SquaredL2Loss should inherit from WeightedSquaredL2Loss about scico

[lukepfister]

I don't think it would break anything. If jitted there would be no performance change.

[bwohlberg]

I expect that the performance change would be small in any event, but are you really confident that the jit analysis would be able to identify a multiplication by a variable that happens to have been assigned as a unit matrix?

Independent of this, is it possible to re-arrange the representation a bit so that we can set W = 1.0 (i.e. a scalar) in the case of no weighting?

[lukepfister]

iirc the default isn’t an IdentityMatrix, but rather the Identity LinOp— it’s eval is a no-op.

[lukepfister]

cf

scico/scico/loss.py

Lines 214 to 215 in 4684909

	if W is None:
	self.W = linop.Identity(y.shape)

[tbalke]

Only concern would be that the diagonal is explicitly set to snp.ones:

scico/scico/linop/_linop.py

Line 211 in 4684909

super().__init__(diagonal=snp.ones(input_shape, dtype=input_dtype), **kwargs)

But can this be scalar 1.0?

Then it could be that the shape of W are different than the shape of W.diagonal. Not sure whether that is a problem.

[lukepfister]

The shape of W and W.diagonal should be the same; we can't stop a user from doing something like W.diagonal = x where x.shape != W.shape, but that's just shooting themselves in the foot. The Diagonal class really treats the .diagonal attribute as a private thing, but Python doesn't have private attributes, so there you go.

As for identity: yes, it is true that Identity(shape).diagonal is a matrix of ones, but look at the eval:

scico/scico/linop/_linop.py

Lines 213 to 217 in 4684909

	def _eval(self, x: Union[JaxArray, BlockArray]) -> Union[JaxArray, BlockArray]:
	return x
	def __rmatmul__(self, x: Union[JaxArray, BlockArray]) -> Union[JaxArray, BlockArray]:
	return x

the .diagonal never gets used.

In earlier times, Identity did not inherit from Diagonal and this wasn't a problem

[tbalke]

In earlier times, Identity did not inherit from Diagonal and this wasn't a problem

What then is the reasoning to inherit from diagonal? Seems like in other contexts (other than the issue at hand) it may also be wasteful to store ones.

In other words, if SquaredL2Loss inherits from WeightedSquaredL2Loss and it is slightly inefficient, that is not exactly because of the inheritance but because the identity operator has some inefficiency?

[tbalke]

Also, this means that I(x) is very different than I.diagonal * x, if I is the Identity operator.

[lukepfister]

I suppose you could relax W.diagonal to be something that broadcasts up to input_shape.