-
Notifications
You must be signed in to change notification settings - Fork 0
Learned budget assignments in Volume are softmax normalised #144
Copy link
Copy link
Open
Labels
OptimisationIssue affects the optimisation of the detectorIssue affects the optimisation of the detectorenhancementNew feature or requestNew feature or requestgood first issueGood for newcomersGood for newcomersideaSomething not relevant to current work, but could be useful in the futureSomething not relevant to current work, but could be useful in the futurelow priorityShould be fixed eventually, but isn't urgentShould be fixed eventually, but isn't urgent
Description
Metadata
Metadata
Assignees
Labels
OptimisationIssue affects the optimisation of the detectorIssue affects the optimisation of the detectorenhancementNew feature or requestNew feature or requestgood first issueGood for newcomersGood for newcomersideaSomething not relevant to current work, but could be useful in the futureSomething not relevant to current work, but could be useful in the futurelow priorityShould be fixed eventually, but isn't urgentShould be fixed eventually, but isn't urgent
Volume.assign_budgettakes the learnedbudget_weights[-inf,inf] and normalises them to sum to 1 via a softmax function.This has the advantage that the learnable parameters have no constraint on their values (i.e. don't have to be positive).
It does however meant that there is a non-linear relationship between the learned values and the actual budget that each detector receives; which could lead to unpredictable/unexpected behaviour (e.g. panels suddenly becoming very large/small).
Instead I think it might be worth investigating how well the optimisation handles clamping the parameters in [0,inf] and normalising by their sum.