MSN

Morphing and Sampling Network
[paper]
This code is optimized for users who want to learn only one category competion.

Environmet

The MDS sampling in parallel is only used in original environment.
Please check the section of Compile for detail.

If you use FPS for sampling

• python 3.8.10
• cuda 11.3
• pytorch 1.12.1
• pytorch3d
• open3d 0.13.0

If you want to use MDS in parallel

You should use docker

make docker images

in MSN dir

docker build . -t cuda10-0

make docker container

docker run --name MSN --gpus all -v [path of MSN in host]:/work -it cuda10-0 /bin/bash

Compile

This code uses extention modules used in original code.
I can confirm availability of emd and expansion_penalty in python-3.8.10 cuda-11.3 and pytorch-1.12.1. But MDS was not able to use in that environment.

Dataset

PCN Dataset from [url] or [BaiduYun].(from [PointTr Github]) You can use other datasets if it has json file same style with PCN Dataset.

Usage

training

python train.py

evaluation

python eval.py

Settings for training and evaluation are in options.py.