Thinklab-SJTU/Bench2Drive

关于自建训练集的expert_assessment

wyield opened this issue · 4 comments

    通过浏览anno.md和下载的数据集可以看到数据结构中还包含了expert_assessment,请问该数据是什么结构的?你们是否对自动生成该类数据汇总了比较好的办法?为什么这类数据索引对于其他数据都-1?

@wyield Please read the paper for what those information mean. The -1 issue is because the evaluation of the current environment is from the value function of RL expert in the next step. It is related to RL and Markov chain.

您搞懂了expert_assessment中是什么数据了嘛?我遇到了与您一样的疑问,我搞不清expert_assessment里面保存了什么数据。

Hi @czg2066, the expert_assessment include 3 parts, expert_feature, RL value and action.

  • action = expert_assessment[-1], its range is [0, 38].
  • value = expert_assessment[-2], it is the output from RL value head.
  • feature = expert_assessment[:-2], it is the feature which feeds to value/actor head.

Hi @czg2066, the expert_assessment include 3 parts, expert_feature, RL value and action.您好,expert_assessment包括 expert_feature、RL 值和操作 3 个部分。

  • action = expert_assessment[-1], its range is [0, 38].action = expert_assessment[-1],则其范围为 [0, 38]。
  • value = expert_assessment[-2], it is the output from RL value head.value = expert_assessment[-2],它是 RL value head 的输出。
  • feature = expert_assessment[:-2], it is the feature which feeds to value/actor head.feature = expert_assessment[:-2],它是馈送到 value/actor head 的特征。

ok,thanks.