mindscience.sciops.dft.DFTn

class mindscience.sciops.dft.DFTn(shape, dim=None, norm='backward', modes=None, compute_dtype=mstype.float32)[source]

1/2/3D discrete Fourier transformation on complex number. The results should be same as scipy.fft.fftn() .

Parameters

  • shape (tuple) – The shape of the dimensions to be transformed, other dimensions need not be included.

  • dim (tuple) – Dimensions to be transformed. Default: None, the trailing dimensions will be transformed.

  • norm (str) – Normalization mode, should be one of 'forward', 'backward', 'ortho'. Default: 'backward', same as torch.fft.irfftn

  • modes (Union[tuple, int, None]) – The length of the output transform axis. The modes must be no greater than half of the dimension of input 'x'. Default: None.

  • compute_dtype (mindspore.dtype) – The type of input tensor. Default: mstype.float32.

Inputs:

  • ar (Tensor) - Real part of the tensor to be transformed, with trailing dimensions aligned with shape.

  • ai (Tensor) - Imag part of the tensor to be transformed, with trailing dimensions aligned with shape.

Outputs:

  • br (Tensor) - Real part of the output tensor, with trailing dimensions aligned with shape.

  • bi (Tensor) - Imag part of the output tensor, with trailing dimensions aligned with shape.

Examples

>>> from mindspore import ops
>>> from mindflow.cell import DFTn
>>> ar = ops.rand((2, 32, 512))
>>> ai = ops.rand((2, 32, 512))
>>> dft_cell = DFTn(ar.shape[-2:])
>>> br, bi = dft_cell(ar, ai)
>>> print(br.shape)
(2, 32, 512)