// New creates new GBRBM instance.
func New(numVisibleUnits, numHiddenUnits int) *GBRBM {
rbm := new(GBRBM)
rand.Seed(time.Now().UnixNano())
rbm.NumVisibleUnits = numVisibleUnits
rbm.NumHiddenUnits = numHiddenUnits
rbm.W = nnet.MakeMatrix(numHiddenUnits, numVisibleUnits)
rbm.B = make([]float64, numVisibleUnits)
rbm.C = make([]float64, numHiddenUnits)
rbm.GradW = nnet.MakeMatrix(numHiddenUnits, numVisibleUnits)
rbm.GradB = make([]float64, numVisibleUnits)
rbm.GradC = make([]float64, numHiddenUnits)
rbm.InitParam()
return rbm
}
// NewNeuralNetwork returns a new network instance with the number of
// input units, number of hidden units and number output units
// of the network.
func NewNeuralNetwork(numInputUnits,
numHiddenUnits, numOutputUnits int) *NeuralNetwork {
net := new(NeuralNetwork)
rand.Seed(time.Now().UnixNano())
// Layers
net.InputLayer = make([]float64, numInputUnits+1) // plus bias
net.HiddenLayer = make([]float64, numHiddenUnits)
net.OutputLayer = make([]float64, numOutputUnits)
// Weights
net.OutputWeight = nnet.MakeMatrix(numHiddenUnits, numOutputUnits)
net.HiddenWeight = nnet.MakeMatrix(numInputUnits+1, numHiddenUnits)
net.InitParam()
return net
}
func NewHiddenLayer(numInputUnits, numHiddenUnits int) *HiddenLayer {
h := new(HiddenLayer)
h.W = nnet.MakeMatrix(numInputUnits, numHiddenUnits)
h.NumInputUnits = numInputUnits
h.NumHiddenUnits = numHiddenUnits
h.B = make([]float64, numHiddenUnits)
h.Init()
return h
}
// Gradient returns gradients of GBRBM parameters for a given (mini-batch) dataset.
func (rbm *GBRBM) Gradient(data [][]float64,
miniBatchIndex int) ([][]float64, []float64, []float64) {
gradW := nnet.MakeMatrix(rbm.NumHiddenUnits, rbm.NumVisibleUnits)
gradB := make([]float64, rbm.NumVisibleUnits)
gradC := make([]float64, rbm.NumHiddenUnits)
for i, v := range data {
// Set start state of Gibbs-sampling
var gibbsStart []float64
persistentIndex := i + miniBatchIndex*rbm.Option.MiniBatchSize
if rbm.Option.UsePersistent {
gibbsStart = rbm.PersistentVisibleUnits[persistentIndex]
} else {
gibbsStart = v
}
// Perform reconstruction using Gibbs-sampling
reconstructedVisible := rbm.Reconstruct(gibbsStart,
rbm.Option.OrderOfGibbsSampling, rbm.Option.UseMean)
// keep recostructed visible
if rbm.Option.UsePersistent {
rbm.PersistentVisibleUnits[persistentIndex] =
reconstructedVisible
}
// pre-computation that is used in gradient computation
p_h_given_v1 := rbm.P_H_Given_V_Batch(v)
p_h_given_v2 := rbm.P_H_Given_V_Batch(reconstructedVisible)
// Gompute gradient of W
for i := 0; i < rbm.NumHiddenUnits; i++ {
for j := 0; j < rbm.NumVisibleUnits; j++ {
gradW[i][j] += p_h_given_v1[i]*v[j] -
p_h_given_v2[i]*reconstructedVisible[j]
}
}
// Gompute gradient of B
for j := 0; j < rbm.NumVisibleUnits; j++ {
gradB[j] += v[j] - reconstructedVisible[j]
}
// Gompute gradient of C
for i := 0; i < rbm.NumHiddenUnits; i++ {
gradC[i] += p_h_given_v1[i] - p_h_given_v2[i]
}
}
return rbm.normalizeGradBySizeOfBatch(gradW, gradB, gradC, len(data))
}
func (h *HiddenLayer) Gradient(input, deltas [][]float64) ([][]float64, []float64) {
gradW := nnet.MakeMatrix(len(h.W), len(h.W[0]))
gradB := make([]float64, len(h.B))
// Gradient
for n := range input {
for i := 0; i < h.NumHiddenUnits; i++ {
for j := 0; j < h.NumInputUnits; j++ {
gradW[j][i] -= deltas[n][i] * input[n][j]
}
gradB[i] -= deltas[n][i]
}
}
return gradW, gradB
}
// Train performs Contrastive divergense learning algorithm to train GBRBM.
// The alrogithm is based on (mini-batch) Stochastic Gradient Ascent.
func (rbm *GBRBM) Train(data [][]float64, option TrainingOption) error {
rbm.Option = option
opt := nnet.BaseTrainingOption{
Epoches: rbm.Option.Epoches,
MiniBatchSize: rbm.Option.MiniBatchSize,
Monitoring: rbm.Option.Monitoring,
}
// Peistent Contrastive learning
if rbm.Option.UsePersistent {
rbm.PersistentVisibleUnits = nnet.MakeMatrix(len(data), len(data[0]))
copy(rbm.PersistentVisibleUnits, data)
}
s := nnet.NewTrainer(opt)
return s.UnSupervisedMiniBatchTrain(rbm, data)
}
