WesleyHsieh0806/SS-PRL

SS-PRL: Self-Supervised Pyramid Representation Learning for Multi-Label Visual Analysis (IEEE WACV 2023)

Self-Supervised Pyramid Representation Learning
for Multi-Label Visual Analysis and Beyond

This repository provides the official Pytorch implementation of pretraining and downstream evaluations for SS-PRL.

📎 Paper Link ✏️ Citations

• Learning of Patch-Based Pyramid Representation
• Cross-Scale Patch-Level Correlation Learning with Correlation Learners

---

Table of Contents

• Prepare Dataset
• Usage
• Downstream tasks
• Citations

---

📚 Prepare Dataset

Please refer to Pretrained_Dataset and Downstream Tasks for further details.

Tasks	Datasets:point_down:
Pre-Training	ImageNet COCO
Downstream	Pascal VOC COCO

🏃 Usage - Training

Requirements

• Python 3.6
• PyTorch = 1.4.0
• torchvision = 0.5.0
• CUDA 10.1 (Check with nvcc --version)
• Apex 0.1 (Installation)
• scipy, pandas, numpy

conda install pytorch==1.4.0 torchvision==0.5.0 cudatoolkit=10.1 -c pytorch

git clone "https://github.com/NVIDIA/apex"
cd apex
git checkout 4a1aa97e31ca87514e17c3cd3bbc03f4204579d0
python setup.py install --cuda_ext

Training with the shell script.

For further details, take a look at the source file | dataset definition | utilities

# Training Checklist:
# 1. modify the DATASET_PATH and EXPERIMENT_PATH in the script
# 2. BATCH_PER_GPU denotes the batch size per gpu, while --nproc_per_node denotes the number of gpus
# 3. modify the parameters
cd SS-PRL
bash train_SSPRL.sh

🚴 Downstream tasks

1. Download the pretrained models

   We provide the checkpoint files of SS-PRL and other SoTA used in our experiments, including

   • SwAV
   • MoCo
   • DenseCL
   • BYOL
   • InsLoc
   • MaskCo

   # Download the checkpoints with this command
   bash get_premodels.sh

2. Transferring to Multi-Label Visual Analysis tasks:

   Please Refer to Readme files for Classification, Object-Detection, and Semantic Segmentation tasks.

Citations

@misc{hsieh2022selfsupervised,
    title={Self-Supervised Pyramid Representation Learning for Multi-Label Visual Analysis and Beyond},
    author={Cheng-Yen Hsieh and Chih-Jung Chang and Fu-En Yang and Yu-Chiang Frank Wang},
    year={2022},
    eprint={2208.14439},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

Acknowledgement

Thanks the Facebook-Research-SwAV for its open-source project.