Pytorch version of Dreamer, which follows the original TF v2 codes (https://github.com/danijar/dreamerv2/tree/faf9e4c606e735f32c8b971a3092877265e49cc2).
Due to limitation of resource, we tested it with the official TF version2 codes for 10 tasks (5 DMC / 5 Atari tasks) as below.
As we can see, for almost tasks, it shows similar performance with Dreamer when comparing the reported performance on paper and from running codes.
For freeway, I found that it is slower than Dreamer, maybe I think it is because of random seed. I will find the reason more.
The below logs are from Tensorboard, you also can see through Tensorboard logs.
- DMC walker walk task
- Atari bank heist task
- atari freeway task
pip install -r requirements.txt
you also need to execute drivergym_config/prepare.sh and make sure the data is in ../../../data/drivergym/ (or modify the reference in code)
To debug
python3 dreamer.py --logdir ~/logdir/DriverDream/1 --configs defaults drivergym debug --task drivergym_
for more information options, you can check.
Any feedback are welcome! Please open an issue on this repository or send email to Jaesik Yoon (jaesik.yoon.kr@gmail.com).