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
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):相关问题 更多 >
编程相关推荐