Complexfloat supported in clam pytorch?

vision
1 
def pdist(A, B = None, squared = False, eps = 1e-4):
    B = B if B is not None else A
    prod = torch.mm(A, B.t())
    normA = (A * A).sum(1).unsqueeze(1)
    normB = (B * B).sum(1).unsqueeze(1)
    D = clamp((normA.expand_as(prod) + normB.t().expand_as(prod) - 2 * prod), mn=0)
    return D if squared else clamp(clamp(D, mn=eps).sqrt(), mn=eps)

Getting error,
RuntimeError: “clamp_min_cpu” not implemented for ‘ComplexFloat’

pytorch 1.11.0

How do I bypass this error?

2

I’ve tested it with the current nightly binary (1.12.0.dev20220425) which also fails with the same error message, so could you create a feature request on GitHub, please?
In the meantime, you might want to split the complex tensor into its .real and .imag parts, apply the torch.clamp method on these parts, and rewrap the result into a complex dtype again.

2 Likes