Question

我正在尝试在Tensorflow中编写核密度估计算法。

在拟合KDE模型时，我正在遍历当前批处理中的所有数据，并且对于每个数据，我使用Function ExciseBracketedText(ByVal Text As String, Optional TrimSpaces AS Boolean = False) As String Dim lBracketCount As Long, sExcise As String, lCounter As Long 'How many pairs of Open/Close brackets are there? 'lBracketCount = Len(Text) - WorksheetFunction.Max(Len(Replace(Text, "(", "")), Len(Replace(Text, ")", ""))) 'This does not distinguish between correct & misordered pairs 'New code works in VB environments other than Excel VBA lBracketCount = Len(Replace(Text, "(", "")) If lBracketCount < Len(Replace(Text, ")", "")) Then lBracketCount = Len(Replace(Text, ")", "")) lBracketCount = Len(Text) - lBracketCount ' e.g. ")()(" will return 2, even though only 1 will be trimmed ExciseBracketedText = Text sExcise = "" 'This is the text to cut out in each loop While lBracketCount > 0 'Once for each bracket pair For lCounter = 1 To Len(ExciseBracketedText) If Mid(ExciseBracketedText, lCounter, 1) = "(" Then sExcise = "(" 'Reset the text-to-excise every time we hit an open bracket ElseIf Mid(ExciseBracketedText, lCounter, 1) = ")" Then If Len(sExcise) > 0 Then 'Replace the text in brackets when we hit a close bracket ExciseBracketedText = Replace(ExciseBracketedText, sExcise & ")", "") Exit For End If ElseIf Len(sExcise) > 0 Then 'If we already have an open bracket, at to the text we will remove sExcise = sExcise & Mid(ExciseBracketedText, lCounter, 1) End If Next lCounter lBracketCount = lBracketCount - 1 'Next pair sExcise = "" 'Reset the text to excise 'If the first open bracket is AFTER the last close bracket, then we can exit early If InStr(ExciseBracketedText, "(") > InStrRev(ExciseBracketedText, ")") Then lBracketCount = 0 Wend If TrimSpaces Then ExciseBracketedText=Application.Trim(ExciseBracketedText) End Function对象创建内核： tensorflow.contrib.distributions.MultivariateNormalDiag

后来，当我试图预测数据点相对于上面拟合的模型的可能性时，对于我正在评估的每个数据点，我将上面每个内核给出的概率加在一起： self.kernels = [MultivariateNormalDiag(loc=data, scale=bandwidth) for data in X]

这种方法仅在tf.reduce_sum([kernel._prob(X) for kernel in self.kernels], axis=0)是一个numpy数组时才有效，因为TF不允许你迭代Tensor。我的问题是，是否有一种方法可以使上述算法与X或X一起使用tf.Tensor？

Answer 1

我在这个问题上找到的一个答案解决了适应KDE并一举预测概率的问题。但是，实施有点过时了。

def fit_predict(self, data):
    return tf.map_fn(lambda x: \ 
        tf.div(tf.reduce_sum(
            tf.map_fn(lambda x_i: self.kernel_dist(x_i, self.bandwidth).prob(x), self.fit_X)),
            tf.multiply(tf.cast(data.shape[0], dtype=tf.float64), self.bandwidth[0])), self.X)

第一个tf.map_fn遍历我们正在计算可能性的数据，将每个单独内核的概率加在一起。

第二个tf.map_fn遍历我们用于调整模型的所有数据，并创建tf.contrib.distributions.Distribution（此处由kernel_dist参数化）。

self.X和self.fit_X是初始化KernelDensity对象时创建的占位符。

如何在Tensorflow中执行核密度估计

1 个答案: