x_dgcnn

A pytorch implementation of DGCNN, more efficient and memory-saving than dgcnn.pytorch.

The performance in comparison with dgcnn.pytorch could be found in the comparison_with_dgcnn.pytorch branch.

Update2: use keops to accelerate the computation of the pairwise distance matrix. It is memory-efficient, thus can handle larger point clouds.

Update: use GELU rather than LeakyReLU. Rewrite the message passing part to make it more efficient. Move the normalization and activation after the max pooling in all instances. Make the categorical embedding learnable. We try LayerNorm and InstanceNorm, and find that BatchNorm beats them by a large margin.

Installation

pip install x-dgcnn

If you suffer from the error No matching distribution found for x-dgcnn using a mirror source, try the following (if you know how to solve this problem, please open an issue):

pip install x-dgcnn -i https://pypi.org/simple

Usage

Classification.

import torch
from x_dgcnn import DGCNN_Cls

model = DGCNN_Cls(3, 10, 20)
x = torch.randn(8, 3, 2048)
xyz = x.clone()
out = model(x, xyz)

Semantic segmentation.

import torch
from x_dgcnn import DGCNN_Seg

# model = DGCNN_Seg(3, 10, 40, global_pooling=False)    # disable global pooling if batch size is too small
model = DGCNN_Seg(3, 10, 40)
x = torch.randn(8, 3, 2048)
xyz = x.clone()
out = model(x, xyz)

Part segmentation.

import torch
from x_dgcnn import DGCNN_Seg, SpatialTransformNet

# use stn and do part segmentation
stn = SpatialTransformNet(k=40)
model = DGCNN_Seg(k=40, in_dim=9, out_dim=4, n_category=10, stn=stn)
x = torch.randn(8, 9, 2048)  # keep xyz at the first 3 channels if using stn
xyz = x[:, :3].clone()
category = torch.randint(0, 10, (10,))
out = model(x, xyz, category)

You can disable keops by calling disable_keops() when exporting the model to onnx.

from x_dgcnn import disable_keops
disable_keops()

TODO

• Add differentiable subset operator and coordinate descent as another option to fuse features.
• Scale up model size and test.
• Add sampling to get hierarchical features.
• Do tasks on other tasks, e.x. rotation prediction, normal prediction, etc.

References

@article{wang2019dynamic,
  title={Dynamic graph cnn for learning on point clouds},
  author={Wang, Yue and Sun, Yongbin and Liu, Ziwei and Sarma, Sanjay E and Bronstein, Michael M and Solomon, Justin M},
  journal={Acm Transactions On Graphics (tog)},
  year={2019},
}