storeblank/dfs

DFS: distillation from state-based policy

DFS: Distillation From State-based Policy

A project under developing and exploration

Structure

configs: config yaml files for starting an experiment

Agent: An entity that can act or learn in an environment, like a living creature. Parts of the agents (like organ) please put in module.py. Now have state-based sac, visual-based sac and vanilla bc agent

Algorithm: A blueprint introduces how to train an agent. Now have only vanilla state-based and visual-based sac

Environment: Openai gym-styled environments, have methods like step(), reset() and so on. Now have dm-control and distracting cs (mention that you need to download dm-control pkg: pip install dm-control)

utils.py: Other reusable tools and components are here, specially ReplayBuffer and ContrastBuffer

TODO

Agent:

1. Enable bc agent to update from representation, refer to https://github.com/HobbitLong/RepDistiller, you may first complete utils.py. Don't use AliasMethod in that repo, just uniformly sample. -- Done!

Algorithm:

1. (Ours) First get a state-based sac, then use DAgger (http://arxiv.org/abs/1011.0686) and contrastive representation distillation (crd) to train a visual-based actor. You can treat it as to use crd_loss + bc_loss in the update step of DAgger. -- distill actor but not critic (which still need state infomation in training)

Others:

1. maniskill2 env
2. crd in both critic and actor? so that we can continue training on just observation