Skip to content

Latest commit

 

History

History
147 lines (108 loc) · 4.66 KB

README.md

File metadata and controls

147 lines (108 loc) · 4.66 KB

Code base used for our distillation project.

Extended from the:

Official codebase for the paper Teachable Reinforcement Learning via AdviceDistillation.

The codebase contains elements from the BabyAI platform, the meta-mb codebase, the d4rl codebase, and denisyarats/pytorch_sac codebase.

Replicating NeurIPS Results

The main branch of this repository is updated frequently. If you are looking to replicate or compare against the results from the NeurIPS paper, please use the code on the neurips_21 branch.

Installation

First, clone the repo.

Run the following commands:

git clone https://github.com/AliengirlLiv/teachable.git
conda create --name teachable_rl python=3.7
cd teachable
conda activate teachable_rl
sudo apt install libglew-dev libsm6 libxrender1 libosmesa6-dev patchelf
pip install -r reqs.txt

Add the project directory MY/PATH/TO/PROJECT/teachable to your PYTHONPATH.

If you want to use the Point Maze or Ant MAze environments, install Mujoco.

Add these to your ~/.bashrc file (note version numbers may have changed):

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:~/.mujoco/mujoco210/bin

Then run:

cd envs/d4rl
pip install -v --editable .

The most common installation error is mujoco not installing correctly. To check if this is the problem, open a python interpreter and run import mujoco_py.

Usage

To train a model, run the train_model.py script from the teachable directory.

Example: python scripts/train_model.py --prefix EXP_NAME --collect_with_rl_policy --level 2 --frames_per_proc 200 --num_envs 20 --rl_teacher OffsetWaypoint --env point_mass

Useful Commands

RL training

python scripts/train_model.py --prefix EXP_NAME --collect_with_rl_policy --train_level --frames_per_proc 200 --num_envs 20 --rl_teacher OffsetWaypoint --env point_mass --n_advice 1000000 --seed 1 --reward_type oracle_dist

Distillation to no advice

python scripts/train_model.py --buffer_capacity 1000000 --frames_per_proc 40 --num_envs 20 --collect_policy PATH_TO_RL_TRAINED_POLICY --distill_teacher none --collect_teacher OffsetWaypoint --env point_mass --level 2 --n_advice 300000 --seed 1 --prefix EXP_NAME

Distilling from one advice form to another

python scripts/train_model.py --prefix EXP_NAME --buffer_capacity 1000000 --train_level --frames_per_proc 200 --num_envs 20 --distillation_steps 15 --collect_policy PATH_TO_RL_TRAINED_POLICY --distill_teacher OTHER_TEACHER_NAME --collect_teacher OffsetWaypoint --env point_mass --n_advice 1000000 --seed 1

--prefix
persisted_models_distill/original_distillation
--collect_policy
logs/persisted_models_distill/claire
--rl_policy
logs/persisted_models_distill/claire
--level
2
--frames_per_proc
200
--num_envs
100
--collect_teacher Direction
--distill_teacher OffsetWaypoint
--env point_mass
--save_option all
--n_itr 10000

Running experiments faster

  • Use the easiest form of advice (for PointMaze, try OffsetWaypoint or Direction advice)
  • Use dense rewards for RL training. This should be enabled by default. In the experiments spreadsheet we default to using a semi-sparse reward which improves advice efficiency but decreases sample efficiency.
  • Use a small maze size.
  • Consider training on a single maze rather than a procedurally generated set of mazes (though I'd expect this change would mean advice doesn't get learned as well/will be less likely to generalize to other mazes)
  • If you use babyAI, there are even smaller mazes than those used in the paper. These could be faster.
  • Use a GPU, if you have one available.
  • You might see small efficiency gains if you sweep over learning rate and frames_per_proc

Experiments we ran

Distillation

--prefix persisted_models_distill/original_distillation_cont
--collect_policy logs/persisted_models_distill/claire
--collect_teacher Direction
--rl_policy logs/persisted_models_distill/claire
--level 2
--frames_per_proc 200
--num_envs 20

--distill_teacher OffsetWaypoint
--env point_mass
--save_option all
--n_itr
10000

This runs with: rl_agent will be None distilling_agent is a newly initialized PPO Agent relabel_policy is None

Collect Teacher is Direction -> collect_policy is the Grounded Agent with weights

--collect_with_rl_policy
--rl_policy logs/persisted_models_distill/grounded
--rl_teacher Direction
--use_input_converter
--env point_mass
--level 2
--frames_per_proc 200
--num_envs 20
--save_option
all
--n_itr
10000