infant-aagcn

PyTorch Implementation of Modeling 3D Infant Kinetics Using Adaptive Graph Convolutional Networks.

Environment

Cluster

module load pytorch/1.13
pip install -Ur requirements.txt
. ./env.sh

Docker

docker run -v $(pwd):/work/infant-stgnn -w /work/infant-stgnn --user $(id -u):$(id -g) --gpus all --shm-size 16g -it infant-aagcn

Preprocessing

Data (2 zip archives, 4.8GB) available upon request. Preprocessing expects .csv files with joint coordinates over time.

./unzip.sh
./preprocess.sh

• Extract zip archives
• Select time intervals
• Center skeleton around neck joint
• Rotate around principal axes
• Create feature streams

Training

By running the env script bin is put on path where all executables reside. The modules folder contain the models and dataloader etc.

train.py \
  --data-dir data/streams/combined \
  --output-dir results/aagcn \
  --age-file metadata/combined.csv \
  --learning-rate 0.01 \
  --batch-size 32 \
  --num-workers 16 \
  --streams j \
  --k-folds 10 \
  --epochs 20 \
  --adaptive \
  --attention

Slurm: sbatch run/submit.sh or for the full comparison sbatch run/experiment.sh

Training creates a results folder with all runs. The notebook folder then contains separate files for the ml baseline, aagcn inference and metrics calculation.

An example of how the models can be called to make predictions is avaialable in the submit script run/predict.sh.

Cite

@article{holmberg2024modeling,
  title={Modeling 3D Infant Kinetics Using Adaptive Graph Convolutional Networks},
  author={Daniel Holmberg and Manu Airaksinen and Viviana Marchi and Andrea Guzzetta and Anna Kivi and Leena Haataja and Sampsa Vanhatalo and Teemu Roos},
  journal={arXiv preprint arXiv:2402.14400},
  year={2024}
}