用目前最新网络
leelaz.exe -g --noponder -w leelaz-model-3486475-64000.txt
执行 heatmap，输出结果为
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 25 0 0 0 0 0 0 0 0 0 0 0 25 0 0 0
0 0 34 169 0 0 0 0 0 0 0 0 0 0 0 191 30 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 31 151 0 0 0 0 0 0 0 0 0 0 0 176 30 0 0
0 0 1 24 0 0 0 0 0 0 0 0 0 0 0 26 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
pass: 0
winrate: 0.495564
明显策略网络不是旋转对称不变的。
可为什么？难道是输入数据没有经过对称化（即将1盘棋变成8盘棋）？
如果确实如此，难道作者想要让程序自己学会对称性？这一点重要吗？
关键是，很浪费啊。。。