Add an IterationControl.warmup(control) wrapper for controls #76

Activity

1. ablaom

   added

   enhancementNew feature or requestNew feature or request

   on Nov 2, 2026
2. aviralgarg05 commented on Nov 10, 2026

   

   aviralgarg05

   on Nov 10, 2026

   More actions

   Hi, I’d like to take this if it’s available. I reviewed the existing wrapper protocol and the current EarlyStopping.Warmup implementation. My understanding is that IterationControl.warmup(control; n=1) should leave an arbitrary wrapped control inactive for the first n control cycles, initialize it on cycle n + 1, and then forward its public state, stopping, teardown, and needs_* traits normally.

   I’d add focused unit and integration coverage using both WithLossDo and a stopping control, along with the control-wrapper documentation update. During the warmup period, should its trace remain empty like skip, or should it expose a warm field like EarlyStopping.Warmup? Is this issue available for me to work on?

   👍React with 👍1Reacted by Anthony Blaom, PhD
3. ablaom commented on Nov 11, 2026

   

   ablaom

   on Nov 11, 2026

   Last edited by ablaom

   MemberAuthor

   More actions

   Thanks for your interest and offer of help @aviralgarg05 .

   During the warmup period, should its trace remain empty like skip, or should it expose a warm field like EarlyStopping.Warmup?

   I expect mimicking skip is the way to go here. I don't think any additional field is necessary.

   Is this issue available for me to work on?

   Confirm you are still interested and I'll assign you.

   FYI: MLJIteration and hence MLJ have not yet been updated to use the latest IterationControl.jl version, which is breaking. The MLJIteration update may come soon, but I'm waiting on some other breaking changes to the MLJ ecosystem before updating the MLJ umbrella package, and that won't happen for a little while yet. I'll try and post some issues outlining these plans a little better soon.

   If you don't mind, pick one issue (this or the other one I assigned to you at StatisticalMeasures.jl) for us to work on through completion, and after that we can multitask if you that's your preference. Again, thanks for your offer of help.