func rbmExactExpectation(r *RBM, layer []bool, hidden bool) linalg.Vector {
var normalizer kahan.Summer64
var outcomeSum []kahan.Summer64
if hidden {
outcomeSum = make([]kahan.Summer64, rbmTestHiddenSize)
} else {
outcomeSum = make([]kahan.Summer64, rbmTestVisibleSize)
}
for i := 0; i < (1 << uint(len(outcomeSum))); i++ {
variableVec := boolVecFromInt(i, len(outcomeSum))
var prob float64
if hidden {
prob = math.Exp(-rbmEnergy(r, layer, variableVec))
} else {
prob = math.Exp(-rbmEnergy(r, variableVec, layer))
}
normalizer.Add(prob)
for j, b := range variableVec {
if b {
outcomeSum[j].Add(prob)
}
}
}
expectation := make(linalg.Vector, len(outcomeSum))
norm := 1.0 / normalizer.Sum()
for i, s := range outcomeSum {
expectation[i] = norm * s.Sum()
}
return expectation
}
// ExpectedHidden returns the expected value of
// the hidden layer given a visible vector.
func (r *RBM) ExpectedHidden(visible []bool) linalg.Vector {
result := make(linalg.Vector, len(r.HiddenBiases))
for i := range result {
var sum kahan.Summer64
for j, v := range visible {
if v {
sum.Add(r.Weights.Get(i, j))
}
}
result[i] = sum.Sum()
}
result.Add(r.HiddenBiases)
mapSigmoid(result)
return result
}
// ExpectedVisible returns the expected value of
// the visible layer given a hidden vector.
func (r *RBM) ExpectedVisible(hidden []bool) linalg.Vector {
result := make(linalg.Vector, len(r.VisibleBiases))
for i := range result {
var sum kahan.Summer64
for j, h := range hidden {
if h {
sum.Add(r.Weights.Get(j, i))
}
}
result[i] = sum.Sum()
}
result.Add(r.VisibleBiases)
mapSigmoid(result)
return result
}