pytorch 错误：“预期 isFloatingType(grad.scalar_type()) || (input_is_complex == grad_is_complex) 为真，但结果为假”

Question

I am working on the implementation of the classification problem. My code works well before inserting the complex noise into the output layer.

The following code does not work with Pytorch error

def forward(self, x):
    #Encoder Process
    out = self.feature(x)
    out = self.encoder(out)
    out = self.last(out)

    #### Start of error point ######
    batch_size, y = out.size()
    out_real = x[:, 0:int(y / 2)]
    out_comp = x[:, int(y / 2) : y + 1]
    out_comp = out_comp * 1j
    symbols = out_real + out_comp

    n = torch.randn(symbols.shape. dtype=torch.cfloat).to(device) 
    out = symbols + n 
    out_real = out.real 
    out_imag = out.imag 
    out = torch.cat((out_real, out_imag),1)
    #### End of error point ######

    #Decoder process
    out0 = self.decoder0(out)

    out1 = self.decoder1(out0)
    out1 += self.shortcut(out0)

    out2 = self.decoder2(out1)
    out2 += self.shortcut(out1)

Before adding '### start of error point ###... ### end of error point ###' lines the code works well and classification works. But I get the following error on last line after adding codes: RuntimeError: Expected isFloatingType(grad.scalar_type()) || (input_is_complex == grad_is_complex) to be true, but got false. (Could this error message be improved? If so, please report an enhancement request to PyTorch.) RuntimeError: Expected isFloatingType(grad.scalar_type()) || (input_is_complex == grad_is_complex) to be true, but got false. (Could this error message be improved? If so, please report an enhancement request to PyTorch.)

What could that be? What should I do? I want to add noise with complex, not real.

Answer 1

I think this error message is the least of your worries.

What are you trying to do? What is the meaning of adding complex noise, when you treat your features as two real parts? How is it different than just adding noise to out ?

If you really want to make your decoder complex, you should leave out as a complex tensor and allow the weights of the decoders and shortcuts to be complex as well. This will actually create interactions between the real parts and the imaginary parts of your tensor. Right now, there is no "complex" meaning for the real and imaginary parts of out , they are just additional dimensions.