SIRF incompatibilities

[paskino]

There are a number of workarounds we have put in place in CIL to overcome the residual incompatibilities with SIRF objects. This discussion is to try to list them and address them upstream, implementing what is needed in SIRF.

• SIRF data containers do not have a geometry member/property.
• SIRF AcquisitionModel do not have a caching mechanism for the norm.
• check usage of as_array and fill
• other places may contain hasattr calls to differentiate between CIL and SIRF objects.
• CI: test docker CIL SyneRBI/SIRF-SuperBuild#838 (comment)

[MargaretDuff]

SIRF priors and CIl function incompatibility from @epapoutsellis #1522

[casperdcl]

SIRF data containers' allocate methods don't always accept None SyneRBI/SIRF-SuperBuild#838 (comment) or SyneRBI/SIRF#1212

[KrisThielemans]

closed by SyneRBI/SIRF#1222

[MargaretDuff]

Checks in the Blockoperator for SIRFDataContainer:

CIL/Wrappers/Python/cil/optimisation/operators/BlockOperator.py

Lines 237 to 241 in 5406f33

	if issubclass(out.__class__, DataContainer) or \
	( has_sirf and issubclass(out.__class__, SIRFDataContainer) ):
	self.get_item(row, col).adjoint(
	x_b.get_item(row),
	out=out)

CIL/Wrappers/Python/cil/optimisation/operators/BlockOperator.py

Lines 248 to 251 in 5406f33

	if issubclass(out.__class__, DataContainer) or \
	( has_sirf and issubclass(out.__class__, SIRFDataContainer) ):
	out += self.get_item(row,col).adjoint(
	x_b.get_item(row))

[KrisThielemans]

@epapoutsellis @MargaretDuff if there is anything here that needs a SIRF fix, please create an issue there.

The geometry prlbem I think is more or less unfixable though. PET/MR/SPECT meta-info is more than geometry.

[MargaretDuff]

  ```python
    try:
        self.operator.range_geometry().allocate(value=1.0).divide(
            self.array, where=np.abs(self.array.as_array()) > 0, out=self.array)
    except:  # Due to CIL/SIRF compatibility and SIRF divide not taking kwargs
        sensitivity_np = self.array.as_array()
        self.pos_ind = np.abs(sensitivity_np) > 0
        array_np = np.zeros(self.operator.range_geometry().allocate().shape)
        array_np[self.pos_ind] = (1./sensitivity_np[self.pos_ind])
        self.array.fill(array_np)
        
        ```

SIRF does not divide (multiply, etc etc) by passing kwargs to numpy and required this work around