Question on Training Epochs

[mhwang003]

Hi, Thanks for the great work.

It looks like in the paper, training is terminated after 100 epochs of training. However, in the code it looks like training is stopped after 1000 epochs. Do we get diminishing returns after 100 epochs when doing standard training with 4M samples including 5 different tasks? Also I'm curious if more training might result in over fitting as well.

Thanks in advance!

[WhoKnowsssss]

In this specific paper, we found that training above 100 epochs don't give too much benefits. For more difficult tasks such as humanoids, you do get benefits when trained longer :)