torch.fft.ihfft2#

torch.fft.ihfft2(input,s=None,dim=(-2,-1),norm=None,*,out=None)→Tensor#

Computes the 2-dimensional inverse discrete Fourier transform of realinput. Equivalent toihfftn() but transforms only thetwo last dimensions by default.

Note

Supports torch.half on CUDA with GPU Architecture SM53 or greater.However it only supports powers of 2 signal length in every transformed dimensions.

Parameters

• input (Tensor) – the input tensor

• s (Tuple[int],optional) – Signal size in the transformed dimensions.If given, each dimensiondim[i] will either be zero-padded ortrimmed to the lengths[i] before computing the Hermitian IFFT.If a length-1 is specified, no padding is done in that dimension.Default:s=[input.size(d)fordindim]

• dim (Tuple[int],optional) – Dimensions to be transformed.Default: last two dimensions.

• norm (str,optional) –

  Normalization mode. For the backward transform(ihfft2()), these correspond to:

  • "forward" - no normalization

  • "backward" - normalize by1/n

  • "ortho" - normalize by1/sqrt(n) (making the Hermitian IFFT orthonormal)

  Wheren=prod(s) is the logical IFFT size.Calling the forward transform (hfft2()) with the samenormalization mode will apply an overall normalization of1/n betweenthe two transforms. This is required to makeihfft2()the exact inverse.

  Default is"backward" (normalize by1/n).

Keyword Arguments

out (Tensor,optional) – the output tensor.

Example

>>>T=torch.rand(10,10)>>>t=torch.fft.ihfft2(t)>>>t.size()torch.Size([10, 6])

Compared against the full output fromifft2(), theHermitian time-space signal takes up only half the space.

>>>fftn=torch.fft.ifft2(t)>>>torch.allclose(fftn[...,:6],rfftn)True

The discrete Fourier transform is separable, soihfft2()here is equivalent to a combination ofifft() andihfft():

>>>two_ffts=torch.fft.ifft(torch.fft.ihfft(t,dim=1),dim=0)>>>torch.allclose(t,two_ffts)True