训练的时候出现irfft2不兼容的问题，请问怎么解决呢？

[ouyangxi0925]

adaptive_weights = torch.fft.irfft2(torch.view_as_complex(DFT_map), dim=(1, 2)).reshape(batch_size, 1,
RuntimeError: cuFFT error: CUFFT_INTERNAL_ERROR

[Linwei-Chen]

Hi, this error is often caused by non-contiguous memory layouts when using view_as_complex with cuFFT.

Please try adding .contiguous() before the view operation:

adaptive_weights = torch.fft.irfft2(torch.view_as_complex(DFT_map.contiguous()), dim=(1, 2)).reshape(batch_size, 1, ...)

If you are using mixed precision (AMP) or half-precision inputs, cuFFT may also fail if the dimensions are not powers of two. In that case, casting to float32 helps:

adaptive_weights = torch.fft.irfft2(torch.view_as_complex(DFT_map).float(), dim=(1, 2)).reshape(batch_size, 1, ...)