Pytorch 與matlab 中的傅里葉變換 fft

最近，pytorch 更新了 1.7.1， 支持了復數。並且torch.fft支持的文檔也說明的很清楚。https://pytorch.org/docs/stable/search.html?q=fft&check_keywords=yes&area=default

    # x = create_complex_number()
    # # dataset = MRBrainS18Dataset()
    # real_t = torch.tensor([0.+1.j, 2.+3.j, 4.+5.j, 6.+7.j])
    # comp_t = torch.fft.fft(real_t)
    # print(comp_t)
    # ireal_t = torch.fft.ifft(comp_t)
    # print(ireal_t)

    import scipy.io as scio
    from skimage import io

    path = r".\train_data\TRAIN\DATA_11.mat"
    item = scio.loadmat(path)
    CS_K_Data, IM, K_Data, mask = item['CS_K_Data'], item['IM'], item['K_Data'], item['mask']
    CS_K_Data, IM, K_Data, mask = CS_K_Data.astype(np.complex64), IM.astype(np.complex64), K_Data.astype(np.complex64), mask.astype(np.double)
    CS_K_Data, IM, K_Data = CS_K_Data[np.newaxis, np.newaxis, ...], IM[np.newaxis, np.newaxis, ...], K_Data[np.newaxis, np.newaxis, ...]
    # IM255 = 255*(IM-np.min(IM))/(np.max(IM)-np.min(IM))
    IM_tensor = torch.tensor(IM, dtype=torch.complex64)
    K_Data_tensor = torch.tensor(K_Data, dtype=torch.complex64)
    # K_Data_tensor = torch.ifft.shift(K_Data_tensor)

    # ! from IM to k-space
    fake_k = torch.fft.ifftn(IM_tensor, dim=(2,3), norm="backward")
    # "forward" - no normalization
    # "backward" - normalize by 1/n  Default is "backward", (normalize by 1/n).
    # "ortho" - normalize by 1/sqrt(n) (making the IFFT orthonormal)

    # print(torch.max(K_Data_tensor), torch.max(fake_k))
    # print(torch.min(K_Data_tensor), torch.min(fake_k))
    print('k', torch.mean(torch.abs(K_Data_tensor)), " and ", torch.mean(torch.abs(fake_k)))
    # ! from k-space to IM

    fake_im = torch.fft.fftn(K_Data_tensor, dim=(2,3), norm="backward")
    #     
    # "forward" - normalize by 1/n
    # "backward" - no normalization Default is "backward" (no normalization).
    # "ortho" - normalize by 1/sqrt(n) (making the FFT orthonormal)
    # print(torch.max(IM_tensor), torch.max(fake_im))
    # print(torch.min(IM_tensor), torch.min(fake_im))
    print('IM', torch.mean(torch.abs(IM_tensor)), " and ", torch.mean(torch.abs(fake_im)))
    # from shift to ishift
    xx = 1

print 的若兩者相等，則傅里葉變換成立。這個數據是之前比賽的。太坑了。fft和ifft 用反了。

    # matlab example
    # magic = torch.tensor([[8,1,6],[3,5,7],[4,9,2]])
    # fmagic = torch.fft.fftn(magic)
    # print(fmagic)
    # imagic = torch.fft.ifftn(fmagic)
    # print(imagic)