Yifeng Zhang comments

Results 4 comments of


                                            Yifeng Zhang

请问人机对弈的时候，为什么不保留之前的树的统计数据呢?

> > 人机对弈的时候还要考虑人一方的move，另外即使保留也只是两个move之后的子树，考虑到每一层的分支数量，两层之后的一个分支中的数据其实很少 > > 那在self_play的过程中为什么就要保存下来呢？是因为self_play的过程中页是经过类似的搜索过程对整颗mcts树上的值有变动吗？个人理解如果人机的话，只有一方搜索，相当于树更多的包含了对机器有利的分支，因为人很可能走对他有利的分支，这个分支很可能在树里面比较少探索的深度比较有限，所以保留下来意义也不大，数据太少。而如果self play 的时候，两个方向都有一定深度，所以适合共享

关于KL散度控制学习率的问题

能问下里面的kl的阈值和学习率调整参数的值是凭经验吗？还是有一定计算方法？谢谢

FLUME-3149 reduce cpu cost for taildir file source while still maintaining reliability by using posFile in memory channel

I think this change is essentially reaching reliability in MemoryChannel when using taildirSource. So the performance diff is essentially the diff of MemoryChannel and FileChannel. In my tests it saves...

fluid.io.load_inference_model 载入多个模型的时候会报错Error Occurs, info:enforce version == 0U failed, 1015534361 != 0 Only version 0 is supported

solved by myself. for those who need it: use a new scope for every model ```python scope = fluid.Scope() with fluid.scope_guard(scope): place = fluid.CPUPlace() exe = fluid.Executor(place) [inference_program, _, fetch_targets]...