TorchSparse

News

2020/09/20: We released torchsparse v1.1, which is significantly faster than our torchsparse v1.0 and is also achieves 1.9x speedup over MinkowskiEngine v0.5 alpha when running MinkUNet18C!

2020/08/30: We released torchsparse v1.0.

Overview

We release torchsparse, a high-performance computing library for efficient 3D sparse convolution. This library aims at accelerating sparse computation in 3D, in particular the Sparse Convolution operation.

The major advantage of this library is that we support all computation on the GPU, especially the kernel map construction (which is done on the CPU in latest MinkowskiEngine V0.4.3).

Installation

You may run the following command to install torchsparse.

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

Note that this library depends on Google's sparse hash map project. In order to install this library, you may run

sudo apt-get install libsparsehash-dev

on Ubuntu servers. If you are not sudo, please clone Google's codebase, compile it and install locally. Finally, add the path to this library to your CPLUS_INCLUDE_PATH environmental variable.

For GPU server users, we currently support PyTorch 1.6.0 + CUDA 10.2 + CUDNN 7.6.2. For CPU users, we support PyTorch 1.6.0 (CPU version), MKLDNN backend is optional.

Usage

Our SPVNAS project (ECCV2020) is built with torchsparse. You may navigate to this project and follow the instructions in that codebase to play around.

Here, we also provide a walk-through on some important concepts in torchsparse.

Sparse Tensor and Point Tensor

In torchsparse, we have two data structures for point cloud storage, namely torchsparse.SparseTensor and torchsparse.PointTensor. Both structures has two data fields C (coordinates) and F (features). In SparseTensor, we assume that all coordinates are integer and do not duplicate. However, in PointTensor, all coordinates are floating-point and can duplicate.

Sparse Quantize and Sparse Collate

The way to convert a point cloud to SparseTensor so that it can be consumed by networks built with Sparse Convolution or Sparse Point-Voxel Convolution is to use the function torchsparse.utils.sparse_quantize. An example is given here:

inds, labels, inverse_map = sparse_quantize(pc, feat, labels, return_index=True, return_invs=True)

where pc, feat, labels corresponds to point cloud (coordinates, should be integer), feature and ground-truth. The inds denotes unique indices in the point cloud coordinates, and inverse_map denotes the unique index each point is corresponding to. The inverse map is used to restore full point cloud prediction from downsampled prediction.

To combine a list of SparseTensors to a batch, you may want to use the torchsparse.utils.sparse_collate_fn function.

Detailed results are given in SemanticKITTI dataset preprocessing code in our SPVNAS project.

Computation API

The computation interface in torchsparse is straightforward and very similar to original PyTorch. An example here defines a basic convolution block:

class BasicConvolutionBlock(nn.Module):
    def __init__(self, inc, outc, ks=3, stride=1, dilation=1):
        super().__init__()
        self.net = nn.Sequential(
            spnn.Conv3d(inc, outc, kernel_size=ks, dilation=dilation, stride=stride),
            spnn.BatchNorm(outc),
            spnn.ReLU(True)
        )

    def forward(self, x):
        out = self.net(x)
        return out

where spnndenotes torchsparse.nn, and spnn.Conv3d means 3D sparse convolution operation, spnn.BatchNorm and spnn.ReLU denotes 3D sparse tensor batchnorm and activations, respectively. We also support direct convolution kernel call via torchsparse.nn.functional, for example:

outputs = torchsparse.nn.functional.conv3d(inputs, kernel, stride=1, dilation=1, transpose=False)

where we need to define inputs(SparseTensor), kernel (of shape k^3 x OC x IC when k > 1, or OC x IC when k = 1, where k denotes the kernel size and IC, OC means input / output channels). The outputs is still a SparseTensor.

Detailed examples are given in here, where we use the torchsparse.nn.functional interfaces to implement weight-shared 3D-NAS modules.

Sparse Hashmap API

Sparse hash map query is important in 3D sparse computation. It is mainly used to infer a point's memory location (i.e. index) given its coordinates. For example, we use this operation in kernel map construction part of 3D sparse convolution, and also sparse voxelization / devoxelization in Sparse Point-Voxel Convolution. Here, we provide the following example for hash map API:

source_hash = torchsparse.nn.functional.sphash(torch.floor(source_coords).int())
target_hash = torchsparse.nn.functional.sphash(torch.floor(target_coords).int())
idx_query = torchsparse.nn.functional.sphashquery(source_hash, target_hash)

In this example, sphash is the function converting integer coordinates to hashing. The sphashquery(source_hash, target_hash) performs the hash table lookup. Here, the hash map has key target_hash and value corresponding to point indices in the target point cloud tensor. For each point in the source_coords, we find the point index in target_coords which has the same coordinate as it.

Dummy Training Example

We here provides an entire training example with dummy input here. In this example, we cover

How we start from point cloud data and convert it to SparseTensor format;
How we can implement SparseTensor batching;
How to train a semantic segmentation SparseConvNet.

You are also welcomed to check out our SPVNAS project to implement training / inference with real data.

Mixed Precision (float16) Support

Mixed precision training is supported via torch.cuda.amp.autocast and torch.cuda.amp.GradScaler. Enabling mixed precision training can speed up training and reduce GPU memory usage. By wrapping your training code in a torch.cuda.amp.autocast block, feature tensors will automatically be converted to float16 if possible. See here for a complete example.

Speed Comparison Between torchsparse and MinkowskiEngine

We benchmark the performance of our torchsparse and latest MinkowskiEngine V0.4.3 here, latency is measured on NVIDIA GTX 1080Ti GPU:

Network	Latency (ME V0.4.3)	Latency (torchsparse V1.0.0)
MinkUNet18C (MACs / 10)	224.7	124.3
MinkUNet18C (MACs / 4)	244.3	160.9
MinkUNet18C (MACs / 2.5)	269.6	214.3
MinkUNet18C	323.5	294.0

Citation

If you find this code useful, please consider citing:

@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},
  year = {2020}
}

Acknowledgements

This library is inspired by MinkowskiEngine, SECOND and SparseConvNet.

ootts/torchsparse