JSparse

Introduction

JSparse is a high-performance auto-differentiation library for sparse voxels computation and point cloud processing based on Jittor, TorchSparse and Torch Cluster.

Installation

If you use cpu version, you need to install Google Sparse Hash, and choose the convolution algorithm with "jittor".

The latest JSparse can be installed by

python setup.py install # or
python setup.py develop

Getting Started

Architecture

- jsparse
    - nn
        - functional
        - modules
    - utils
        collate/quantize/utils.py

You can use the modules from jsparse/modules .

Sparse Tensor

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

• Coordinates (indices): a 2D integer tensor with a shape of $N \times 4$, where the first dimension denotes the batch index, and the last three dimensions correspond to quantized $x, y, z$ coordinates.

• Features (values): a 2D tensor with a shape of $N \times C$, 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 JSparse.utils.quantize) to voxelize $x, y, z$ coordinates and remove duplicates.

  You can also use the initialization method to automatically obtain the discretized features by turning on quantize option

  inputs = SparseTensor(values=feats, indices=coords, voxel_size=self.voxel_size, quantize=True)

  We can then use sparse_collate_fn (provided in JSparse.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

We finished many common modules in jsparse.nn such like GlobalPool.

The neural network interface in jsparse is similar to Jittor:

import jsparse.nn as spnn
def get_conv_block(self, in_channel, out_channel, kernel_size, stride):
    return nn.Sequential(
        spnn.Conv3d(
            in_channel,
            out_channel,
            kernel_size=kernel_size,
            stride=stride,
        ),
        spnn.BatchNorm(out_channel),
        spnn.LeakyReLU(),
    )

You can get the usage of most of the functions and modules from the example examples/MinkNet/classification_model40.py.

BenchMark

We test several networks between JSparse(v0.5.0) and TorchSparse(v1.4.0).

Because the Jittor framework is fast, inference and training are faster than PyTorch on many operators. We also speed up quantize with jittor’s operations and get better performance.

We test the speed on the following model and choose 10 scenes from ScanNet as the dataset.

import jsparse.nn as spnn
from jittor import nn
algorithm = "cuda"
model = nn.Sequential(
    spnn.Conv3d(3, 32, 2),
    spnn.BatchNorm(32),
    spnn.ReLU(),
    spnn.Conv3d(32, 64, 3, stride=1, algorithm=algorithm),
    spnn.BatchNorm(64),
    spnn.ReLU(),
    spnn.Conv3d(64, 128, 3, stride=1, algorithm=algorithm),
    spnn.BatchNorm(128),
    spnn.ReLU(),
    spnn.Conv3d(128, 256, 2, stride=2, algorithm=algorithm),
    spnn.BatchNorm(256),
    spnn.ReLU(),
    spnn.Conv3d(256, 128, 2, stride=2, transposed=True, algorithm=algorithm),
    spnn.BatchNorm(128),
    spnn.ReLU(),
    spnn.Conv3d(128, 64, 3, stride=1, transposed=True, algorithm=algorithm),
    spnn.BatchNorm(64),
    spnn.ReLU(),
    spnn.Conv3d(64, 32, 3, stride=1, transposed=True, algorithm=algorithm),
    spnn.BatchNorm(32),
    spnn.ReLU(),
    spnn.Conv3d(32, 3, 2),
)

We finnished two versions of Sparse Convolution(completed convolution function with jittor operators or cuda).

We choose attribute batch_size = 2, total_len = 10 and run on RTX3080 to test per iteration’s speed (JSparse’s version is v0.5.0 ). | | JSparse | TorchSparse(v1.4.0) | |——————-|—————|———————-| | voxel_size = 0.50 | 20.05ms | 33.66ms | | voxel_size = 0.10 | 25.15ms | 40.40ms | | voxel_size = 0.02 | 81.37ms | 87.42ms |

We also test the same 200 scenes of ScanNet on VMNet on JSparse and TorchSparse.

We choose attribute batch_size = 3, num_workers=16 and run on RTX Titan and Intel(R) Xeon(R) CPU E5-2678 v3 to test per iteration’s speed.

JSparse(cuda)	TorchSparse(v1.4.0)
0.79s	0.92s

If we ignore the initiation(scannet.py), and just test the speed of network, the speed of JSparse and TorchSparse is similar.

Acknowledgements

The implementation and idea of JSparse refers to many open source libraries, including(but not limited to) MinkowskiEngine and TorchSparse.

If you use JSparse in your research, please cite our and their works by using the following BibTeX entries:

@article{hu2020jittor,
  title={Jittor: a novel deep learning framework with meta-operators and unified graph execution},
  author={Hu, Shi-Min and Liang, Dun and Yang, Guo-Ye and Yang, Guo-Wei and Zhou, Wen-Yang},
  journal={Science China Information Sciences},
  volume={63},
  number={222103},
  pages={1--21},
  year={2020}
}

@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}
}