid(sum)
}
y.Data[algo.Params.Hidden] = 1.0
for i := int64(0); i <= algo.MaxLabel; i++ {
sum := float64(0)
for j := int64(0); j <= algo.Params.Hidden; j++ {
sum += y.GetValue(j) * algo.Model.L2.GetValue(j, i)
}
z.SetValue(i, sum)
}
z = z.SoftMaxNorm()
e.SetValue(int64(sample.Label), 1.0)
e.AddVector(z, -1.0)
for i := int64(0); i <= algo.Params.Hidden; i++ {
delta := float64(0)
for j := int64(0); j <= algo.MaxLabel; j++ {
wij := algo.Model.L2.GetValue(i, j)
sig_ij := e.GetValue(j) * (1 - z.GetValue(j)) * z.GetValue(j)
delta += sig_ij * wij
wij += algo.Params.LearningRate * (y.GetValue(i)*sig_ij - algo.Params.Regularization*wij)
algo.Model.L2.SetValue(i, j, wij)
}
delta_hidden.SetValue(i, delta)
}
for i := int64(0); i < algo.Params.Hidden; i++ {
wi := algo.Model.L1.Data[i]
for _, f := range sample.Features {
wji := wi.GetValue(f.Id)
wji += algo.Params.LearningRate * (delta_hidden.GetValue(i)*f.Value*y.GetValue(i)*(1-y.GetValue(i)) - algo.Params.Regularization*wji)
wi.SetValue(f.Id, wji)
}
}
counter++
if algo.Params.Verbose > 0 && counter%2000 == 0 {
fmt.Printf("Epoch %d %f%%\n", step+1, float64(counter)/float64(total)*100)
}
}
if algo.Params.Verbose > 0 {
algo.Evaluate(dataset)
}
algo.Params.LearningRate *= algo.Params.LearningRateDiscount
}
fmt.Println()
}
/*
Given matrix m and vector v, compute inv(m)*v.
Based on Gibbs and MacKay 1997, and Mark N. Gibbs's PhD dissertation
Details:
A - positive seminidefinite matrix
u - a vector
theta - positive number
C = A + I*theta
Returns inv(C)*u - So you need the diagonal noise term for covariance matrix in a sense.
However, this algorithm is numerically stable, the noise term can be very small and the inversion can still be calculated...
*/
func (algo *GaussianProcess) ApproximateInversion(A *core.Matrix, u *core.Vector, theta float64, dim int64) *core.Vector {
max_itr := 500
tol := 0.01
C := core.NewMatrix()
for key, val := range A.Data {
C.Data[key] = val.Copy()
}
// Add theta to diagonal elements
for i := int64(0); i < dim; i++ {
_, ok := C.Data[i]
if !ok {
C.Data[i] = core.NewVector()
}
C.Data[i].Data[i] = C.Data[i].Data[i] + theta
}
var Q_l float64
var Q_u float64
var dQ float64
u_norm := u.Dot(u) / 2
// Lower bound
y_l := core.NewVector()
g_l := u.Copy()
h_l := u.Copy()
lambda_l := float64(0)
gamma_l := float64(0)
var tmp_f1 float64
var tmp_f2 float64
var tmp_v1 *core.Vector
tmp_f1 = g_l.Dot(g_l)
tmp_v1 = C.MultiplyVector(h_l)
// Upper bound
y_u := core.NewVector()
g_u := u.Copy()
h_u := u.Copy()
lambda_u := float64(0)
gamma_u := float64(0)
var tmp_f3 float64
var tmp_f4 float64
var tmp_v3 *core.Vector
var tmp_v4 *core.Vector
tmp_v3 = g_u.MultiplyMatrix(A)
tmp_v4 = C.MultiplyVector(h_u)
tmp_f3 = tmp_v1.Dot(g_u)
for i := 0; i < max_itr; i++ {
// Lower bound
lambda_l = tmp_f1 / h_l.Dot(tmp_v1)
y_l.AddVector(h_l, lambda_l) //y_l next
Q_l = y_l.Dot(u) - 0.5*(y_l.MultiplyMatrix(C)).Dot(y_l)
// Upper bound
lambda_u = tmp_f3 / tmp_v3.Dot(tmp_v4)
y_u.AddVector(h_u, lambda_u) //y_u next
Q_u = (y_u.MultiplyMatrix(A)).Dot(u) - 0.5*((y_u.MultiplyMatrix(C)).MultiplyMatrix(A)).Dot(y_u)
dQ = (u_norm-Q_u)/theta - Q_l
if dQ < tol {
break
}
// Lower bound var updates
g_l.AddVector(tmp_v1, -lambda_l) //g_l next
tmp_f2 = g_l.Dot(g_l)
gamma_l = tmp_f2 / tmp_f1
for key, val := range h_l.Data {
h_l.SetValue(key, val*gamma_l)
}
h_l.AddVector(g_l, 1) //h_l next
tmp_f1 = tmp_f2 //tmp_f1 next
tmp_v1 = C.MultiplyVector(h_l) //tmp_v1 next
// Upper bound var updates
g_u.AddVector(tmp_v4, -lambda_u) //g_u next
tmp_v3 = g_u.MultiplyMatrix(A) //tmp_v3 next
tmp_f4 = tmp_v3.Dot(g_u)
gamma_u = tmp_f4 / tmp_f3
for key, val := range h_u.Data {
h_u.SetValue(key, val*gamma_u)
}
h_u.AddVector(g_u, 1) //h_u next
tmp_v4 = C.MultiplyVector(h_u) //tmp_v4 next
tmp_f3 = tmp_f4 // tmp_f3 next
}
return y_l
}