This repository contains the implementation for the paper:

PointRTD: Replaced Token Denoising for Robust Point Cloud Pretraining

Installation

1. Clone this repository:

   git clone https://github.com/GunnerStone/PointRTD.git
   cd PointRTD
   

2. Installing Dependencies

conda env create -f environment.yaml

Below is a tested combination of library versions that are compatible. Start here if you are having trouble.

Environment Details

• Python: 3.10.4
• PyTorch: 2.4.1 (py3.10_cuda12.1_cudnn9.1.0_0 build)
• PyTorch-CUDA: 12.1
• PyTorch3D: 0.7.8 (py310_cu121_pyt241 build)
• Torch-Geometric: 2.6.1
• Torch-Cluster: 1.6.3 (+pt24cu121 build)
• Torch-Scatter: 2.1.2 (+pt24cu121 build)
• Torch-Sparse: 0.6.18 (+pt24cu121 build)
• Torch-Spline-Conv: 1.2.2 (+pt24cu121 build)

Notes

• Make sure to install the exact versions listed above to avoid compatibility issues.
• These libraries are designed to work with CUDA 12.1 and cuDNN 9.1.0, so ensure your system supports these versions.
• All other libraries should be fairly easy to pip/conda install.

Training

Pretraining PointRTD on ShapeNetCore.v2

Download the ShapeNetCore.v2 using the instructions found in the README

Run through the provided pretraining notebook:

pretrain_pointRTD.ipynb will produce pretraining checkpoints located in ./checkpoints/Pretrain_PointRTD/CR_XX/pointrtd_epoch_XX_CR_XX.pth

Once you have a satisfactory checkpoint, create a folder ./checkpoints_pointrtd/ and place your checkpoint file within this folder.

Fine-tuning on ModelNet10 or ModelNet40

Download ModelNet10 and ModelNet40 using instructions found in their respective READMEs.

Use the provided training notebook:

train_modelnet10_pointRTD.ipynb or train_modelnet40_pointRTD.ipynb

Make sure these are using the correct path for your desired pretrained model checkpoint file.