Haoping Bai

关于添加Dirichlet Noise的疑问

4

您好！论文中提到对prior probabilities in the root node添加dirichlet noise，按[AlphaGo Zero](https://deepmind.com/documents/119/agz_unformatted_nature.pdf)和[AlphaZero](https://arxiv.org/pdf/1712.01815.pdf)论文里的notation来看，这里的prior probabilities应该指的是policy value network在feedforward相应的game state后输出的move probabilities，而不是mcts playout之后用visit count算出的search probabilities，但看下面这里好像是在visit count算出的search probabilities上加了dirichlet noise？ https://github.com/junxiaosong/AlphaZero_Gomoku/blob/66292c55cc53acfae7f7bc5a15a370571549bdd9/mcts_alphaZero.py#L190-L198 下面是下赢奕心那位的implementation，看起来是在root node的move probabilities加了noise的 (另外注意到他的alpha用了0.04，估计是因为用的15x15的棋盘?) https://github.com/zhongjn/gomokuer/blob/f44a6f3389b9a97647ce3c8fef17d0363cdb609b/search.cpp#L186-L188

Question about H-swish activation and SE layer

2

Hi, Thanks for sharing the exciting work! Two quick questions about the H-swish activation and the SE layer. H-swish activation is troublesome because of its skewed distribution. Specifically for LSQ,...