Author's Pytorch implementation of SCQ (Paper Link).
-
Install Mujoco binaries
-
Create Anaconda environment
conda create -n scq python=3.8.5 pip conda activate scq
We use
python=3.8.5
to do the experiments in the paper. But we believepython=3.9
andpython=3.10
should also work as long as it is compatible with d4rl. -
Install the dependencies
pip install -r requirements.txt
For reference, authors use the following versions:
- gym==0.18.3
- d4rl==1.1
- cython==0.29.37
- torch==2.1.2
- torchrl==0.2.1
- numpy==1.24.4
- wandb==0.16.4
- tqdm==4.66.1
- pyyaml==6.0.1
- argparse==1.1
All the necessary parameters to reproduce the result can be found in the config
folder. You can just run each of the experiment by using the following command.
Ex1. HalfCheetah-medium-v2
python main.py --config config/halfcheetah/halfcheetah-medium-v2.yaml --seed 0
Ex2. Antmaze-medium-play-v2
python main.py --config config/antmaze/antmaze-medium-play-v2.yaml --seed 0
We use Weights and Biases to visualize all of the logging data and results. In order to access these information, you need to set up your own W&B API keys at the first experiment.
Some of our code come from the following repository. We appreciate these authors to share their valuable codes.
If you have any questions, suggestions and improvements, please feel free to send me message to purewater0901[at]berkeley[dot]edu