func (trainer *AdaboostClassifierTrainer) trainNextBaseEstimator(data_set *mlearn.DataSet, step int) (
float64, mlearn.BaseClassifier, float64) {
base_estimator := trainer.baseModelTrainer.TrainClassifierWithWeights(data_set, trainer.weights)
// embedded features ranking
if trainer.options.EnableEmbeddedFeaturesRanking {
for j := range trainer.EmbeddedFeaturesRank {
trainer.EmbeddedFeaturesRank[j] += trainer.baseModelTrainer.GetFeaturesRank()[j]
}
}
// save the predictions for later use
sample := make([]float64, data_set.FeaturesNum)
for i := range trainer.prediction {
data_set.GetSampleInplace(i, sample)
trainer.prediction[i] = base_estimator.PredictProbe(sample)
}
// calculate the weighted prediction error
var err float64
for i := 0; i < data_set.SamplesNum; i++ {
if trainer.prediction[i] != data_set.Classes[i] {
err += trainer.weights[i]
}
}
// check it to prevent math.Log domain violation (err == 0 means that there's nothing to boost)
if err == 0 {
return 1.0, base_estimator, err
}
b := math.Log((1.0-err)/err) + math.Log(float64(data_set.ClassesNum)-1.0)
// update weights according to error
const minimalWeight float64 = 10e-9
var weights_sum float64
for i := range trainer.weights {
if trainer.prediction[i] != data_set.Classes[i] {
trainer.weights[i] *= math.Exp(b)
if trainer.weights[i] < minimalWeight { /* this is to deal with rounding errors */
trainer.weights[i] = 0
}
} else {
//trainer.weights[i] *= math.Exp(-b)
}
weights_sum += trainer.weights[i]
}
// normalize weights to 1.0 sum
var normalized_sum float64
for i := range trainer.weights {
trainer.weights[i] /= weights_sum
normalized_sum += trainer.weights[i]
}
// dirty hack to guarantee 1.0 weights sum which is important to base model (regardless of rounding errors)
trainer.weights[0] += 1.0 - normalized_sum
return b, base_estimator, err
}
