Closed RoyStegeman closed 2 years ago
Also, perhaps we should talk to Juan in case he has some ideas about this. I guess @Radonirinaunimi @giacomomagni will have the opportunity to do so at some point next week?
I don't really have high hopes, but perhaps this added "randomization" can prevent getting "stuck".
I also believe it is unlikely this will solve all issues, but if we stop attempting we won't get nothing, so let's keep trying :)
I also believe it is unlikely this will solve all issues, but if we stop attempting we won't get nothing, so let's keep trying :)
That's the spirit! :smiling_face_with_tear:
I was just thinking that one other thing we could try is to take one step/epoch per experiment instead of calculating the chi2 for all experiments and then taking the step. I don't really have high hopes, but perhaps this added "randomization" can prevent getting "stuck".
Thought I'd open this issue so I don't forget :). Other ideas are of course also welcome.
Sorry for the late reply, I was offline. I am also investigating why the convergence is so poor and slow (at the level of the fit). I will create a new branch for this tomorrow.
Also, perhaps we should talk to Juan in case he has some ideas about this. I guess @Radonirinaunimi @giacomomagni will have the opportunity to do so at some point next week?
We will have our group meeting in person with Juan on Wednesday and we will indeed raise this issue (if not solved by then).
Hi @Radonirinaunimi @RoyStegeman I am happy to help but I really don't have enough information about what is going on ;) Anyway I will talk with Tanjona on wednesday and we can look at the problem together then
Thanks @juanrojochacon! Just as a quick update: by Wednesday we were running fits, but we observed some problems with convergence. We tried understanding this on Thursday and Friday but still have not been able to get to the heart of the issue. I think it's a bit much to explain the details over messages so it would indeed be good if Tanjona can explain the status to you on Wednesday.
I was just thinking that one other thing we could try is to take one step/epoch per experiment instead of calculating the chi2 for all experiments and then taking the step. I don't really have high hopes, but perhaps this added "randomization" can prevent getting "stuck".
Thought I'd open this issue so I don't forget :). Other ideas are of course also welcome.