/LaneGCN

[ECCV2020 Oral] Learning Lane Graph Representations for Motion Forecasting

Primary LanguagePythonOtherNOASSERTION

Caution

The Amazon AWS S3 bucket argoai-argoverse which holds many of the demo files has been compromised. The files may be corrupted.
Files referencing this S3 bucket have been modified, and any retrievals from this bucket are commented out. Please proceed with caution.

LaneGCN: Learning Lane Graph Representations for Motion Forecasting

Paper | Slides | Project Page | ECCV 2020 Oral Video

Ming Liang, Bin Yang, Rui Hu, Yun Chen, Renjie Liao, Song Feng, Raquel Urtasun

Rank 1st in Argoverse Motion Forecasting Competition

img

Table of Contents

Install Dependancy

You need to install following packages in order to run the code:

  1. Following is an example of create environment from scratch with anaconda, you can use pip as well:
conda create --name lanegcn python=3.7
conda activate lanegcn
conda install pytorch==1.5.1 torchvision cudatoolkit=10.2 -c pytorch # pytorch=1.5.1 when the code is release

# install argoverse api
pip install  git+https://github.com/argoai/argoverse-api.git

# install others dependancy
pip install scikit-image IPython tqdm ipdb
  1. [Optional but Recommended] Install Horovod and mpi4py for distributed training. Horovod is more efficient than nn.DataParallel for mulit-gpu training and easier to use than nn.DistributedDataParallel. Before install horovod, make sure you have openmpi installed (sudo apt-get install -y openmpi-bin).
pip install mpi4py

# install horovod with GPU support, this may take a while
HOROVOD_GPU_OPERATIONS=NCCL pip install horovod==0.19.4

# if you have only SINGLE GPU, install for code-compatibility
pip install horovod

if you have any issues regarding horovod, please refer to horovod github

Prepare Data: Argoverse Motion Forecasting

You could check the scripts, and download the processed data instead of running it for hours.

bash get_data.sh

Training

[Recommended] Training with Horovod-multigpus

# single node with 4 gpus
horovodrun -np 4 -H localhost:4 python /path/to/train.py -m lanegcn

# 2 nodes, each with 4 gpus
horovodrun -np 8 -H serverA:4,serverB:4 python /path/to/train.py -m lanegcn

It takes 8 hours to train the model in 4 GPUS (RTX 5000) with horovod.

We also supply training log for you to debug.

[Recommended] Training/Debug with Horovod in single gpu

python train.py -m lanegcn

Testing

You can download pretrained model from here

Inference test set for submission

python test.py -m lanegcn --weight=/absolute/path/to/36.000.ckpt --split=test

Inference validation set for metrics

python test.py -m lanegcn --weight=36.000.ckpt --split=val

Qualitative results

Labels(Red) Prediction (Green) Other agents(Blue)


Quantitative results img

Licence

check LICENSE

Citation

If you use our source code, please consider citing the following:

@InProceedings{liang2020learning,
  title={Learning lane graph representations for motion forecasting},
  author={Liang, Ming and Yang, Bin and Hu, Rui and Chen, Yun and Liao, Renjie and Feng, Song and Urtasun, Raquel},
  booktitle = {ECCV},
  year={2020}
}

If you have any questions regarding the code, please open an issue and @chenyuntc.