shengkelong

I realize that you make the parameters trainable in the MSHF block. Does this fit the paper？

你好，我设法让他在windows上和sabaki连接，但是他自己断开了

4


关于和leelazero对照的问题

1

作者您好，我一直在关注sayuri的训练过程（因为对gumbel是否能真正提升训练很感兴趣），我注意到您在最新的日志中和leelazero的早期网络进行了对比，但是据我所知leelazero在早期很长一段时间的训练有比较严重的问题（具体是什么忘了，很久以前的事了），所以如果想要对比训练速度的话和sai比leelazero更合适，考虑到sayuri使用了一些kata的算法来改进训练，如果想要证明gumbel有效可能和kata的早期网络对比是最合适的。

The training is unstable

2

Thank you for your impressive work. But when I try to recurrent this network(I rewrite the code myself), sometimes the loss will suddenly increase by 10 times. The structure of...

`73 periods = players_subbed_in_at_each_period['PERIOD'].drop_duplicates().values.tolist()` Sometimes the game has no substitutions, which may miss the period. `periods = play_by_play['PERIOD'].drop_duplicates().values.tolist()` This avoids the problem.

About luck adjusted rapm

3

Thanks for your work, I can try to calculate rapm, but I still don't understand how to calculate the lucky adjusted rapm, I understand the basic principle, but how to...

I'm curious if noise takes effect in search_best_move

2

I observed that the policy will be set to noise in "expand_node", but the "update_policy" used during inference (in "process_mini_batch") will directly update the policy to the result of network...