spyrit.core.noise.PoissonApproxGaussSameNoise.forward

PoissonApproxGaussSameNoise.forward(z: tensor) → tensor

Corrupt measurement by Gaussian approximation of Poisson noise

\[y \sim \alpha z + \sqrt{\alpha z} \cdot \mathcal{N}(0, 1) \quad \text{with }z\ge 0\]

Args:

z (torch.tensor): Measurements \(z\) with arbitrary shape.

Returns:

torch.tensor: Noisy measurement \(y\) with the same shape as z.

Important

Contrary to PoissonApproxGauss, different noise realisations apply only to the last dimension of the input tensor. The same noise realisations are repeated to the first dimensions of the input tensor.

Raises:

RuntimeError: If there are negative values in the input tensor.

Example:

Two different noisy measurement vectors

>>> import spyrit.core.noise as sn
>>> import torch
>>> noise = sn.PoissonApproxGaussSameNoise(100)
>>> z = torch.empty(10, 4).uniform_(0, 1)
>>> y = noise(z)
>>> print(y.shape)
torch.Size([10, 4])
>>> print(f"Noiseless measurements in ({torch.min(z):.2f} , {torch.max(z):.2f})")
Noiseless measurements in (...)
>>> print(f"Noisy measurements in ({torch.min(y):.2f} , {torch.max(y):.2f})")
Noisy measurements in (...)

>>> y = noise(z)
>>> print(f"Noisy measurements in ({torch.min(y):.2f} , {torch.max(y):.2f})")
Noisy measurements in (...)