/torchsparse

[MLSys'22] TorchSparse: Efficient Point Cloud Inference Engine

Primary LanguageCudaMIT LicenseMIT

TorchSparse

TorchSparse is a high-performance neural network library for point cloud processing.

Installation

TorchSparse depends on the Google Sparse Hash library.

  • On Ubuntu, it can be installed by

    sudo apt-get install libsparsehash-dev
  • On Mac OS, it can be installed by

    brew install google-sparsehash
  • You can also compile the library locally (if you do not have the sudo permission) and add the library path to the environment variable CPLUS_INCLUDE_PATH.

The latest released TorchSparse (v1.4.0) can then be installed by

pip install --upgrade git+https://github.com/mit-han-lab/torchsparse.git@v1.4.0

If you use TorchSparse in your code, please remember to specify the exact version in your dependencies.

For installation help and troubleshooting, please consult the Frequently Asked Questions before posting an issue.

Benchmark

We compare TorchSparse with MinkowskiEngine (where the latency is measured on NVIDIA GTX 1080Ti):

MinkowskiEngine v0.4.3 TorchSparse v1.0.0
MinkUNet18C (MACs / 10) 224.7 ms 124.3 ms
MinkUNet18C (MACs / 4) 244.3 ms 160.9 ms
MinkUNet18C (MACs / 2.5) 269.6 ms 214.3 ms
MinkUNet18C 323.5 ms 294.0 ms

Getting Started

Sparse Tensor

Sparse tensor (SparseTensor) is the main data structure for point cloud, which has two data fields:

  • Coordinates (coords): a 2D integer tensor with a shape of N x 4, where the first three dimensions correspond to quantized x, y, z coordinates, and the last dimension denotes the batch index.
  • Features (feats): a 2D tensor with a shape of N x C, where C is the number of feature channels.

Most existing datasets provide raw point cloud data with float coordinates. We can use sparse_quantize (provided in torchsparse.utils.quantize) to voxelize x, y, z coordinates and remove duplicates:

coords -= np.min(coords, axis=0, keepdims=True)
coords, indices = sparse_quantize(coords, voxel_size, return_index=True)
coords = torch.tensor(coords, dtype=torch.int)
feats = torch.tensor(feats[indices], dtype=torch.float)
tensor = SparseTensor(coords=coords, feats=feats)

We can then use sparse_collate_fn (provided in torchsparse.utils.collate) to assemble a batch of SparseTensor's (and add the batch dimension to coords). Please refer to this example for more details.

Sparse Neural Network

The neural network interface in TorchSparse is very similar to PyTorch:

from torch import nn
from torchsparse import nn as spnn

model = nn.Sequential(
    spnn.Conv3d(in_channels, out_channels, kernel_size),
    spnn.BatchNorm(out_channels),
    spnn.ReLU(True),
)

Citation

If you use TorchSparse in your research, please use the following BibTeX entries:

@inproceedings{tang2022torchsparse,
  title = {{TorchSparse: Efficient Point Cloud Inference Engine}},
  author = {Tang, Haotian and Liu, Zhijian and Li, Xiuyu and Lin, Yujun and Han, Song},
  booktitle = {Conference on Machine Learning and Systems (MLSys)},
  year = {2022}
}
@inproceedings{tang2020searching,
  title = {{Searching Efficient 3D Architectures with Sparse Point-Voxel Convolution}},
  author = {Tang, Haotian and Liu, Zhijian and Zhao, Shengyu and Lin, Yujun and Lin, Ji and Wang, Hanrui and Han, Song},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year = {2020}
}

Acknowledgements

TorchSparse is inspired by many existing open-source libraries, including (but not limited to) MinkowskiEngine, SECOND and SparseConvNet.