The official implementation of SeMAIL.
conda env create -f SeMAIL.yaml
conda activate SeMAIL
To train agents for each environmnet download the expert data from the provided link and run:
# car racing
python -u SeMAIL.py --expert_dir ../data/car_racing_expert --steps 500000 --seed 2021 --image_size 64 --batch_size 64 --task gym_car_racing_none-none --video_datadir ../driving_car_16/driving_car_16 --log_dir logdir/SeMAIL --disen_rec_scale 1.0
# locomotion tasks
python -u SeMAIL.py --expert_dir ../data/walker_run_expert --steps 1000000 --seed 2021 --image_size 64 --batch_size 64 --task dmc_walker_run_driving-driving --video_datadir ../driving_car_16/driving_car_16 --log_dir logdir/SeMAIL --disen_rec_scale 0.25
To train agents for each environmnet from expert demonstration data with different video backgrounds, run:
python -u SeMAIL.py --expert_dir ../data/walker_run_expert-8 --steps 1000000 --seed 2021 --image_size 64 --batch_size 64 --task dmc_walker_run_driving-driving --video_datadir ../driving_car_16/8 --log_dir logdir/SeMAIL-8 --disen_rec_scale 0.25
The training plots and output visualizations will be in the log folder:
tensorboard --logdir logdir
We appreciate the following github repos where we borrow code from: