RuntimeError: "slow_conv2d_cpu" not implemented for 'ComplexDouble'

1

While training a CNN network, this error raised RuntimeError: "slow_conv2d_cpu" not implemented for 'ComplexDouble'. My datatype are dtype=torch.complex128. So now, which dtype do I need for my network to train? Is there any issue? Or to convert my data tensor in dtype = torch.float16?

trainloader = create_data_loader(data_save, batch_size)
    ntotal_len = len(trainloader)
    for epoch in range(epochs):
        model.train()
        running_loss = 0.0

        for target, labels in trainloader:
            target, labels = target.to(DEVICE), labels.to(DEVICE)

            outputs = model(target)
            loss = criterion(outputs, labels)

333
3331592×698 103 KB

2

I think convolutions for complex numbers are not supported currently.
The easiest approach is to convert them to a 2_channel tensor via
(with spectrograms)

tensor = torch.view_as_real(tensor)
tensor = tensor.permute(0,3,1,2) # B x [Re,Im] x F x T

or just
tensor = torch.view_as_real(tensor) # * x [Re,Im] for a generic tensor of any shape

1 Like
3

Solved. tensor = torch.view_as_real(tensor) is a perfect solution. Thank you Sir

4

Hi Juan (and Vin)!

As an aside, Conv2d has complex support in the latest nightly (version
1.12) (but does not yet in version 1.11).

Best.

K. Frank