selftraining katago:Should I change the lr usually?What is the final learning rate of b10c128？Training b10c128 About how much training s how much to reduce the learning rate?

[gaoruihua111]

Or that sustained lr 1.0 is strong enough.

[lightvector]

For KataGo's main run, the learning rate of b10c128 was never changed from 1.0. If you were to decrease it from 1.0 near the end of training, I would expect it to improve significantly in strength.

We didn't do this, because there wasn't a lot of value to trying to optimize the 10 block network, because we only used it for few days before switching to a bigger network. In general, because the training time of b10c128 was so short and was considered unimportant because the run quickly moved to larger networks, there was not a lot of study on how to optimize b10c128 specifically. You will mostly have to do your own experiments about how much training is necessary and what the optimal learning rate schedule is if you care a lot about about optimizing that size of network as much as possible.