def posPhaseGrad(self, data):
hidden = self.hiddenFromVisible(data)
Blas.mulMatrixOnMatrix(data,
hidden,
out=self.vars["W"].grad,
transpA=True)
if self.useBias:
Blas.sumOnMatrix(data, out=self.vars["b"].grad)
Blas.sumOnMatrix(hidden, out=self.vars["c"].grad)
return hidden
def accGradParams(self, grad, scale=1.0, momentum=0.0):
if self.wmode == GroupMode.full:
if self.useW:
A, B = (grad, self.inData) if self.transpW else (self.inData,
grad)
BlasGroup.mulTensorBatch(A,
B,
out=self.vars["W"].grad,
formatA=self.format,
formatB=self.format,
formatOut="gbp",
transpA=True,
alpha=scale,
beta=momentum)
if self.useBias:
BlasGroup.sumOnTensorGroup(grad,
out=self.vars["b"].grad,
formatT=self.format)
else:
if self.useW:
A, B = (grad, self.inData) if self.transpW else (self.inData,
grad)
wgrad = BlasGroup.mulTensorBatch(A,
B,
transpA=True,
formatA=self.format,
formatB=self.format,
formatOut="gbp",
alpha=scale,
beta=momentum)
Blas.sumOnMatrix(wgrad.reshape(wgrad.shape[0], -1),
out=self.vars["W"].grad.ravel())
if self.useBias:
Blas.sumOnMatrix(grad.reshape(grad.shape[0] * grad.shape[1],
grad.shape[2]),
out=self.vars["b"].grad[0])
def accGradParams(self, grad, scale=1.0, momentum=0.0):
if not self.transpose:
Blas.mulMatrixOnMatrix(self.inData,
grad,
out=self.vars["W"].grad,
transpA=True,
alpha=scale,
beta=momentum)
else:
Blas.mulMatrixOnMatrix(grad,
self.inData,
out=self.vars["W"].grad,
transpA=True,
alpha=scale,
beta=momentum)
if self.useBias:
Blas.sumOnMatrix(grad,
out=self.vars["b"].grad,
alpha=scale,
beta=momentum)
def negPhaseGrad(self, hidden):
visible = self.visibleFromHidden(hidden)
hidden = self.hiddenFromVisible(visible)
Blas.mulMatrixOnMatrix(visible,
hidden,
out=self.vars["W"].grad,
transpA=True,
alpha=-1.0,
beta=1.0)
if self.useBias:
Blas.sumOnMatrix(visible,
out=self.vars["b"].grad,
alpha=-1.0,
beta=1.0)
Blas.sumOnMatrix(hidden,
out=self.vars["c"].grad,
alpha=-1.0,
beta=1.0)
return hidden
def updateGrad(self, grad):
if self.useW:
formatOut = self.format if self.inmode == GroupMode.full else "gbp"
self.grad = BlasGroup.mulTensorBatch(grad,
self.W,
formatA=self.format,
formatB="gbp",
transpB=not self.transpW,
formatOut=formatOut)
if self.inmode != GroupMode.full:
self.grad = Blas.sumOnMatrix(
self.grad.reshape(self.groups,
grad.shape[0] * self.W.shape[1]))
self.grad = self.grad.reshape(grad.shape[0], 1,
self.W.shape[1])
else:
self.grad = grad