numpy整形函数的PyTorch等价

Question

嗨，我有这些函数来展平我的复杂类型数据，将其提供给NN，并将NN预测重建为原始形式

def flatten_input64(Input): #convert (:,4,4,2) complex matrix to (:,64) real vector
  Input1 = Input.reshape(-1, 32, order='F')
  Input_vector=np.zeros([19957,64],dtype = np.float64)
  Input_vector[:,0:32] = Input1.real
  Input_vector[:,32:64] = Input1.imag
  return Input_vector

def convert_output64(Output): #convert (:,64) real vector to (:,4,4,2) complex matrix  
  Output1 = Output[:,0:32] + 1j * Output[:,32:64]
  output_matrix = Output1.reshape(-1, 4 ,4 ,2 , order = 'F')
  return output_matrix

我正在写一个定制的丢失，要求所有操作都在torch中，我应该在PyTorch中重写我的转换函数。问题是Pytork没有“F”顺序整形。我试着写我自己版本的F reorder，但它不起作用。 你知道我犯了什么错吗

def convert_output64_torch(input):
   # number_of_samples = defined
   for i in range(0, number_of_samples):
     Output1 = input[i,0:32] + 1j * input[i,32:64]
     Output2 = Output1.view(-1,4,4,2).permute(3,2,1,0)
   if i == 0:
     Output3 = Output2
   else:
     Output3 = torch.cat((Output3, Output2),0)
return Output3

更新：在@a_guest评论之后，我尝试用转置和整形重新创建矩阵，我得到的代码与numy中的F顺序整形相同：

def convert_output64_torch(input):
   Output1 = input[:,0:32] + 1j * input[:,32:64]
   shape = (-1 , 4 , 4 , 2)
   Output3 = torch.transpose(torch.transpose(torch.reshape(torch.transpose(Output1,0,1),shape[::-1]),1,2),0,3)
return Output3

Answer 1

在Numpy和PyTorch中，您可以通过以下操作获得等价物：a.T.reshape(shape[::-1]).T（其中a是数组或张量）：

>>> a = np.arange(16).reshape(4, 4)
>>> a
array([[ 0,  1,  2,  3],
       [ 4,  5,  6,  7],
       [ 8,  9, 10, 11],
       [12, 13, 14, 15]])
>>> shape = (2, 8)
>>> a.reshape(shape, order='F')
array([[ 0,  8,  1,  9,  2, 10,  3, 11],
       [ 4, 12,  5, 13,  6, 14,  7, 15]])
>>> a.T.reshape(shape[::-1]).T
array([[ 0,  8,  1,  9,  2, 10,  3, 11],
       [ 4, 12,  5, 13,  6, 14,  7, 15]])