JigsawClustering

Jigsaw Clustering for Unsupervised Visual Representation Learning

Pengguang Chen, Shu Liu, Jiaya Jia

Introduction

This project provides an implementation for the CVPR 2021 paper "Jigsaw Clustering for Unsupervised Visual Representation Learning"

Installation

Environment

We verify our code on

• 4x2080Ti GPUs
• CUDA 10.1
• python 3.7
• torch 1.6.0
• torchvision 0.7.0

Other similar envirouments should also work properly.

Install

We use the SyncBN from apex, please install apex refer to https://github.com/NVIDIA/apex (SyncBN from pytorch should also work properly, we will verify it later.)

We use detectron2 for the training of detection tasks. If you are willing to finetune our pretrained model on the detection task, please install detectron2 refer to https://github.com/facebookresearch/detectron2

git clone https://github.com/Jia-Research-Lab/JigsawClustering.git
cd JigsawClustering/
pip install diffdist

Dataset

Please put the data under ./datasets. The directory looks like:

datasets
│
│───ImageNet/
│   │───class1/
│   │───class2/
│   │   ...
│   └───class1000/
│   
│───coco/
│   │───annotations/
│   │───train2017/
│   └───val2017/
│
│───VOC2012/
│   
└───VOC2007/

Results and pretrained model

The pretrained model is available at here.

Task	Dataset	Results
Linear Evaluation	ImageNet	66.4
Semi-Supervised 1%	ImageNet	40.7
Semi-Supervised 10%	ImageNet	63.0
Detection	COCO	39.3

Training

Pre-training on ImageNet

python main.py --dist-url 'tcp://localhost:10107' --multiprocessing-distributed --world-size 1 --rank 0 \
    -a resnet50 \
    --lr 0.03 --batch-size 256 --epoch 200 \
    --save-dir outputs/jigclu_pretrain/ \
    --resume outputs/jigclu_pretrain/model_best.pth.tar \
    --loss-t 0.3 \
    --cross-ratio 0.3 \
    datasets/ImageNet/

Linear evaluation on ImageNet

python main_lincls.py --dist-url 'tcp://localhost:10007' --multiprocessing-distributed --world-size 1 --rank 0 \
    -a resnet50 \
    --lr 10.0 --batch-size 256 \
    --prefix module.encoder. \
    --pretrained outputs/jigclu_pretrain/model_best.pth.tar \
    --save-dir outputs/jigclu_linear/ \
    datasets/ImageNet/

Semi-Supervised finetune on ImageNet

10% label

python main_semi.py --dist-url 'tcp://localhost:10102' --multiprocessing-distributed --world-size 1 --rank 0 \
    -a resnet50 \
    --batch-size 256 \
    --wd 0.0 --lr 0.01 --lr-last-layer 0.2 \
    --syncbn \
    --prefix module.encoder. \
    --labels-perc 10 \
    --pretrained outputs/jigclu_pretrain/model_best.pth.tar \
    --save-dir outputs/jigclu_semi_10p/ \
    datasets/ImageNet/

1% label

python main_semi.py --dist-url 'tcp://localhost:10101' --multiprocessing-distributed --world-size 1 --rank 0 \
    -a resnet50 \
    --batch-size 256 \
    --wd 0.0 --lr 0.02 --lr-last-layer 5.0 \
    --syncbn \
    --prefix module.encoder. \
    --labels-perc 1 \
    --pretrained outputs/jigclu_pretrain/model_best.pth.tar \
    --save-dir outputs/jigclu_semi_1p/ \
    datasets/ImageNet/

Transfer to COCO detection

Please convert the pretrained weight first

python detection/convert.py

Then start training using

python detection/train_net.py --config-file detection/configs/R50-JigClu.yaml --num-gpus 4

VOC detection

python detection/train_net.py --config-file detection/configs/voc-R50-JigClu.yaml --num-gpus 4

Citation

Please consider citing JigsawClustering in your publications if it helps your research.

@inproceedings{chen2021jigclu,
    title={Jigsaw Clustering for Unsupervised Visual Representation Learning},
    author={Pengguang Chen, Shu Liu, and Jiaya Jia},
    booktitle={IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
    year={2021},
}