You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Hi, If I needed to sample entire trajectories for a single env from the RolloutStorage, what would be an easy way to do so?
For example, would it be easier to adapt the recurrent_generator -- any hints would be really appreciated (haven't dealt much with PPO, so this may be a really stupid question)!
Context: I want to randomly sample 2 trajectories (possibly from different envs) and add an auxiliary loss which depends on these two trajectories.
The text was updated successfully, but these errors were encountered:
The only issue is that you might end up with some partial trajectories since the first observation in rollouts might be a continuation of a trajectory collected during the previous update (so it's not necessarily the initial observation in an environment). But I think this can be fixed by removing the first trajectory that you are adding i.e. start from prev_index = indices[0] rather than prev_index = 0.
Hi, If I needed to sample entire trajectories for a single env from the
RolloutStorage
, what would be an easy way to do so?For example, would it be easier to adapt the
recurrent_generator
-- any hints would be really appreciated (haven't dealt much with PPO, so this may be a really stupid question)!Context: I want to randomly sample 2 trajectories (possibly from different envs) and add an auxiliary loss which depends on these two trajectories.
The text was updated successfully, but these errors were encountered: