Traffic4cast 2021

This repository includes 3DResNets and SparseUNet used for NeurIPS Traffic4Cast Competition 2021. A short paper that describes all details can be found at http://arxiv.org/abs/2111.05990.

Update 14 Dec 2021: Our works have been presented at Traffic4Cast Special Session at NeurIPS 2021. The slides can be found under this repository. Thank all organizers and speakers who made this fabulous event. More information: https://www.iarai.ac.at/traffic4cast/

Requirements

The conda environment with PyTorch1.9.1 and cuda11.1 is recommended.

Other essential packages:

pip install einops GPUtil tensorboard argparse pandas h5py

To run the SparseUnet, you will need to build and install MinkowskiEngine with the following commands:

conda install openblas-devel -c anaconda
find ${CONDA_PREFIX}/include -name "cblas.h"
export CUDA_HOME=$(dirname $(dirname $(which nvcc)))
export MAX_JOBS=1
pip install -U git+https://github.com/NVIDIA/MinkowskiEngine -v --no-deps --install-option="--force_cuda" --install-option="--blas_include_dirs=${CONDA_PREFIX}/include"  --install-option="--blas=openblas"

Model Training

1. Make sure all training files are correctly stored under the data folder. For example, a typical training file should be like: data/raw/BERLIN/training/2019-01-03_BERLIN_8ch.h5
2. Run the following command under the models folder for training different models:

python Resnet3D.py
python SparseUNet.py

3. Check the training logs

tensorboard --logdir="logs"

Checkpoints

Trained models can be found in models/checkpoints folder. Run the following codes to load the model state:

model_state_dict = torch.load('./checkpoints/Resnet3D.pk') # or SparseUNet
model.load_state_dict(model_state_dict)

The complete submission files can be generated by using models/submission.ipynb.

Have fun!