This repository is the official implementation of Dypre. Our implementation is based on SAC+AE.
Required dependencies of this repo can be installed by running:
conda env create --name dypre python==3.8
pip install -r requirements.txt
pip install git+https://github.com/denisyarats/dmc2gym.git
conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.3 -c pytorch
conda install tensorboard Then you can activate the environment by running:
source activate py3.6 To train a Dypre agent on the cartpole swingup task with images, run the command:
python train.py \
--domain_name cartpole \
--task_name swingup \
--action_repeat 8 \
--save_tb \
--seed 0 \
--eval_freq 1250 \
--batch_size 512 \
--results_dir ./logs \
--time_step 2 \
--omega_dypre_loss 0.01 \
--fc_output_logits True \
--kl_use_target True \
To evaluate Dypre on all 6 control tasks, run the train.sh scriptor:
bash train.sh