Skip to content

Latest commit

 

History

History
53 lines (33 loc) · 1.99 KB

README.md

File metadata and controls

53 lines (33 loc) · 1.99 KB
Raw

DePO codes on NGSIM driving experiments

Official Pytorch implemetation of ICML2022 paper Depo (Plan Your Target and Learn Your Skills: Transferable State-Only Imitation Learning via Decoupled Policy Optimization) on NGSIM driving dataset. Since some settings of the driving environment are quite different from the Mujoco environments, we maintain the experiment code on NGSIM separately in this repo. The code for the main experiments can be found in DePO. Current simulation platform is a (slightly) modified version of PPUU. We plan to open source the training code based on the SMARTS simulator, so be sure to stay tuned!

Important Notes

This repository is based on ILSwiss. If you are looking for DePO for Mujoco experiments, check here.

Reproducing Results

Download and Transform Driving Dataset

Go to this address and download the TGZ file (330 MB) on your machine. Place the downloaded folder as pytorch-PPUU/traffic-data.

Then transform the dataset so that it can fit in ILSwiss.

cd ILSwiss/
python run_experiment.py -e exp_specs/expert_ppuu.yaml
python run_experiment.py -e exp_specs/expert_ppuu_multitype.yaml

Example scripts

Co-training of DePO on multiple action spaces

python run_experiment.py -e exp_specs/dpo_ppuu_multitype.yaml

Transferring of DePO

  1. Pretrain on Normal
python run_experiment.py -e exp_specs/dpo_ppuu_pretrain.yaml
  1. Transfer to Inverse and Transpose

Remember to change state_predictor_path in exp_specs/dpo_ppuu_transfer.yaml to the path of best.pkl file obtained from pretraining.

python run_experiment.py -e exp_specs/dpo_ppuu_transfer.yaml

Baseline Algorithm

Run GAIfO.

python run_experiment.py -e exp_specs/gailfo_ppuu.yaml