This is the code for CVPR 2021 oral paper: Jigsaw Clustering for Unsupervised Visual Representation Learning

Last update: Sep 18, 2022

Related tags

Deep Learning JigsawClustering

Overview

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},
}

Comments

Some question about trainning

Hi~Thanks for your excellent work! I have a machine with 2 1080Ti,and I want to train your model on CIFAR10 with resnet18.

I use the parmeters like this ,but it seems don't work.

The program is stuck in this situation.

opened by zbw0329 10
Some details about the training

Hi, I have recently read your paper and find it very interesting. There are still some confusions about the experiments.

The experiments require 4 2080ti for training. Does it mean we must have 4 2080ti on one single machine? What if I have 4 2080ti on different machines? Is there any suggestion for this situation? BTW, how long does it take when you train on ImageNet1k?

Much appreciation for your reply.

Best wishes!

opened by Hanzy1996 3
Some questions about the results of ImageNet100

Thank you for your wonderful work, I want to do some more works based on your code. But I meet some questions about the results. I use the JigsawClustering and the dataset ImageNet100 to train the model. I only changed one line in the model to fit this dataset(I added model.fc = nn.Linear(2048, 100) in line 162 of main_lincls.py). However, despite using 4 GPUs, and did not change the configuration file. I only got an accuracy of 79.24. There is still a certain gap between this and the 80.9 reported in the paper. How can I achieve the accuracy reported in the paper now? Once again, thank you for your excellent work and code. I am looking forward to your reply.

opened by WilyZhao8 1
Results of Faster-RCNN R50-FPN with model pretrained on ImageNet with standard cross-entropy loss

Hi, thanks for your work! In Objection Detection, do you apply ResNet-50 model pretrained on ImageNet with standard cross-entropy loss to Faster-RCNN R50-FPN?

opened by fzfs 1
Training the model on a single GPU

Hi! I'm aware that the question has been asked previously, but could you guide how to modify jigclu to remove the distributeddataparallel depedency?

Thanks!

opened by shuvam-creditmate 2
It seems that the model has not learned anything，What should I do？

Thanks for your excellent work! I change the dataloader to use JigClu in CIFAR-10,and train the model on it by 1000epoch. But the prediction of my model is all the same. It seem that model always cluster into the same cluster

opened by zbw0329 10

Releases(1.0)

1.0(May 11, 2021)

Release pretrained models.
Source code(tar.gz)
Source code(zip)
JigClu_200e.pth(213.72 MB)

Owner

DV Lab

Deep Vision Lab

GitHub Repository

Price-Prediction-For-a-Dream-Home - A machine learning based linear regression trained model for house price prediction.

Price-Prediction-For-a-Dream-Home ROADMAP TO THIS LINEAR REGRESSION BASED HOUSE PRICE PREDICTION PREDICTION MODEL Import all the dependencies of the p

1 Dec 29, 2021

DEMix Layers for Modular Language Modeling

DEMix This repository contains modeling utilities for "DEMix Layers: Disentangling Domains for Modular Language Modeling" (Gururangan et. al, 2021). T

43 Nov 11, 2022

A Planar RGB-D SLAM which utilizes Manhattan World structure to provide optimal camera pose trajectory while also providing a sparse reconstruction containing points, lines and planes, and a dense surfel-based reconstruction.

ManhattanSLAM Authors: Raza Yunus, Yanyan Li and Federico Tombari ManhattanSLAM is a real-time SLAM library for RGB-D cameras that computes the camera

117 Dec 28, 2022

This is the code for CVPR 2021 oral paper: Jigsaw Clustering for Unsupervised Visual Representation Learning

Related tags

Overview

JigsawClustering

Introduction

Installation

Environment

Install

Dataset

Results and pretrained model

Training

Pre-training on ImageNet

Linear evaluation on ImageNet

Semi-Supervised finetune on ImageNet

Transfer to COCO detection

Citation

Comments

Some question about trainning

Some details about the training

Some questions about the results of ImageNet100

Results of Faster-RCNN R50-FPN with model pretrained on ImageNet with standard cross-entropy loss

Training the model on a single GPU

It seems that the model has not learned anything，What should I do？

Releases(1.0)

1.0(May 11, 2021)

Owner

DV Lab

Price-Prediction-For-a-Dream-Home - A machine learning based linear regression trained model for house price prediction.

DEMix Layers for Modular Language Modeling

A Planar RGB-D SLAM which utilizes Manhattan World structure to provide optimal camera pose trajectory while also providing a sparse reconstruction containing points, lines and planes, and a dense surfel-based reconstruction.

A fast Evolution Strategy implementation in Python

CM-NAS: Cross-Modality Neural Architecture Search for Visible-Infrared Person Re-Identification (ICCV2021)

Auxiliary data to the CHIIR paper Searching to Learn with Instructional Scaffolding

CIFAR-10 Photo Classification

BraTs-VNet - BraTS(Brain Tumour Segmentation) using V-Net

EgoNN: Egocentric Neural Network for Point Cloud Based 6DoF Relocalization at the City Scale

CoCosNet v2: Full-Resolution Correspondence Learning for Image Translation

Pytorch implementation of paper "Efficient Nearest Neighbor Language Models" (EMNLP 2021)

Lane assist for ETS2, built with the ultra-fast-lane-detection model.

Weighted K Nearest Neighbors (kNN) algorithm implemented on python from scratch.

A small library of 3D related utilities used in my research.

[SDM 2022] Towards Similarity-Aware Time-Series Classification

Official implementation of "Open-set Label Noise Can Improve Robustness Against Inherent Label Noise" (NeurIPS 2021)

Tensorflow implementation of Human-Level Control through Deep Reinforcement Learning

Populating 3D Scenes by Learning Human-Scene Interaction https://posa.is.tue.mpg.de/

A state-of-the-art semi-supervised method for image recognition

Lux AI environment interface for RLlib multi-agents