add check for A not linop in prox

[tbalke]

See #215

Note that check for self.has_prox, namely that W is Diagonal, was redundant and so I removed it.

[tbalke]

In test_functional.py we get the following functions failing:

    def test_squared_l2(self):
        L = loss.SquaredL2Loss(y=self.y, A=self.Ao)
        assert L.has_eval == True
        assert L.has_prox == False  # not diagonal

    def test_weighted_squared_l2(self):
        L = loss.WeightedSquaredL2Loss(y=self.y, A=self.Ao, W=self.W)
        assert L.has_eval == True
        assert L.has_prox == False  # not diagonal

with

        assert L.has_eval == True
>       assert L.has_prox == False  # not diagonal
E       AssertionError: assert True == False
E        +  where True = <class 'scico.loss.WeightedSquaredL2Loss'>\nhas_eval = True\nhas_prox = True\n        .has_prox

However, self.Ao is a LinearOperator so in this case we indeed have a prox. Why are we asserting that has_prox is False?

[bwohlberg]

Note that check for self.has_prox, namely that W is Diagonal, was redundant and so I removed it.

In what way was it redundant?

[tbalke]

Note that check for self.has_prox, namely that W is Diagonal, was redundant and so I removed it.

In what way was it redundant?

We had

if isinstance(self.A, linop.Diagonal) and isinstance(self.W, linop.Diagonal):
    self.has_prox = True

At this point isinstance(self.W, linop.Diagonal) == True always because we set it to a Diagonal in case it is None.

[bwohlberg]

At this point isinstance(self.W, linop.Diagonal) == True always because we set it to a Diagonal in case it is None.

But what if W is not None, but is set to something other than a linop.Diagonal?

[tbalke]

At this point isinstance(self.W, linop.Diagonal) == True always because we set it to a Diagonal in case it is None.

But what if W is not None, but is set to something other than a linop.Diagonal?

Then raise TypeError(f"W must be None or a linop.Diagonal, got {type(W)}")

So it is either Diagonal or None in which case we set it to Identity which is a Diagonal

[bwohlberg]

However, self.Ao is a LinearOperator so in this case we indeed have a prox. Why are we asserting that has_prox is False?

Previously (some months ago), the prox was only defined if the linear operator was diagonal. That changed when you (I think) added a CG solution for non-diagonal case. I would have assumed that this test was a remnant from before, but that doesn't explain why it only started failing now.

[bwohlberg]

But what if W is not None, but is set to something other than a linop.Diagonal?

Then raise TypeError(f"W must be None or a linop.Diagonal, got {type(W)}")

So it is either Diagonal or None in which case we set it to Identity which is a Diagonal

OK, that makes sense. The full context is not apparent in the diff!

[tbalke]

However, self.Ao is a LinearOperator so in this case we indeed have a prox. Why are we asserting that has_prox is False?

Previously (some months ago), the prox was only defined if the linear operator was diagonal. That changed when you (I think) added a CG solution for non-diagonal case. I would have assumed that this test was a remnant from before, but that doesn't explain why it only started failing now.

Correct. I think that the test fails because this is not an appropriate test. I suggest changing the test to the negation. If the A is a linop then has_prox should be true.

[bwohlberg]

Correct. I think that the test fails because this is not an appropriate test. I suggest changing the test to the negation. If the A is a linop then has_prox should be true.

That seems reasonable, but I'm a bit concerned that we don't understand why the tests only started failing now.

[tbalke]

Correct. I think that the test fails because this is not an appropriate test. I suggest changing the test to the negation. If the A is a linop then has_prox should be true.

That seems reasonable, but I'm a bit concerned that we don't understand why the tests only started failing now.

The previous test asserted that has_prox == False (incorrectly) and the loss.py code set the has_prox attribute to False (incorrectly) when A was a linop.

Fixing the loss.py code made fixing the test code necessary.

[bwohlberg]

That makes sense w.r.t. tests, but it is a bit strange that we did not previously noticed that the new prox capability that was added was crippled by this oversight. Nevertheless, you may as well go ahead and make the proposed changes to the tests.

[codecov]

Codecov Report

Merging #220 (88a67f3) into main (912b460) will decrease coverage by 0.02%.
The diff coverage is 66.66%.

@@            Coverage Diff             @@
##             main     #220      +/-   ##
==========================================
- Coverage   92.44%   92.42%   -0.03%     
==========================================
  Files          49       49              
  Lines        3429     3431       +2     
==========================================
+ Hits         3170     3171       +1     
- Misses        259      260       +1     

Flag	Coverage Δ
unittests	92.42% <66.66%> (-0.03%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
scico/loss.py	88.34% <66.66%> (-0.76%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 912b460...88a67f3. Read the comment docs.

[bwohlberg]

Looks good, but let's ask @Michael-T-McCann to take a look too.

[Michael-T-McCann]

While it still strikes me as odd every time we call CG "a prox," this PR seems in keeping with the currently design.