VN-Transformer Paper | VectorNeuron Paper
Download the data from here. Run the following script to train the model (it takes around 1 min per epoch on NVIDIA 2080).
python train_modelnet_cls.py --data_path=path/to/modelnet40_normal_resampled
So far, I have not been able to replicate the results from the paper. Currently, running a hyperparameter search based on Table 5. The best I have gotten is 82.7% test acc after 1000 epochs (32hidden, 4heads, No latent).
If you use this repo, please consider citing the original works:
@article{assaad2022vn,
title={VN-Transformer: Rotation-Equivariant Attention for Vector Neurons},
author={Assaad, Serge and Downey, Carlton and Al-Rfou, Rami and Nayakanti, Nigamaa and Sapp, Ben},
journal={arXiv preprint arXiv:2206.04176},
year={2022}
}
@article{deng2021vn,
title={Vector Neurons: a general framework for SO(3)-equivariant networks},
author={Deng, Congyue and Litany, Or and Duan, Yueqi and Poulenard, Adrien and Tagliasacchi, Andrea and Guibas, Leonidas},
journal={arXiv preprint arXiv:2104.12229},
year={2021}
}
Many of the vector neuron modules are taken from the VNN repo.
- Replicate results on ModelNet40 Classification
- Implement late-fusion model architectures (Figure 4)
- Test with non-spatial attributes