示例#1
0
    def backward(ctx, grad_output):
        """
        In the backward pass we receive a Tensor containing the gradient of the loss
        with respect to the output, and we need to compute the gradient of the loss
        with respect to the input.
        """

        I = torch.ifft(grad_output, 2)
        I[:, :, :, 1] = 0
        F = torch.fft(I, 2)
        #F = grad_output

        out = fft2vec(F, ctx.apply_scaling)

        return out
示例#2
0
        N = 64
        tensor = 1000 * torch.randn(B, N, N)
        #tensor[:,:,:,1] = 0 # no imaginary
        #tensor = torch.zeros(B, N, N, 2)
        #for i in range(N):
        #    for j in range(N):
        #        tensor[:,i,j,0] = i*N+j

        print('tensor', tensor[0, :, :])

        tensor_fft = torch.rfft(tensor, 2, onesided=False)

        print('tensor_fft', tensor_fft[0, :, :, 0])
        print('tensor_fft', tensor_fft[0, :, :, 1])

        vec = fft2vec(tensor_fft, apply_scaling=True)

        #print(vec.squeeze())

        fft_recovered = vec_fft(vec)

        print('rec_tensor_fft', fft_recovered[0, :, :, 0])
        print('rec_tensor_fft', fft_recovered[0, :, :, 1])

        print(torch.abs(tensor_fft - fft_recovered).sum())
        print(torch.abs(tensor_fft - fft_recovered).max())

        tensor_recovered = torch.irfft(fft_recovered, 2, onesided=False)

        print('rec_tensor', tensor_recovered[0, :, :])