bramton / point-transformer

This is an unofficial implementation of the Point Transformer paper.

Point-Transformer PyTorch

• Implemention of Point Transformer.
• Code based on PointNet2 Pytorch repository.

Setup

• Install python -- This repo is tested with {3.6, 3.7}

• Install pytorch with CUDA -- This repo is tested with {1.4, 1.5, 1.7.1}. It may work with versions newer than 1.7.1, but this is not guaranteed.

• Install dependencies

  pip install -r requirements.txt
  

Training

Install with: pip install -e .

The example training script can be found in point_transformer/train.py. The training examples are built using PyTorch Lightning and Hydra.

You can train a Point Transformer model on various tasks as,

# train Point Transformer for classification task on ModelNet40
python -m point_transformer.train task=cls

# train Point Transformer for part segmentation task on ShapeNet
python -m point_transformer.train task=partseg

# train Point Transformer for semantic segmentation task on S3DIS
python -m point_transformer.train task=semseg

If you want to override the default config, you can pass the command line arguments,

# Change the batch size to 32
python -m point_transformer.train task=cls batch_size=32

Building only the CUDA kernels

pip install point_transformer_lib/.

Experiment Results

• Classification on ModelNet40

Model	mAcc	OA
Paper	90.6	93.7
Our Implemention		87.2

• Part Segmentation on ShapeNet

Model	cat. mIoU	ins. mIoU
Paper	83.7	86.6
Our Implemention		83.2

• Semantic Segmentation on S3DIS Area5

Model	mAcc	OA	mIoU
Paper	76.5	90.8	70.4
Our Implemention		88.0

Contributing

This repository uses black for linting and style enforcement on python code. For c++/cuda code, clang-format is used for style. The simplest way to comply with style is via pre-commit

pip install pre-commit
pre-commit install