def __init__(self,inputSize, outputSize, activationType, bias, **kwargs):
self.bias = bias
self.inputSize = inputSize
self.outputSize = outputSize
##Was epsilonInit passed in?
if 'epsilonInit' in kwargs.keys():
self.epsilonInit = kwargs['epsilonInit']
else:
self.epsilonInit = 1./(2*inputSize)
#Create W on initialization:
if bias:
self.W = randInit(self.inputSize+1, self.outputSize, self.epsilonInit)
else:
self.W = randInit(self.inputSize, self.outputSize, self.epsilonInit)
# Create Activation Function Object(s):
# If Activation is spline, we need to compute the total number of params within this layer.
# We also need to create mutliple spline instances.
if activationType == 'spline':
raise Exception('Spline param counting not implemented yet - SW')
#Create as many instance as outputs here.
self.numParams = 0
else:
self.aF = ActivationFunction(activationType)
self.numParams = self.W.size
class Layer(object):
def __init__(self,inputSize, outputSize, activationType, bias, **kwargs):
self.bias = bias
self.inputSize = inputSize
self.outputSize = outputSize
##Was epsilonInit passed in?
if 'epsilonInit' in kwargs.keys():
self.epsilonInit = kwargs['epsilonInit']
else:
self.epsilonInit = 1./(2*inputSize)
#Create W on initialization:
if bias:
self.W = randInit(self.inputSize+1, self.outputSize, self.epsilonInit)
else:
self.W = randInit(self.inputSize, self.outputSize, self.epsilonInit)
# Create Activation Function Object(s):
# If Activation is spline, we need to compute the total number of params within this layer.
# We also need to create mutliple spline instances.
if activationType == 'spline':
raise Exception('Spline param counting not implemented yet - SW')
#Create as many instance as outputs here.
self.numParams = 0
else:
self.aF = ActivationFunction(activationType)
self.numParams = self.W.size
#Each layer has it's own parameter updater and getter:
#Here you can pass in a matrx the same size as W, or a param vector of size (numParams)
def setParams(self, params):
if params.shape == (self.numParams,):
self.W = np.reshape(params, self.W.shape)
elif params.shape == self.W.shape:
self.W = params
else:
raise Exception('params must be of size (' + str(self.numParams) + ',) or ' + str(self.W.shape) + '-SW')
def getParams(self):
return self.W.ravel()
#Unravel derivatives with respect to wieghts:
def getDW(self):
return self.dW.ravel()
#Pass inputs through network:
def forward(self, **kwargs):
#Takes in optional argument x, if no arg, use previous x.
if 'x' in kwargs.keys():
self.x = kwargs['x']
#Only concatente ones with new inputs!
if self.bias:
self.x = np.hstack((np.ones((self.x.shape[0],1)), self.x))
self.z = np.dot(self.x, self.W)
self.a = self.aF.forward(self.z)
return self.a
