Revert primal dual test to meaningful result

[mmcleod89]

• Brings primal dual test into line with others (checks the MSE rather than a strict regression test)
• Reverts previous changes made to measurement operator
• Now produces a reasonable looking image

The commit message attached to the code that had to be changed back was: "updated unit tests to be shorted and use un normed operators", which suggests that the un-normalised operators may be an issue for the implementation of the PD algorithm? (Commit 40be6b8)

[mmcleod89]

Original primal dual solution (solution.jpg) and current primal dual solution (pd_result.jpg) for context. I don't get exactly the same result but it looks very close.

[tkoskela]

It looks like the test still fails. Are we supposed to be multiplying average_intensity by 1e-3 in

purify/cpp/tests/algo_factory.cc

Line 130 in d8c9152

CHECK(mse <= average_intensity * 1e-3);

Is this a unit conversion?

[mmcleod89]

Not a unit conversion, just an loose attempt to check that the average error per pixel was significantly less than the average value per pixel as a first attempt at checking for similarity of solutions.

[mmcleod89]

I will play around a bit more with these tests and the UQ tomorrow

[mmcleod89]

These tests should pass when SOPT is updated to account for the change in the initial guess for PD

[mmcleod89]

The issue was just that just that the noise was too high for the tolerance because the majority of the image is black and therefore strongly biases the average value of the pixels. Visually inspecting the results they seem consistent and the algorithms are validated in SOPT, but I think a more robust testing framework would need to be built around synthetic data with a known solution and an expected level of noise

[tkoskela]

LGTM