ComplexDenoiserWrapper#

class deepinv.models.ComplexDenoiserWrapper(denoiser, mode='real_imag', *args, **kwargs)[source]#

Bases: Denoiser

Complex-valued wrapper for a real-valued denoiser \(\denoisername(\cdot, \sigma)\).

This class lifts any real-valued denoiser to the complex domain by applying it separately to a chosen pair of real representations of the complex input and recombining the outputs.

Let the input be \(x \in \mathbb{C}^{B\times C\times H\times W}\) and a noise level \(\sigma > 0\) (scalar or batch of size \(B\)). The underlying denoiser (given by denoiser) \(\denoisername\) acts on real tensors only. Two processing modes are supported:

  1. 'real_imag' mode

We decompose

\[x = x_{\mathrm{real}} + i x_{\mathrm{imag}}.\]

The denoiser is applied on the real and imaginary parts (same \(\sigma\) broadcast across both halves). The complex reconstruction is

\[\hat x = \denoisername(x_{\mathrm{real}}, \sigma) + i \, \denoisername(x_{\mathrm{imag}}, \sigma).\]

If the provided input tensor is real (i.e. torch.is_complex(x) is False), it is interpreted as \(x_{\mathrm{real}}\) with \(x_{\mathrm{imag}}=0\) and the output is returned as \(\denoisername(x_{\mathrm{real}},\sigma) + i 0\) (complex dtype ensured).

  1. 'abs_angle' mode

We use the polar decomposition

\[x = m \exp(i\phi), \qquad m = |x|,\; \phi = \mathrm{arg}(x) \in (-\pi,\pi].\]

The denoiser is applied on the magnitude and phase parts (same \(\sigma\) broadcast across both halves). The reconstructed complex output is

\[\hat x = \denoisername(m, \sigma) \exp \big(i\, \denoisername(\phi, \sigma)\big).\]

Note that the phase estimate \(\denoisername(\phi,\sigma)\) is clipped back to \((-\pi,\pi]\).

Note

This wrapper can only process complex inputs that are compatible with the underlying real-valued denoiser. For example, if the wrapped denoiser supports only single-channel (grayscale) real images, then the corresponding complex input must also be single-channel.

Examples:
>>> import deepinv as dinv
>>> import torch
>>> from deepinv.models import ComplexDenoiserWrapper, DRUNet
>>> denoiser = DRUNet()
...
>>> complex_denoiser = ComplexDenoiserWrapper(denoiser, mode="real_imag")
>>> y = torch.randn(2, 3, 32, 32, dtype=torch.complex64)  # complex input
>>> sigma = 0.1
>>> with torch.no_grad():
...     denoised = complex_denoiser(y, sigma)
>>> print(denoised.dtype)  # should be complex dtype
torch.complex64
Parameters:

  • denoiser (deepinv.models.Denoiser) – Real-valued denoiser \(\denoisername\) to wrap.

  • mode (str) – Either 'real_imag' or 'abs_angle'. Default 'real_imag'.

Raises:

ValueError – If an unsupported mode string is provided.

Returns:

Complex denoised output \(\hat x\) with same spatial shape as the input.

forward(x, sigma)[source]#

Applies the complex-valued denoiser. If a real tensor is provided, it is treated as a complex tensor with zero imaginary part.

Parameters:
Returns:

Denoised images, with the same shape as the input and will always be in complex dtype.

Return type:

Tensor

Examples using ComplexDenoiserWrapper:#

Random phase retrieval and reconstruction methods.

Random phase retrieval and reconstruction methods.