Numpy 函数类型错误：只能将大小为 1 的数组转换为 Python 标量

问题描述

class GCN:
  def __init__(self,alpha,adj,feature,hiddenlayer_neurons,output_layer_neurons):
    self.alpha=alpha
    self.adj=adj
    self.feature=feature
    self.hiddenlayer_neurons=hiddenlayer_neurons
    self.output_layer_neurons=output_layer_neurons
  
  def weightlayers(self):
    self.weights1= np.random.normal(loc=0,scale=0.5,size=(features.shape[1],self.hiddenlayer_neurons))
    print(features.shape)
    print(adj.shape)
    self.weights2= np.random.normal(loc=0,size=(self.hiddenlayer_neurons,self.output_layer_neurons))
    self.bias1= np.random.normal(loc=0,scale=0.05,size=self.hiddenlayer_neurons)
    self.bias2=np.random.normal(loc=0,size= self.output_layer_neurons)
    return self.weights1,self.weights2,self.bias1,self.bias2

  def sigmoid(self,x):
    sigma=1/(1+np.exp(-x))
    return sigma
  
  def softmax(self,inputs):
    inputs=inputs.astype(np.float)
    inputs=np.vectorize(inputs)
    f=np.exp(inputs) / float(sum(np.exp(inputs)))
    #f2 = np.vectorize(f)
    return f

  def forwardpropagation(self):
    self.weights1,self.bias2=self.weightlayers()

    self.bias1=(np.reshape(self.bias1,(-1,1))).T
    self.bias2=(np.reshape(self.bias2,1))).T
    print(self.bias1.ndim)
    #self.sigmoid=self.sigmoid()
    self.adj=self.adj.T
    self.input= self.adj.dot(self.feature).dot(self.weights1) + (self.bias1)
    print(self.input.shape)
    self.sigmaactivation= self.sigmoid(self.input)
    self.hiddeninput=(self.sigmaactivation @ self.weights2 ) + (self.bias2)
    self.output=self.softmax(self.hiddeninput)
    return self.output

对于 softmax 函数，它抛出上述错误。按照之前对类似问题的回答，我尝试将其矢量化并将其转换为浮点数。但这似乎不起作用。

当我对其进行矢量化时，出现此错误：

TypeError: loop of ufunc does not support argument 0 of type vectorize which has no callable exp method.

解决方法

对于作为二维数值数组的 inputs，您不需要所有矢量化或浮点转换。

考虑一个小的二维数组（整数类型，但无所谓）：

In [110]: arr = np.arange(6).reshape(2,3)
In [111]: np.exp(arr)
Out[111]: 
array([[  1.,2.71828183,7.3890561 ],[ 20.08553692,54.59815003,148.4131591 ]])

sum 是一个 python 函数，它进行一维求和 - 注意结果是 (3,) 形状数组。尝试对其进行标量 float 转换会产生错误：

In [112]: sum(np.exp(arr))
Out[112]: array([ 21.08553692,57.31643186,155.8022152 ])
In [113]: float(sum(np.exp(arr)))
Traceback (most recent call last):
  File "<ipython-input-113-0972ef0e1a76>",line 1,in <module>
    float(sum(np.exp(arr)))
TypeError: only size-1 arrays can be converted to Python scalars

np.sum 对所有值求和，返回一个值。这是浮动，但这并不重要。

In [114]: np.sum(np.exp(arr))
Out[114]: 234.2041839862982

可用于缩放单个值：

In [115]: f=np.exp(arr)
     ...: f/np.sum(f)
Out[115]: 
array([[0.00426978,0.01160646,0.03154963],[0.08576079,0.23312201,0.63369132]])

deep-learning matrix matrix numpy python softmax