def __init__(self, Para):
PassThrough.__init__(self, Para)
try:
self.activationType = Para['activationType']
if not (self.activationType in Activation.activationTypes):
self.printError('Activation type not defined')
except:
self.printError('Activation type not defined')
def __init__(self, para):
PassThrough.__init__(self, para)
try:
self.poolType = para['poolType']
if not (self.poolType in self.poolTypes):
self.printError('Pool type not defined.')
self.stride = para['stride']
self.kernelShape = para['kernelShape']
except:
self.printError('Parameters not complete.')
def createAndInitializeStruct(self):
PassThrough.createAndInitializeStruct(self)
# set bias to zeros
self.sBeta = np.zeros(self.outChannel,
dtype=float) # bias intialized to zeros
self.sGamma = np.ones(self.outChannel,
dtype=float) # Gamma initialized to ones
self.normD = Scale.normDimList[len(self.inpShape) -
2] # axises to calculate mean and std
self.dGamma = np.zeros(self.outChannel, dtype=float)
self.dBeta = np.zeros(self.outChannel, dtype=float)
self.paraDict['sBeta'] = self.sBeta
self.paraDict['sGamma'] = self.sGamma
self.dParaDict['sBeta'] = self.dBeta
self.dParaDict['sGamma'] = self.dGamma
self.createOptimizerStruct()
def createAndInitializeStruct(self):
PassThrough.createAndInitializeStruct(self)
self.normD = Normalize.normDimList[len(
self.inpShape) - 2] # axises to calculate mean and std
self.mean = np.zeros(self.outChannel, dtype=float)
self.var = np.zeros(self.outChannel, dtype=float)
self.std = np.zeros(self.outChannel, dtype=float)
self.dMean = np.zeros(self.outChannel, dtype=float)
self.dVar = np.zeros(self.outChannel, dtype=float)
self.meanList = np.zeros((Normalize.numStats, self.outChannel),
dtype=float)
self.stdList = np.zeros((Normalize.numStats, self.outChannel),
dtype=float)
self.tmp = np.zeros(self.inpShape, dtype=float)
self.tmp2 = np.zeros(self.outChannel, dtype=float)
self.count = 0
self.normComputed = False
class Fork2(PassThrough):
def __init__(self, para):
PassThrough.__init__(self, para)
self.skip = PassThrough(
{'instanceName': para['instanceName'] + '_skip'})
self.main = PassThrough(
{'instanceName': para['instanceName'] + '_main'})
def fork(self, topSkip, topMain, bot):
self.skip.stack(topSkip, bot)
self.main.stack(topMain, bot)
self.bottom = bot.topInterface
def createAndInitializeStruct(self):
self.skip.createAndInitializeTopStruct()
self.main.createAndInitializeTopStruct()
self.createAndInitializeBotStruct()
def forward(self):
cbs = Layer.currentBatchSize
np.copyto(self.skip.out[:cbs], self.bottom.getOutput()[:cbs])
np.copyto(self.main.out[:cbs], self.bottom.getOutput()[:cbs])
def backward(self):
cbs = Layer.currentBatchSize
np.add(self.skip.top.getInputDerivative()[:cbs],
self.main.top.getInputDerivative()[:cbs],
out=self.inpDeriv[:cbs])
def printIOShape(self):
print(colors.BOLD + self.instanceName + colors.END, end='')
print(' input:',
self.bottom,
' ',
self.bottom.getOutput().shape,
end='')
print(' skip:', self.skip.getOutput().shape, end='')
print(' main:', self.main.getOutput().shape)
def createAndInitializeStruct(self):
self.inpShape = self.bottom.getOutput().shape
(_, hIn, wIn, cIn) = self.inpShape
(kH, kW) = self.kernelShape
self.kArea = kH * kW
hOut = (hIn - kH) // self.stride + 1
wOut = (wIn - kW) // self.stride + 1
self.outChannel = cIn
self.outShape = (Layer.batchSize, hOut, wOut, self.outChannel)
self.zerosArray = np.zeros(self.inpShape[1:], dtype=float)
if self.top.isPadded():
(pHOut, pWOut) = self.top.getPadShape()
pOutShape = (Layer.batchSize, self.outShape[1] + 2 * pHOut,
self.outShape[2] + 2 * pWOut, self.outChannel)
self.pOut = np.zeros(pOutShape, dtype=float) # padded activations
self.out = self.pOut[:, pHOut:-pHOut, pWOut:-pWOut, :]
else:
self.pOut = np.zeros(self.outShape, dtype=float)
self.out = self.pOut
self.tmpDA = np.zeros(self.outShape, dtype=float)
PassThrough.createAndInitializeBotStruct(self)
class Sum2(PassThrough):
def __init__(self,para):
PassThrough.__init__(self,para)
self.skip = PassThrough({'instanceName':para['instanceName']+'_skip'})
self.main = PassThrough({'instanceName':para['instanceName']+'_main'})
def sum(self,top, botSkip,botMain):
self.skip.stack(top,botSkip)
self.main.stack(top,botMain)
self.top = top.bottomInterface
# set the bottom to the main path. for collecting dimension only
self.bottom = self.main.bottom
def createAndInitializeStruct(self):
self.skip.createAndInitializeBotStruct()
self.main.createAndInitializeBotStruct()
# check the skip and main path input dimensions
if self.skip.bottom.getOutput().shape != self.main.bottom.getOutput().shape:
self.printError("Skip and main input dimensions do not match.")
self.createAndInitializeTopStruct()
def forward(self):
cbs = Layer.currentBatchSize
np.add(self.skip.bottom.getOutput()[:cbs],self.main.bottom.getOutput()[:cbs],
out=self.out[:cbs])
def backward(self):
cbs = Layer.currentBatchSize
np.copyto(self.main.inpDeriv[:cbs],self.top.getInputDerivative()[:cbs])
np.copyto(self.skip.inpDeriv[:cbs],self.top.getInputDerivative()[:cbs])
def printIOShape(self):
print(colors.BOLD+self.instanceName+colors.END,end='')
print(' skip:',self.skip.bottom,' ',self.skip.bottom.getOutput().shape,end='')
print(' main:',self.main.bottom,' ',self.main.bottom.getOutput().shape,end='')
print(' output:',self.out.shape)
def __init__(self, para):
PassThrough.__init__(self, para)
self.skip = PassThrough(
{'instanceName': para['instanceName'] + '_skip'})
self.main = PassThrough(
{'instanceName': para['instanceName'] + '_main'})
def __init__(self, para):
PassThrough.__init__(self, para)
def createAndInitializeStruct(self):
PassThrough.createAndInitializeStruct(self)
self.lossOut = []