Question on Training Time
4ndrelim opened this issue · 2 comments
Hi, i'm really fascinated by what you've done and was hoping to recreate the reported results. But i seem to be having trouble re-creating them (granted, i've only run the command a few times, some with a few params tweak. But the longest i have run the default command for was an entire day). The accuracy seems to be incrementing slowly, and even after a day, it reaches around 20% and seem to oscillate up and down.
May i ask, how long did you train the model for? and under what parameter/specification did you use to achieve the high results mentioned in the paper?
Hi, thank you for this question.
My I know what platform you are using (e.g., GPUs)? We used 4xA6000 GPUs to run the experiment.
Oscillating is not observed in our experiment. Could you share your log with us?
Hi, sure, below is a screenshot. Im running on only 2x A4000 GPUs but was hoping if the same number of epochs is hit, the accuracy could be replicated (or marginally smaller). I ran the default command.
