Number of sampling steps
Closed this issue · 2 comments
Jiet-97 commented
Hi,
If I want to train a model which can achieve good result in only 4 steps, should I set args.distill_learning_steps into 4? Because the default args.distill_learning_steps is 50. Thanks for your attention.
MKFMIKU commented
You should leave the distill_learning_steps as 50 or 20. distill_learning_steps
is the hyper-paramter for determining the sampling length of the teacher model.
Jiet-97 commented
Thanks for your reply!