search

google-research / batch_rl

Offline Reinforcement Learning (aka Batch Reinforcement Learning) on Atari 2600 games
https://offline-rl.github.io/
Apache License 2.0
528 stars 74 forks source link

Native Resolution Images #13

Closed MatthewChang closed 3 years ago

MatthewChang commented 3 years ago

Hi,

I've been trying to use the data from rl unplugged in its native resolution (210x160). I hoping to replay the rl unplugged actions from the sequential data release for dopamine into the environment to regenerate the frames. However I think sticky actions is making the behavior not align perfectly. Is the script that was used to generate the data publicly available somewhere? Or is there another way I can use the rl unplugged data in the native resolution? Thanks!

agarwl commented 3 years ago

The agents in LoggedDQNAgent in this repo here were used to collect the data using the gin config present here. You can just use the run_experiment.py script to launch these experiments. It will take about 3-5 days to collect the entire data. Please note that Dopamine does preprocessing of Atari frames to grayscale images and you'd have to write additional code for getting the native resolution images).

Regarding RL Unplugged, If I remember correctly, it used the data from DQN Replay data, so it probably doesn't contain the native resolution images. As an aside, I think prior work on regenerating the frames has probably used the preprocessed frames for regeneration, although I am not sure.