def cal_derivative(self, W_w, W_u, item):
"""
calculate derivative only with l2 norm
:param weight:
:return:
"""
label, featureDic = item
dir_W = []
dir_U = []
temp_eux = []
temp_sywx = []
sum_eux = 0.0
sum_eux_sywx = 0.0
for i in range(self.M):
#get all the temp exp(uj * x) and sigmoid(y * wj * x)
#and get sum at the same time
eux = np.exp(fc.dotProduct(W_u[i], featureDic))
sywx = fc.sigmoid(label * fc.dotProduct(W_w[i], featureDic))
temp_eux.append(eux)
temp_sywx.append(sywx)
sum_eux += eux
sum_eux_sywx += eux * sywx
for i in range(self.M):
#calculate array uj and array wj
dir_w = {}
dir_u = {}
for index in featureDic:
dir_u[index] = temp_eux[i] * featureDic[index] / sum_eux - \
temp_eux[i] * temp_sywx[i] * featureDic[index] / sum_eux_sywx
dir_w[index] = label * temp_sywx[i] * (
temp_sywx[i] - 1) * featureDic[index] / sum_eux_sywx
dir_W.append(dir_w)
dir_U.append(dir_u)
return dir_W, dir_U
def shift(self,data, sList, yList, roList, W, newW, LW, LU):
newLW, newLU = fc.sumCalDerivative(newW[:len(newW)//2],newW[len(newW)//2:],data)
newGradient = newLW+newLU
gradient = LW + LU
for i in range(len(sList)):
slist = sList[i]
ylist = yList[i]
rolist = roList[i]
w = W[i]
neww = newW[i]
g = gradient[i]
newg = newGradient[i]
size = len(slist)
if size == self.memoryNum:
# print >> sys.stdout, "pop 老的S, Y, RO"
slist.pop(0)
ylist.pop(0)
rolist.pop(0)
nextS = {}
nextY = {}
fc.addMultInto(self.feaNum, nextS, neww, w, -1)
# print "newG: %s" % newGradient
fc.addMultInto(self.feaNum, nextY, newg, g, -1)
# print "nextS: %s" % nextS
# print "nextY: %s" % nextY
ro = fc.dotProduct(nextS, nextY)
slist.append(nextS)
ylist.append(nextY)
rolist.append(ro)
def cal_loss(self, W_w, W_u, data):
"""
calculate the loss over all data
:param w_w:
:param w_u:
:param data:
:return:
"""
loss = 0.0
for label, featureDic in data:
#sum_u is sum(exp(uj * x))
#sum_us is sum(exp(uj * x) * sigmoid(y * wi * x))
sum_u = 0
sum_us = 0
for i in range(self.M):
wx = fc.dotProduct(W_w[i], featureDic)
eux = np.exp(fc.dotProduct(W_u[i], featureDic))
sum_u += eux
sum_us += eux * fc.sigmoid(label * wx)
loss += np.log(sum_u) - np.log(sum_us)
print("loss is: %s" % loss)