func (_ SigmoidCECost) Cost(x linalg.Vector, a autofunc.Result) autofunc.Result {
logsig := autofunc.LogSigmoid{}
log := logsig.Apply(a)
invLog := logsig.Apply(autofunc.Scale(a, -1))
xVar := &autofunc.Variable{x}
oneMinusX := autofunc.AddScaler(autofunc.Scale(xVar, -1), 1)
sums := autofunc.Add(autofunc.Mul(xVar, log), autofunc.Mul(oneMinusX, invLog))
return autofunc.Scale(autofunc.SumAll(sums), -1)
}
func (_ CrossEntropyCost) Cost(x linalg.Vector, a autofunc.Result) autofunc.Result {
return autofunc.Pool(a, func(a autofunc.Result) autofunc.Result {
xVar := &autofunc.Variable{x}
logA := autofunc.Log{}.Apply(a)
oneMinusA := autofunc.AddScaler(autofunc.Scale(a, -1), 1)
oneMinusX := autofunc.AddScaler(autofunc.Scale(xVar, -1), 1)
log1A := autofunc.Log{}.Apply(oneMinusA)
errorVec := autofunc.Add(autofunc.Mul(xVar, logA),
autofunc.Mul(oneMinusX, log1A))
return autofunc.Scale(autofunc.SumAll(errorVec), -1)
})
}
func (s *LogSoftmaxLayer) Apply(in autofunc.Result) autofunc.Result {
return autofunc.Pool(in, func(in autofunc.Result) autofunc.Result {
// Compute the log of the sum of the exponents by
// factoring out the largest exponent so that all
// the exponentials fit nicely inside floats.
maxIdx := maxVecIdx(in.Output())
maxValue := autofunc.Slice(in, maxIdx, maxIdx+1)
exponents := autofunc.AddFirst(in, autofunc.Scale(maxValue, -1))
expSum := autofunc.SumAll(autofunc.Exp{}.Apply(exponents))
expLog := autofunc.Log{}.Apply(expSum)
denomLog := autofunc.Add(expLog, maxValue)
return autofunc.AddFirst(in, autofunc.Scale(denomLog, -1))
})
}
func (d *DropoutLayer) Apply(in autofunc.Result) autofunc.Result {
if d.Training {
return autofunc.Mul(in, d.dropoutMask(len(in.Output())))
} else {
return autofunc.Scale(in, d.KeepProbability)
}
}
func (r *RegularizingCost) Cost(a linalg.Vector, x autofunc.Result) autofunc.Result {
regFunc := autofunc.SquaredNorm{}
cost := r.CostFunc.Cost(a, x)
for _, variable := range r.Variables {
norm := regFunc.Apply(variable)
cost = autofunc.Add(cost, autofunc.Scale(norm, r.Penalty))
}
return cost
}
// ApplyBlock applies the block to an input.
func (g *GRU) ApplyBlock(s []State, in []autofunc.Result) BlockResult {
stateVars, stateRes := PoolVecStates(s)
var gateInputs []autofunc.Result
for i, x := range stateRes {
gateInputs = append(gateInputs, in[i], x)
}
n := len(in)
gateInput := autofunc.Concat(gateInputs...)
stateIn := autofunc.Concat(stateRes...)
resetMask := g.resetGate.Batch(gateInput, n)
updateMask := g.updateGate.Batch(gateInput, n)
maskedByReset := autofunc.Mul(resetMask, stateIn)
inputValue := autofunc.PoolSplit(n, maskedByReset,
func(newStates []autofunc.Result) autofunc.Result {
var newGateInputs []autofunc.Result
for i, input := range in {
newGateInputs = append(newGateInputs, input, newStates[i])
}
newIn := autofunc.Concat(newGateInputs...)
return g.inputValue.Batch(newIn, n)
})
newState := autofunc.Pool(updateMask, func(umask autofunc.Result) autofunc.Result {
updateComplement := autofunc.AddScaler(autofunc.Scale(umask, -1), 1)
return autofunc.Add(autofunc.Mul(umask, stateIn),
autofunc.Mul(updateComplement, inputValue))
})
return &gruResult{
InStates: stateVars,
Output: newState,
}
}
func (r *RescaleLayer) Apply(in autofunc.Result) autofunc.Result {
return autofunc.Scale(autofunc.AddScaler(in, r.Bias), r.Scale)
}
func (_ HyperbolicTangent) Apply(r autofunc.Result) autofunc.Result {
stretched := autofunc.Scale(autofunc.Sigmoid{}.Apply(autofunc.Scale(r, 2)), 2)
return autofunc.AddScaler(stretched, -1)
}
func (_ DotCost) Cost(x linalg.Vector, a autofunc.Result) autofunc.Result {
xVar := &autofunc.Variable{x}
return autofunc.Scale(autofunc.SumAll(autofunc.Mul(xVar, a)), -1)
}