Dual eval #17
Dual eval #17
Conversation
abopt/base.py
Outdated
| self.complete_g(state) | ||
|
|
||
| if self.problem.problem_dual_eval: | ||
| self.complete_y_g(state) |
There was a problem hiding this comment.
I see what you are concerned about. accessing _member from outside of the class.
The branching can be done inside Problem.f_g():
if self._objective_gradient:
return ...
else:
return self.f(), self.g()
There was a problem hiding this comment.
done. The branching is now in Problem.fg(x). In Proposal I removed complete_y and complete_g and replaced it by complete_y_g, this required replacing complete_y in the linesearch by complete_y_g as well
There was a problem hiding this comment.
is that what you meant?
abopt/base.py
Outdated
| g = self._gradient(x) | ||
| return g | ||
|
|
||
| def f_g(self, x): |
There was a problem hiding this comment.
Perhaps just call it fg? f_g may imply $f_g$. No strong preference here.
abopt/testing/__init__.py
Outdated
| def vjp(x, v): return v.dot(J) * phiprime(x) | ||
| def jvp(x, v): return J.dot(v * phiprime(x)) | ||
|
|
||
| def objective(x): |
There was a problem hiding this comment.
Shall we do a test for each of the following cases:
- only objective_gradient
- objective and objective_gradient
- objective and gradient
There was a problem hiding this comment.
Will do, I have also added a statement that checks that objective and gradient are not None is objective_gradient is.
|
Thanks! I'll rename the tests (dual is less explicit than WithObjectGradient) |
2 participants
I added a ChiSquareProblem_dual test which is called during the LBFGs tests. It passes (and I made sure it's being called). Should I look at anything else in the tests?