spyrit.core.prep.SplitPoisson.forward_expe

SplitPoisson.forward_expe(x: tensor, meas_op: LinearSplit | HadamSplit) → Tuple[tensor, tensor]

Preprocess to compensate for image normalization and splitting of the measurement operator.

It computes \(m = \frac{x[0::2]-x[1::2]}{\alpha}\), where \(\alpha = \max H^\dagger (x[0::2]-x[1::2])\).

Contrary to forward(), the image intensity \(\alpha\) is estimated from the pseudoinverse of the unsplit measurements. This method is typically called for the reconstruction of experimental measurements, while forward() is called in simulations.

The method returns a tuple containing both \(m\) and \(\alpha\)

Args:

x: batch of measurement vectors

meas_op: measurement operator (required to estimate \(\alpha\))

Output (\(m\), \(\alpha\)): preprocess measurement and estimated intensities.

Shape:

x: \((B, 2M)\) where \(B\) is the batch dimension

meas_op: the number of measurements meas_op.M should match \(M\).

\(m\): \((B, M)\)

\(\alpha\): \((B)\)

Example:

>>> x = torch.rand([10,2*400], dtype=torch.float)
>>> Perm = torch.rand([32,32])
>>> meas_op = HadamSplit(400, 32,  Perm)
>>> split_op = SplitPoisson(10, meas_op)
>>> m, alpha = split_op.forward_expe(x, meas_op)
>>> print(m.shape)
>>> print(alpha.shape)
torch.Size([10, 400])
torch.Size([10])