A few months ago, I noticed that the LR finder, one of the most handy tools characterizing fastai, does not work properly (actually it does not work at all) on small datasets. In my case, it was Kaggle’s Monkeys classification dataset.
@jeremy encouraged me in making modifications to fastai code towards an improvement on small datasets, and submit them to him for review. I did it, but perhaps he was a bit busy back then, and the specific thread was left unanswered (actually, I even sent a PM to Jeremy). I thought he would have had the code improved by himself once he’d manage to find the time and stopped bothering him.
Regardless “my” improvement was quite crude: it made the lr finder iterate for more epochs on user’s request.
Unfortunately, after my last env update a few days ago, I still see the lr finder does not work on small datasets.
I don’t mean to bother you, but I think this should be addressed either by incorporating my (admittedly, quite mediocre) little bit of code, or by writing a more professional solution by yourself @jeremy, @rachel, @reshama.
Thanks a lot.