MundVetter / vnn_dl2

Vector Neurons: A General Framework for SO(3)-Equivariant Networks

Vector Neurons++: Extending Neural Dimensionality and Generalizing Activation Functions for Vector Neuron Networks

This code base is forked from VN-PointNet Deng et al.. We introduce the addition of arbitrary activation and inclusion of normals in the VN-layers, specifically for PointNet. For DGCNN we refer to https://github.com/CSteigstra/vnn-pc which is based on its correct implementation from https://github.com/FlyingGiraffe/vnn-pc .

Overview

vnn++ is the author's implementation of Vector Neuron Networks with PointNet and DGCNN backbones. The current version only supports PointNet for Modelnet40 classification.

Environment

conda env create -f dl2_gpu.yml
conda activate dl2
# or 
source activate dl2

Data Preparation

• Classification: Download ModelNet40 and save in data/modelnet40_normal_resampled/.

Usage

Classification on ModelNet40

Training

# Author's (Deng et al.) LeakyReLU
python train_cls.py --log_dir LOG_DIR --model vn_pointnet_cls --optimizer Adam --rot z
python train_cls.py --log_dir LOG_DIR --model vn_pointnet_cls --optimizer Adam --rot z --normal
# Ours
python train_cls.py --log_dir LOG_DIR --model vn_pointnet_cls --optimizer Adam --rot z --activ leaky_relu
python train_cls.py --log_dir LOG_DIR --model vn_pointnet_cls --optimizer Adam --rot z --activ leaky_relu --normal

Evaluation

# Author's (Deng et al.) LeakyReLU
python test_cls.py --log_dir LOG_DIR --model vn_pointnet_cls --rot so3
python test_cls.py --log_dir LOG_DIR --model vn_pointnet_cls --rot so3 --normal
# Ours
python test_cls.py --log_dir LOG_DIR --model vn_pointnet_cls --rot so3 --activ leaky_relu
python test_cls.py --log_dir LOG_DIR --model vn_pointnet_cls --rot so3 --activ leaky_relu --normal

Citation

In the works. Refer to our github for now. ^.^

License

MIT License

Acknowledgement

The structure of this codebase is borrowed from this pytorch implementataion of PointNet/PointNet++ and DGCNN as well as this implementation.