func TrainRBM() {
train_file := "./data_0.dat"
validation_file := "./data_1.dat"
class_sizes := []int{4, 2, 5, 1}
hidden_layer_size := 2
train_data_accessor := rbm.NewInstanceLoader(train_file, len(class_sizes))
defer train_data_accessor.Close()
validation_data_accessor := rbm.NewInstanceLoader(validation_file, len(class_sizes))
defer validation_data_accessor.Close()
class_biases, y_bias := rbm.GetBiases(class_sizes, train_data_accessor)
var rbm_m rbm.SparseClassRBM
(&rbm_m).Initialize(class_sizes, class_biases, hidden_layer_size, y_bias)
var trainer rbm.RBMTrainer
learning_rate := rbm.WeightT(0.001)
regularization := rbm.WeightT(0.0)
momentum := rbm.WeightT(0.0)
gen_learning_imp := rbm.WeightT(0.0)
gibs_chain_len := 1
trainer.Initialize(&rbm_m, train_data_accessor, validation_data_accessor,
learning_rate, regularization, momentum, gen_learning_imp, gibs_chain_len)
trainer.Train()
//Evaluate RBM
// test_file := "./data_2.dat"
}
// Predict makes a prediction with the given input.
func (m *SimpleModel) Predict(x []int, noise rbm.WeightT) rbm.WeightT {
w := rbm.WeightT(0)
for i, k := range x {
w += m.feature_weights[i][k]
}
w += noise
return Sigmoid(w)
}
// GenerateInstance generates a data sample form the given model.
func (m *SimpleModel) GenerateInstance() ([]int, int) {
x := make([]int, len(m.feature_distribution))
for i, v := range m.feature_distribution {
k := SelectKFromDist(rbm.WeightT(rand.Float64()), v)
x[i] = k
}
y := 0
if m.Predict(x, m.noise_std) > m.threshold {
y = 1
}
return x, y
}
func Sigmoid(x rbm.WeightT) rbm.WeightT {
x_exp := math.Exp(float64(x))
return rbm.WeightT(x_exp / (1 + x_exp))
}