Question

class TISP:

def __init__ (self, x, y, w, k, l, it):
#predict is a 2d list of observed values
    #actual is a 1d list corresponding to a 0-1 value normalized 
    #and descriptive of each observations outcome
    self.predicter = x
    self.actual = y
    self.normal_predicter = []
    self.normal_actual = []
    self.weights = w
    self.kap = k
    self.lam = l
    self.iterations = it
    return

def normalize_predicter (self):
    #create a listed array of predicter, with each value replaced with
    #z score from standardized collumn
    #also cuts off first column
    self.normal_predicter = []
    nppredict = numpy.array(self.predicter)
    nppredict_collumns = nppredict.T
    nppredict_collumns = numpy.delete(nppredict_collumns, 0, 0)
    for col in nppredict_collumns:
        self.normal_predicter.append(scipy.stats.mstats.zscore(col, axis=0))
    self.normal_predicter = numpy.array(self.normal_predicter).T
    return


def normalize_actual (self):
    self.normal_actual = []
    npactual = numpy.array(self.actual)
    self.normal_actual = scipy.stats.mstats.zscore(npactual)
    return

def predicter_matrix (self):
    #(I - 1/k^2 * X^t * X)
    identity_n = self.normal_predicter.shape[1]
    return numpy.identity(identity_n) - self.kap**-2 * numpy.dot(self.normal_predicter.T, self.normal_predicter)


def predicter_actual_matrix (self):
    #1/k^2 * X^t * Y
    return self.kap**-2 * numpy.dot(self.normal_predicter.T, self.normal_actual)


def pseudo_weights (self):
    #theta on slides
    return numpy.dot(self.predicter_matrix(), self.weights) + self.predicter_actual_matrix()



def adjust_lambda (self):
    return self.lam * self.kap**-2


def hard_threshhold (self, pseudo_weight):
    #put pseudo weights through hard threshhold
    adj_lam = self.adjust_lambda()
    if abs(pseudo_weight) < adj_lam:
        return adj_lam**2 - (abs(pseudo_weight) - adj_lam)**2
    else:
        return adj_lam**2


def update_weights (self):
    #update weights, call all
    weights = []
    for col, pw in enumerate(self.pseudo_weights()):
        weights.append(self.hard_threshhold(pw))
    self.weights = weights
    return


def run_all (self):
    self.normalize_all()
    for i in range(self.iterations):
        self.update_weights()
        print self.weights
    return

我省略了规范化功能，但是他们设置了属性，似乎工作正常。

obs 0的前几个列：

1，-0.0057548，-0.13457,0.027341，-0.23031，-0.25396，-0.084033,0.50101，-0.15747，-0.20175,0.027615，-0.24926,0.0090723，-0.056621，-0.10185,0.028461,0.1346

我正在实现一个名为tisp的阈值算法。

从权重的零开始，我第一次更新到其他一些低权重：

[0.0004，0.0004，0.0004，0.00020879693209322012，0.0004，0.00010554674122089871，0.0002128103652660953，0.00018560221711856617，0.00020718718704488368，0.00025472546603877807，0.0004，0.00032653814690485235，0.00028773649339636052，0.00014511605239718185，0.00026039641181569817，0.00032812782641074132，0.00010513077856280048，0.00020507079525590476，0.0002843078256187282，0.00018170990580009818]

然而，随着它不断进步，它将不再改变（我正在运行10次迭代）。

导致这种情况的原因是什么？

为什么这个属性会在一次迭代后更新，而不是其余的？

0 个答案: