Lisheng Wu

Hi, thanks for the shared code. I assume `gaussian_identity_variance` distribution of the[ decoder distribution](https://github.com/rail-berkeley/rlkit/blob/master/rlkit/torch/vae/conv_vae.py#L238) is a standard normal distribution. However, in the function [`compute_log_p_log_q_log_d`](https://github.com/rail-berkeley/rlkit/blob/354f14c707cc4eb7ed876215dd6235c6b30a2e2b/rlkit/torch/vae/vae_trainer.py#L68), the variance was taken as logvar...

I found that it may be not that fast in darknet

1

I want to know how fast YOLO extract features, however, I found it quite slow than darknet.

For example, if I want to run the script ` python baselines/her/experiment/train.py --num_cpu 19 --env_name HandManipulateBlock-v0 --n_epochs 200 --replay_strategy future` Will each mpi worker hold an individual replay buffer? Within...