// Forward performs a forward transfer algorithm of Neural network
// and returns the output.
func (net *NeuralNetwork) Forward(input []float64) []float64 {
output := make([]float64, len(net.OutputLayer))
if len(input)+1 != len(net.InputLayer) {
panic("Dimention doesn't match: The number units of input layer")
}
// Copy
for i := range input {
net.InputLayer[i] = input[i]
}
net.InputLayer[len(net.InputLayer)-1] = Bias
// Transfer to hidden layer from input layer
for i := 0; i < len(net.HiddenLayer)-1; i++ {
sum := 0.0
for j := range net.InputLayer {
sum += net.HiddenWeight[j][i] * net.InputLayer[j]
}
net.HiddenLayer[i] = nnet.Sigmoid(sum)
}
net.HiddenLayer[len(net.HiddenLayer)-1] = Bias
// Transfer to output layer from hidden layer
for i := 0; i < len(net.OutputLayer); i++ {
sum := 0.0
for j := range net.HiddenLayer {
sum += net.OutputWeight[j][i] * net.HiddenLayer[j]
}
output[i] = nnet.Sigmoid(sum)
}
net.OutputLayer = output
return output
}
// P_H_Given_V returns p(h=1|v), the conditinal probability of activation
// of a hidden unit given a set of visible units.
func (rbm *GBRBM) P_H_Given_V(hiddenIndex int, v []float64) float64 {
sum := 0.0
for j := 0; j < rbm.NumVisibleUnits; j++ {
sum += rbm.W[hiddenIndex][j] * v[j]
}
return nnet.Sigmoid(sum + rbm.C[hiddenIndex])
}
// PseudoLogLikelihood returns pseudo log-likelihood for a given input sample.
func (rbm *RBM) PseudoLogLikelihoodForOneSample(v []float64) float64 {
bitIndex := rand.Intn(len(v))
fe := rbm.FreeEnergy(v)
feFlip := rbm.FreeEnergy(flip(v, bitIndex))
cost := float64(rbm.NumVisibleUnits) * math.Log(nnet.Sigmoid(feFlip-fe))
return cost
}
// P_V_Given_H returns p(v=1|h) the conditinal probability of activation
// of a visible unit given a set of hidden units.
func (rbm *RBM) P_V_Given_H(visibleIndex int, h []float64) float64 {
sum := 0.0
for i := 0; i < rbm.NumHiddenUnits; i++ {
sum += rbm.W[i][visibleIndex] * h[i]
}
return nnet.Sigmoid(sum + rbm.B[visibleIndex])
}