RFC: more generic convergence assessment #530
Conversation
|
I love that the commits have two versions of you as authors :) I like the change. Does this pass all tests locally @ you? |
Codecov Report
@@ Coverage Diff @@
## master #530 +/- ##
=======================================
Coverage 89.87% 89.87%
=======================================
Files 35 35
Lines 1777 1777
=======================================
Hits 1597 1597
Misses 180 180
Continue to review full report at Codecov.
|
|
I don't know why there's two me there... tests are passing yes, I had a small mistake in the PR. |
| @@ -13,6 +13,14 @@ update_h!(d, state, method::SecondOrderOptimizer) = hessian!(d, state.x) | |||
|
|
|||
| after_while!(d, state, method, options) = nothing | |||
|
|
|||
| initial_convergence(d, state, method, initial_x, options) = false | |||
| initial_convergence(d, state, method::ZerothOrderOptimizer, initial_x, options) = false | |||
| converged = x_converged || f_converged || g_converged | ||
|
|
||
| return x_converged, f_converged, g_converged, converged, f_increased | ||
| end | ||
|
|
||
| gradient_convergence_assessment(state::AbstractOptimizerState, d, options) = g_residual(gradient(d)) ≤ options.g_tol | ||
| gradient_convergence_assessment(state::ZerothOrderState, d, options) = false |
There was a problem hiding this comment.
Yes, in Jonathan's solver outside of Julia :) He wants to make some ZerothOrderSolver's outside of Optim but hook into everything else in Optim
|
@jonathanBieler could you add a few tests? Just to make sure everything works as expected. |
|
I've added a couple of tests, I had to make DummyMethod type and change a few things, since those were not used anywhere, I assumed it's fine. |
I've tried to implement some of the changes discussed in #309, which allows my two minimal examples of ZerothOrderOptimizer and FirstOrderOptimizer defined outside of Optim to run.