-
Notifications
You must be signed in to change notification settings - Fork 506
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[MRG] (f)gw barycenter solvers new features and cleaning #578
Conversation
Codecov Report
Additional details and impacted files@@ Coverage Diff @@
## master #578 +/- ##
==========================================
+ Coverage 96.61% 96.64% +0.03%
==========================================
Files 74 74
Lines 15037 15187 +150
==========================================
+ Hits 14528 14678 +150
Misses 509 509 |
if conv_criterion == 'barycenter': | ||
log_['err_feature'] = [] | ||
log_['err_structure'] = [] | ||
log_['Ts_iter'] = [] |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I am really not convinced that Ts_iter
is useful to anyone, knowing that it uses a lot of memory relatively ^^'
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Nice PR. I think we shoumd call it stop_criterion
instead of conv it is mor eclear to me WDYT?
Types of changes
add a new option to (un)regularized (f)gw barycenter solvers :
conv_criterion
taking values into{"barycenter", "loss"}
so the user can choose convergence criterions either based on the absolute norm variations of the barycenter, or relative loss variations, respectively : The last criterion is most likely meaningful for most applications, plus suit to the ones used in inner (f)gw solvers based on conditional gradient. This feature is also added to the entropic (f)gw barycenter solvers and would come with an overhead of computing the loss at the last iteration of the inner entropic solvers. Whereas for CG inner functions, computing the barycenter loss is performed in O(#samples).Trying to harmonise the behavior across all (f)gw barycenter solvers :
fixed_structure
andfixed_features
toentropic_fused_gromov_barycenter
, same than unregularized solver.Adapt and improve these functions documentation.
Motivation and context / Related issue
How has this been tested (if it applies)
PR checklist