Why the real and imaginary number are subtracted and added together to make the output?

[kronee0516]

Can't understand what is happening
Is that some kind of theory need to do that?

one more question
what is the pytorch version you use? i encountered different errors when about start to train

[pheepa]

Hi
Referring to the article

Complex-valued Building Blocks. Given a complex-valued convolutional filter W = A+iB with
real-valued matrices A and B, the complex convolution operation on complex vector h = x + iy
with W is done by W ∗h = (A∗x−B ∗y)+i(B ∗x+A∗y)

Basically, it's just a way of multiplying complex numbers:
(a + ib) (c + id) = ac + iad + ibc + i2bd ⇒ (a + ib) (c + id) = (ac - bd) + i(ad + bc) [Because i2 = -1]

Unfortunately, it is impossible to find out which pytorch version I used (no artefacts like requirements.txt left) because at that time I did not have enough experience to write reproducible code.
In any case, you can share your errors and we will try to solve them, it can be usefull for someone.

[JangraManisha]

Hi

I am writing about the errors i got while reproducing the code:

1. error required to set return_complex= True in torch.stft method. (Done)
2. if return_complex=True, the error asks to convert bias also to be of complex datatype. (Done by setting Bias=False while defining conv2D layer)
3. The next error required to set datatype of conv2D layer : self.real_conv = nn.Conv2d(in_channels=self.in_channels,
   out_channels=self.out_channels,
   kernel_size=self.kernel_size,
   padding=self.padding,
   stride=self.stride, bias=False,dtype=torch.cfloat)
   I don't remember the error.
4. Now I am getting the error of mismatch size of kernel and input. given groups=1, weight of size [45, 1, 7, 5], expected input[1, 2, 1, 1539] to have 1 channels, but got 2 channels instead. I am stuck here. After dataloader, the input shape that I am getting is torch.Size([2, 1, 1539, 214]). Can anyone help?