facebookresearch / deep_bisim4control

Learning Invariant Representations for Reinforcement Learning without Reconstruction
Other
145 stars 37 forks source link

Reproducing the Generalization Experiments #5

Closed rraileanu closed 3 years ago

rraileanu commented 3 years ago

Hi!

I would like to reproduce the distractor and natural video environments used in the generalization experiments in Section 6.2 (as well as perhaps the results themselves). What are the commands for creating and training on these environments? It seems like one needs to specify args.total_frames as well as a directory with videos to replace the standard backgrounds (i.e. the args.resource_files). Could you specify how to generate these videos or add a download link?

Thank you!

amyzhang commented 3 years ago

Hi!

Apologies for the late reply. The hyperparameters to try are the ones in the scripts, you can try run_local.sh. You can download the videos from here: https://deepmind.com/research/open-source/kinetics It is the Kinetics 400 dataset, and we used the 'driving_car' train set. I will add a download link. You can just pass the directory you save those videos to to --resource_files