Fix PANOCplus #66
Fix PANOCplus #66
Conversation
Codecov Report
@@ Coverage Diff @@
## master #66 +/- ##
==========================================
- Coverage 88.59% 88.57% -0.02%
==========================================
Files 22 22
Lines 894 893 -1
==========================================
- Hits 792 791 -1
Misses 102 102
Continue to review full report at Codecov.
|
| @@ -178,7 +177,6 @@ function Base.iterate(iter::PANOCplusIteration{R}, state::PANOCplusState) where | |||
| break | |||
| end | |||
| state.tau = tau_backtracks >= iter.max_backtracks - 1 ? R(0) : state.tau / 2 | |||
| tau_backtracks += 1 | |||
There was a problem hiding this comment.
This value is used in the lines immediately above this one, in some checks: if the value is instead updated earlier (at the top of the loop) probably the inequalities involving it need to be adjusted? (In order to maintain the exact same behavior, that is)
There was a problem hiding this comment.
The algorithm is identical in this regard, since it starts always with can_update_direction = true and so it takes tau_backtracks = 0. Then the update tau_backtracks += 1, or the reset, is effectively performed after the checks you mention.
This PR fixes a small yet major error:
set_next_direction!should useres_previnstead ofres, as the latter is tentatively updated.The solver with acceleration was working correctly on simple problems only.
Minor change:
tau_backtracksis restarted every time a new direction is set.Note that the algorithm remains well-defined, namely every iteration terminates, as long as
iter.minimum_gamma > 0.