📣 News

• [02/Mar/2024] ✨ Release the code and checkpoint. 😊😊
• [26/Feb/2024] 🎉 Our paper DAPT is accepted by CVPR 2024! 🥳🥳

Abstract

Point cloud analysis has achieved outstanding performance by transferring point cloud pre-trained models. However, existing methods for model adaptation usually update all model parameters, i.e., full fine-tuning paradigm, which is inefficient as it relies on high computational costs (e.g., training GPU memory) and massive storage space. In this paper, we aim to study parameter-efficient transfer learning for point cloud analysis with an ideal trade-off between task performance and parameter efficiency. To achieve this goal, we first freeze the parameters of the default pre-trained models and then propose the Dynamic Adapter, which generates a dynamic scale for each point token, considering the token significance to the downstream task. We further seamlessly integrate Dynamic Adapter with Prompt Tuning (DAPT) by constructing Internal Prompts, capturing the instance-specific features for interaction. Extensive experiments conducted on five challenging datasets demonstrate that the proposed DAPT achieves superior performance compared to the full fine-tuning counterparts while significantly reducing the trainable parameters and training GPU memory by 95% and 35%, respectively.

Overview

Getting Started

Installation

We recommend using Anaconda for the installation process:

$ git clone https://github.com/LMD0311/DAPT.git
$ cd DAPT
# Create virtual env and install PyTorch
$ conda create -y -n dapt python=3.9
$ conda activate dapt
(dapt) $ pip install torch==1.8.0+cu111 torchvision==0.9.0+cu111 torchaudio==0.8.0 -f https://download.pytorch.org/whl/torch_stable.html

# Install basic required packages
(dapt) $ pip install -r requirements.txt

# Chamfer Distance & emd
(dapt) $ cd ./extensions/chamfer_dist && python setup.py install --user
(dapt) $ cd ../..
(dapt) $ cd ./extensions/emd && python setup.py install --user

# PointNet++
(dapt) $ pip install "git+https://github.com/erikwijmans/Pointnet2_PyTorch.git#egg=pointnet2_ops&subdirectory=pointnet2_ops_lib"

# GPU kNN
(dapt) $ pip install --upgrade https://github.com/unlimblue/KNN_CUDA/releases/download/0.2/KNN_CUDA-0.2-py3-none-any.whl

Datasets

See DATASET.md for details.

Pretrain

To fine-tune on downstream tasks, you may need to download or reproduce the pre-trained checkpoint.

Main Results (Point-MAE)

The evaluation commands with checkpoints should be in the following format:

CUDA_VISIBLE_DEVICES=<GPU> python main.py --test --config <yaml_file_name> --exp_name <output_file_name> --ckpts <path/to/ckpt>

Fine-tuning on downstream tasks

ModelNet40

CUDA_VISIBLE_DEVICES=<GPU> python main.py --config cfgs/finetune_modelnet_dapt.yaml --ckpts <path/to/pre-trained/model> --finetune_model --exp_name <name>

# further enable voting mechanism
CUDA_VISIBLE_DEVICES=<GPU> python main.py --config cfgs/finetune_modelnet_dapt.yaml --test --vote --exp_name <name> --ckpts <path/to/best/model>

The voting strategy is time-consuming and unfair for various compute platforms; hence, we prioritize reporting overall accuracy without voting. We recommend not to use the voting strategy, despite the promising results. 🙏🙏

ScanObjectNN

# For fine-tuning on OBJ-BG variant
CUDA_VISIBLE_DEVICES=<GPU> python main.py --config cfgs/finetune_scan_objbg_dapt.yaml --ckpts <path/to/pre-trained/model> --finetune_model --exp_name <name>

# For fine-tuning on OBJ-ONLY variant
CUDA_VISIBLE_DEVICES=<GPU> python main.py --config cfgs/finetune_scan_objonly_dapt.yaml --ckpts <path/to/pre-trained/model> --finetune_model --exp_name <name>

# For fine-tuning on PB-T50-RS variant
CUDA_VISIBLE_DEVICES=<GPU> python main.py --config cfgs/finetune_scan_hardest_dapt.yaml --ckpts <path/to/pre-trained/model> --finetune_model --exp_name <name>

t-SNE visualization

You can use t-SNE to visualize the results obtained on ScanObjectNN test sets.

# t-SNE on ScanObjectNN
CUDA_VISIBLE_DEVICES=<GPU> python main.py --config cfgs/tsne/finetune_scan_hardest_dapt_tsne.yaml --ckpts <path/to/ckpt> --tsne --exp_name <name>

You can also make your own config for other visualization. 😍😍

To Do

• Release code for classification.
• Release checkpoints.
• Release code for part segmentation.

Acknowledgements

This project is based on Point-BERT (paper, code), Point-MAE (paper, code), ACT(paper, code), ReCon (paper, code), IDPT (paper, code). Thanks for their wonderful works.

Citation

If you find this repository useful in your research, please consider giving a star ⭐ and a citation

@inproceedings{zhou2024dynamic,
  title={Dynamic Adapter Meets Prompt Tuning: Parameter-Efficient Transfer Learning for Point Cloud Analysis},
  author={Xin Zhou and Dingkang Liang and Wei Xu and Xingkui Zhu and Yihan Xu and Zhikang Zou and Xiang Bai},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2024}
}