EQNet: Boost Efficiency for Query-Centric Multi-Agent Motion Forecasting

SETUP

PREPARE DATASET

• You need to first download Argoverse 2 motion forecasting dataset.

• You need to modify the data root in EQNet/env_config.yml:machines/default/argoverse_v2_dataroot to the correct location.

• The dataset folder structure example:

  /mnt/data/argoverse2/
  ├── test
  │   ├── processed
  │   └── raw
  ├── train
  │   ├── processed
  │   └── raw
  └── val
      ├── processed
      │   ├── 6a24bd-1ad8-44c7-9a0e-5e93bb41d5ae.pkl
      |   ...
      └── raw
          ├── 6a24bd-1ad8-44c7-9a0e-5e93bb41d5ae
          ...
  

  The processed folder contains the cache data, which is generated from the raw data. It will be automatically generated on the initial run. The raw folder contains the original data.

PREPARE ENVIRONMENT

git clone <this-repo>

# we use conda to manage system packages
# to compile cuda extensions: gcc=9.5.0 gxx=9.5.0
# evalai dependencies: lxml
# av2-api dependencies: rust
conda create --name py310rust python=3.10 gcc=9.5.0 gxx=9.5.0 lxml rust virtualenv -c conda-forge 
conda activate py310rust
pip install virtualenv  # we use virtualenv to manage python packages

cd <repo_root>
python3 -m venv .venv  # to be able to navigate the library code in the explorer tree, create the .venv folder locally
source .venv/bin/activate
pip install -U setuptools pip

<install-pytorch-via-pip-with-proper-version-of-cuda>  # accord with /usr/local/cuda version
<install-pytorch-geometric-via-pip-with-proper-version-of-cuda>  # https://pytorch-geometric.readthedocs.io/en/1.6.3/notes/installation.html

pip install https://github.com/argoverse/av2-api.git  # mkdir .venv/src ; cd .venv/src; git clone [email protected]:argoverse/av2-api.git; pip install -e av2-api; cd ../..
pip install https://github.com/JonathonLuiten/TrackEval.git  # mkdir .venv/src ; cd .venv/src; git clone [email protected]:JonathonLuiten/TrackEval.git; pip install -e TrackEval; cd ../..
pip install beautifulsoup4==4.7.1 beautifultable==0.7.0 "boto3>=1.9.88" docker==3.6.0 validators==0.12.6 termcolor==1.1.0 "tqdm>=4.49.0"  # manually manage dependencies for evalai to avoid conflicts
pip install evalai --no-deps

pip install lightning einops torchscale huggingface_hub tensorboard==2.12.1

pip install GitPython json5 shapely descartes imageio omegaconf colour pyyaml docstring_parser typing_inspect ipdb matplotlib click

pip install "numpy<2.0"

TRAINING

conda activate py310rust
source .venv/bin/activate
bash EQNet/train.sh -m

TESTING

TBA

BENCHMARK SPEED

TBA

VISUALIZATION

• checkout visualize.ipynb.

CITATION

TBA