Reward loss timescale

[roggirg]

Hi,

I believe the reward loss should be based on rewards[1:] instead of rewards[:-1] :

dreamer-pytorch/main.py

Line 209 in 7e9050e

reward_loss = -reward_dist.log_prob(rewards[:-1]).mean(dim=(0, 1))

If not, can you please explain your reasoning?
Thanks,

[sumwailiu]

I think you are right. After fixing it (refers to the pull request #12) and other related operations, I found the issue #4 is solved (refers to #4 (comment))