def BackProbGradients(self, output, networkOutput, layerOutputs):
weightGradients = [None] * (self.noOfLayers + 1)
if self.outputFunctionName == "SoftMax" and self.lossFunctionName == "CrossEntropy":
gradientsWRTActivation = self.LossAndOutputGradients[
'CrossEntropyWithSoftMax'](networkOutput, output)
weightGradients[self.noOfLayers] = np.matmul(
gradientsWRTActivation,
np.transpose(
fns.IntergrateBiasAndData(layerOutputs[self.noOfLayers])))
else:
gradientsWRTActivation = self.LossGradients[self.lossFunctionName](
networkOutput, output)
gradientsWRTActivation = self.OutputGradients[
self.outputFunctionName](networkOutput, gradientsWRTActivation)
weightGradients[self.noOfLayers] = np.matmul(
gradientsWRTActivation,
np.transpose(
fns.IntergrateBiasAndData(layerOutputs[self.noOfLayers])))
for i in reversed(range(0, self.noOfLayers)):
backProbGradient = np.matmul(
np.transpose(
fns.DisIntergrateBiasFromWeights(self.weights[i + 1])),
gradientsWRTActivation)
gradientsWRTActivation = self.ActivationGradients[
self.activationFunctionNames[i]](layerOutputs[i + 1],
backProbGradient)
weightGradients[i] = np.matmul(
gradientsWRTActivation,
np.transpose(fns.IntergrateBiasAndData(layerOutputs[i])))
return weightGradients
def FeedForward(self, data):
layersOutputs = [data]
data = fns.IntergrateBiasAndData(data)
for i in range(0, self.noOfLayers):
data = self.Activation[self.activationFunctionNames[i]](
data, self.weights[i])
layersOutputs.append(data)
data = fns.IntergrateBiasAndData(data)
return self.OutputFunction[self.outputFunctionName](
data, self.weights[self.noOfLayers]), layersOutputs