Aleksei Petrenko
Aleksei Petrenko
On a fresh Ubuntu 20 installation, after installing all the Linux dependencies from the README, the linking during installation still fails with the following error message: ``` [ 35%] Linking...
Multiplayer recording does not seem to work. I tried to run the example script `record_multiplayer.py`, it goes through 1 multiplayer episode successfully, but then the recorded .lmp file is nowhere...
* Values of certain variables (DEATHCOUNT, DEATHCOUNT, HITCOUNT) are not reset to 0 on the episode boundary. They keep growing as the agent plays an episode after episode. * ViZDoom...
@edbeeching I will leave this here so we don't forget. Notes: - cpu_affinity() returns only half of the cores (40 instead of 80) which leads to workers only using 50%...
Good results on Atari is a must for any popular RL framework. Scope of this work: 1) Create `train_atari.py` and `enjoy_atari.py` scripts similar to `train_vizdoom` or `train_isaacgym` that would register...
The env info caching should be improved too. First of all, we need to figure out env info before we can create any of the actual sf2 processes because we...
* Make sure we have an example folder and we can run 6 standard Mujoco environments * Try to match results from OpenRLBenchmark in terms of sample efficiency * Try...
This is a very desirable feature, especially to push the throughput of single-agent training to 200K FPS and beyond. Plan: use NCCL and/or Torch DistributedDataParallel. We can spawn one learner...
This adds a new feature that allows real-time evaluation and visualization of agents during the training session. The evaluation worker is supposed to run in a separate process from the...
Ideally, we want a worker that can render an episode every x minutes and post a gif animation or video on Tensorboard. The best solution (I think) is to repurpose...