Task-Driven Graph attention for Hierarchical Object Navigation

Installation

Note: You may use conda instead of mamba

git clone --recursive [email protected]:mjlbach/ssg
cd ssg
conda env create -f environment.yml
conda activate ssg
bash install.sh
source .env
dvc pull

Running

Before running, you will want to specify the path to ig_assets and ig_dataset:

export GIBSON_ASSETS_PATH=/home/michael/Documents/ig_data/assets
export IGIBSON_DATASET_PATH=/home/michael/Documents/ig_data/ig_dataset

This can be done automatically via a .envrc file and direnv.

Uses hydra, select the experiment with +experiment=path

python scripts/train.py +experiment=search ++experiment_save_path=/svl/u/mjlbach/ray_results ++experiment_name=search_test

Note: ++eval_frequency=0 and ++num_envs=1 will help when running locally

Updating dependencies

Requires git 1.8.2 or above:

git submodule update --recursive --remote

If you did not clone the repo with --recursive, you will need to run this first:

git submodule update --init --recursive

Updating

conda env update --file environment.yml --prune

Technical debt

Notes

• Should use ~ 5 gb RAM per worker on full Rs_int scene
• Needs about 50 gb RAM for 8 workers + driver