Implement user-friendly stopping criterions like StopWhenFreeEnergyChangeLess

[Nimrais]

It would be nice to have stopping criteria for the model better than just running N iterations. I like how it's actually done in the ManOpt ecosystem https://manoptjl.org/stable/plans/stopping_criteria/, where you just provide some callbacks that check the state of the solver and then stop once this condition is satisfied.

[bvdmitri]

We already have it :)
You can stop the inference. Read the callbacks section here, specifically before_iteration and after_iteration

[Nimrais]

I mean the fact that I can implement it myself is fine.

But what do you mean we have it? I can not write right now stopping_criterion=StopWhenFreeEnergyChangeLess(...).

[bvdmitri]

The issue is titled Implement callback-based stopping criteria not Implement "stop when free energy change less" stopping criteria". We have callback based stopping already, just return a boolean flag inside the after_iteration callback and it will stop. Is as easy as

infer(
    ...,
    callbacks = (
        after_iteration = some_stopping_criteria,
    )
)

if you meant specific criterias, like StopWhenFreeEnergyChangeLess I would suggest to open a separate issue (or rename this one and change the description), but the way the issue is titled and described is already implemented and is not an issue IMO. Also, include which criterias you have in mind to make it more actionable

[bvdmitri]

I do think its a great addition so don't get me wrong :) We only need a list of potential candidates here. Perhaps some can also find typing callbacks = ... too verbose? I don't know. We can indeed have something like stopping_criterion = ... which you suggested. Can be then stopping_criterion = FixedNumberOfIterations(100). Can be combined later on with

stopping_criterion = FixedNumberOfIterations(100) | StopWhenFreeEnergyChangeLess(1e-4)

would be super cool tbh