Softmax transiogram fit and multistart in Matrix Exponential path#86
Softmax transiogram fit and multistart in Matrix Exponential path#86pkofod wants to merge 19 commits into
Conversation
Co-authored-by: Patrick Kofod Mogensen <patrick.mogensen@gmail.com>
Co-authored-by: Patrick Kofod Mogensen <patrick.mogensen@gmail.com>
…ltistart Matrix Exponential fits because they appeared to have multiple local minima.
|
Thank you @pkofod for attempting an alternative solution with a softmax rewrite. Appreciate the time you put into this! I would like to avoid core changes to the original implementation though because it has been used in past projects with great success. There is nothing special about the box constraints in #85, yet we still get test failures with Optim.jl v2.x, which seems to indicate an issue on the optimization side. Can we try to debug it further over there? I will close the PR here to avoid confusion. |
|
@pkofod I am starting to appreciate your proposed solution with softmax. I've noticed that our current implementation has several flaws and that your approach is superior. |
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## main #86 +/- ##
==========================================
- Coverage 77.76% 77.36% -0.41%
==========================================
Files 47 47
Lines 1390 1396 +6
==========================================
- Hits 1081 1080 -1
- Misses 309 316 +7 ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
@juliohm please read this carefully as it may not be what you want. I changed a few things. First, proportions can be fit as you do, but the softmax version is more straight forward imho because this way they sum to 1 by construction. The multistart in the mat exp can be removed, but when running the tests I found that even if a minimum was found it was not always the global.
Some caveats: