WIFS Tutorial 2017 – Understanding DNNs and their Predictions
DRAFT VERSION
modules.py In [ ]: import numpy # ------------------------# Feed-forward network # ------------------------class Network: def __init__(self,layers): self.layers = layers def forward(self,Z): for l in self.layers: Z = l.forward(Z) return Z def gradprop(self,DZ): for l in self.layers[::-1]: DZ = l.gradprop(DZ) return DZ def update(self,lr): for l in self.layers: l.update(lr) def dump(self): for l in self.layers: l.dump() # ------------------------# ReLU activation layer # ------------------------class ReLU: def def def def
forward(self,X): self.Z = X>0; return X*self.Z gradprop(self,DY): return DY*self.Z update(self,lr): pass dump(self): pass
# ------------------------# Sum-pooling layer # ------------------------class Pooling: def forward(self,X): self.X = X self.Y = 0.5*(X[:,::2,::2,:]+X[:,::2,1::2,:]+X[:,1::2,::2,:]+X[:,1::2,1::2,:]) return self.Y def gradprop(self,DY): self.DY = DY DX = self.X*0 for i,j in [(0,0),(0,1),(1,0),(1,1)]: DX[:,i::2,j::2,:] += DY*0.5 return DX def update(self,lr): pass def dump(self): pass # ------------------------# Convolution layer 1
# ------------------------class Convolution: def __init__(self,name,write=False): if write: wshape = map(int,list(name.split("-")[-1].split("x"))) w,h,m,n = wshape self.W = numpy.random.normal(0,1/(w*h*m)**.5,wshape) self.B = numpy.zeros([n]) self.name = name else: wshape = map(int,list(name.split("-")[-1].split("x"))) self.W = numpy.loadtxt(name+’-W.txt’).reshape(wshape) self.B = numpy.loadtxt(name+’-B.txt’) def forward(self,X): self.X = X mb,wx,hx,nx = X.shape ww,hw,nx,ny = self.W.shape wy,hy = wx-ww+1,hx-hw+1 Y = numpy.zeros([mb,wy,hy,ny],dtype=’float32’) for i in range(ww): for j in range(hw): Y += numpy.dot(X[:,i:i+wy,j:j+hy,:],self.W[i,j,:,:]) return Y+self.B def gradprop(self,DY): self.DY = DY mb,wy,hy,ny = DY.shape ww,hw,nx,ny = self.W.shape DX = self.X*0 for i in range(ww): for j in range(hw): DX[:,i:i+wy,j:j+hy,:] += numpy.dot(DY,self.W[i,j,:,:].T) return DX def update(self,lr): mb,wx,hx,nx = self.X.shape mb,wy,hy,ny = self.DY.shape ww,hw = wx-wy+1,hx-hy+1 self.DW = self.W*0 for i in range(ww):
2
for j in range(hw): x = self.X[:,i:i+wy,j:j+hy,:] dy = self.DY self.DW[i,j,:,:] += numpy.tensordot(x,dy,axes=([0,1,2],[0,1,2])) self.DB = self.DY.sum(axis=(0,1,2)) self.W -= lr*self.DW self.B -= lr*self.DB self.B = numpy.minimum(0,self.B) def dump(self): numpy.savetxt(self.name+’-W.txt’,self.W.flatten(),fmt=’%.3f’) numpy.savetxt(self.name+’-B.txt’,self.B.flatten(),fmt=’%.3f’)
3
