just a small thing I’ve noticed that might help:
I’ve noticed that running a colab during the day in my timezone (GMT+2.00) is several times faster than if I run in during night.
example: learn.lr_find() on a language_model_learner (based on AWD_LSTM) takes about 16 minutes as opposed to ~55 minutes when I ran it during night.
that could mean a couple of things: there are less users during those hours on colab for timezones between GMT-5.00 to GMT-9.00, or colab have added more computing power during the last 2 days
or could it be some penalty/quota colab algorithm? if someone has internals on that, i’d be glad to know