This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.
Implementation of Change-Based Exploration Transfer (C-BET), as presented in Interesting Object, Curious Agent: Learning Task-Agnostic Exploration.
This code was built on the RIDE repository.
conda create -n cbet python=3.8.10
conda activate cbet
git clone [email protected]:sparisi/cbet.git
cd cbet
pip install -r requirements.txt
- Follow the official guide
and do a full install with
habitat_baselines. - Download and extract Replica scenes in the root folder of
cbet
WARNING! The dataset is very large!
sudo apt-get install pigz
git clone https://github.com/facebookresearch/Replica-Dataset.git
cd Replica-Dataset
./download.sh replica-path
If the script does not work, manually unzip with
cat replica_v1_0.tar.gz.part* | tar -xz
-
Intrinsic-only pre-training:
OMP_NUM_THREADS=1 python main.py --model cbet --env <ENV_NAME> --no_reward --intrinsic_reward_coef=0.005 -
Extrinsic-only transfer with pre-trained model:
OMP_NUM_THREADS=1 python main.py --model cbet --env <ENV_NAME> --intrinsic_reward_coef=0.0 --checkpoint=path/to/model.tar
Pre-trained exploration policies are available as code release. -
Tabula-rasa training with summed intrinsic and extrinsic reward:
OMP_NUM_THREADS=1 python main.py --model cbet --env <ENV_NAME> --intrinsic_reward_coef=0.005
See src/arguments.py for the full list of hyperparameters.
For MiniGrid, <ENV_NAME> can be MiniGrid-DoorKey-8x8-v0, MiniGrid-Unlock-v0, ...
For Habitat, <ENV_NAME> can be HabitatNav-apartment_0, HabitatNav-hotel_0, ...
-
Offline testing (single):
python test_checkpoint.py --checkpoint=<PATH_TO_TAR> -
Offline testing (batch):
python test_models.py --logdir=<PATH_TO_LOG>(this tests all models in a log directory produced by a Slurm batch run).
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent