Cannot reproduce Atari Pong scores

[mlinda96]

I'm currently exploring the Dreamer models and papers and wanted to get some insight from the code.

However, I'm encountering a few issues with reproducibility.

I initially attempted to replicate the issue in Google Colab using gym==0.19.0, but found that this version is no longer installable via pip on new instances. Instead, I installed gym==0.22.0 and made minor modifications to common.envs.py. As a result, the code now runs smoothly in the Colab environment. https://colab.research.google.com/drive/1xpqS4ovnohsqrZoWAmA3N6cRtA0MPtX4#scrollTo=BASD1Km60aKk

The main issue I'm having is reproducing the performance metrics you've previously published. Specifically, the reward on Pong remains stagnant, and the image loss is consistently around 3772. This leads me to believe there may be an underlying issue or bug affecting the model's performance.

I'm just running the basic example with python train.py --logdir .logdir/atari_pong/ --configs atari --task atari_pong

Am I doing something wrong? I love the idea behind Dreamer and would love to play around with the code.

Thank you for your time and effort