Datou
Datou
> I think Colossal-AI currently only supports Linux environment, i you are using Windows and you do not have a server, maybe you can try to use Colab or something....
Win11, 9900K, 32G. The file was downloaded completely without any issues. I tried downloading it again using the browser, but the file size and error message were the same. 
You got "bad magic" too.
 It doesn't work :(
I got the crop data from the mask and the problem is solved, thank you. 
> 现在我大概明白为什么作者要一个AI的指令持续6帧然后才输出奖惩值。这个是因为在游戏中,做出一个动作并不会马上打到对手并得分,要大概几个Frame的时间才能看到效果。如果我们把每一帧都丢给AI,当AI做对一个动作,并在几帧(step)后得到一个奖励,这个做对的动作实际得到的discounted return (刚刚学到的术语)就会相当小, 反而离奖励近的无效动作得到的discounted return多,这样并不利于AI训练。 > > 所以我有个想法,既要用作者的这个方法,也要让AI会发波,升龙。 原来是这样啊,我把这个加回去试试。
把动作持续几帧似乎立竿见影,随便一连就通关了,等会模型出来了我验证一下是真通关还是假通关。 
> 果然没有这种好事,代码写错了 T_T
 OpenAI官方支持聊天记录导出了
gym-retro官方好像放弃声音了,对于他们训练AI来说这不重要。这里有段声音代码,我试了一下,加不对地方,只能断断续续出点声。 https://github.com/openai/retro/issues/76#issuecomment-591928467