question about step setting
Closed this issue · 2 comments
Wangbo-CASIA commented
您好 请教一个问题 dp3.yaml中设置的
horizon: 4
n_obs_steps: 2
n_action_steps: 4
实际获取的每个样本为连续的四帧；
怎么做到 2步观测+4步推理的 （正常我们不应该取连续的6帧吗）
YanjieZe commented
it is like this:
| o | o |
| a | a | a | a |
we actually use the last 3 actions.
(let me update the config maybe hhh, it makes people confused)
Wangbo-CASIA commented
那为啥网络预测的action 依然是(batch, 4, action_dim) 并且我看loss mask也没有对第一个action进行掩码
