chensisi0730
chensisi0730
> Not a bug, a feature: this beat-tracking algorithm is causal, so that it can be used in real-time. It may take up to 6 seconds to pick-up a beat...
> Hi @Lord-Kamina > > The beats should be synchronized: the algorithm outputs **predicted** beats at their expected time. Have a look at aubiotrack output to see if it matches...
can you test first beat of the first seconds ?
> @qiwang067 > @amy-jiang-fiirst 感谢您的反馈，回复如下图所示： >  > > > > 文中有一句“就拿实际执行的这个动作，先取个 one-hot 向量，然后再拿到神经网络预测的动作概率，这两个一相乘，我就可以拿到算法里面的那个 ln𝜋(𝐴𝑡∣𝑆𝑡,𝜃)”，请问神经网络预测的动作概率不就是𝜋(𝐴𝑡∣𝑆𝑡,𝜃)吗？算法里的ln𝜋(𝐴𝑡∣𝑆𝑡,𝜃)直接对神经网络输出取对数不就可以了嘛？ > > 有推理的依据吗？还是拍脑袋出来的？
就我觉得符号体系混乱吗？策略一会是p 一会是π,并且和前三章体系也不同，
i think connected to edition ， in v0.8.8 , it is "mobile_calculators"
i have , can you need i t ?
so now are you get it ？
特别是增加train_eps到原来的10倍后，能看得很清楚！！ actor_loss = self.critic(state, self.actor(state)) # 个人理解critic不应该更新啊？？？！！