
摘要
本文介绍了SimCLR:一种用于对比学习视觉表示的简单框架。我们简化了最近提出的对比自监督学习算法,无需专门的架构或内存库。为了理解是什么使得对比预测任务能够学习到有用的表示,我们系统地研究了该框架的主要组成部分。研究表明:(1)数据增强组合在定义有效的预测任务中起着关键作用;(2)在表示和对比损失之间引入可学习的非线性变换显著提高了所学表示的质量;(3)与有监督学习相比,对比学习从更大的批量和更多的训练步骤中获益更多。通过结合这些发现,我们在ImageNet上大幅超越了先前的自监督和半监督学习方法。基于SimCLR学到的自监督表示训练的线性分类器达到了76.5%的Top-1准确率,比之前的最先进方法相对提高了7%,与有监督ResNet-50的表现相当。当仅使用1%的标签进行微调时,我们实现了85.8%的Top-5准确率,超过了AlexNet,而使用的标签数量仅为AlexNet的1/100。
代码仓库
hannaiiyanggit/unicon
pytorch
GitHub 中提及
sidwa/ae_thesis
pytorch
GitHub 中提及
alexk1704/scclv2
tf
GitHub 中提及
JoonHyung-Park/SimCLR
pytorch
GitHub 中提及
acholston/SimCLR
pytorch
GitHub 中提及
delyan-boychev/grad-cache-con-learning
pytorch
GitHub 中提及
jramapuram/SimCLR
pytorch
GitHub 中提及
giakou4/pyssl
pytorch
dewenzeng/clsp
pytorch
GitHub 中提及
Mayurji/SimCLR
pytorch
Pangoraw/GeoSSL
pytorch
wvangansbeke/Unsupervised-Classification
pytorch
GitHub 中提及
AidenDurrant/MoCo-Pytorch
pytorch
GitHub 中提及
sylviesss/pytorch-simclr
pytorch
GitHub 中提及
rajatkoner08/oodformer
pytorch
GitHub 中提及
google-research/simclr
官方
tf
GitHub 中提及
cjrd/self-supervised-pretraining
pytorch
GitHub 中提及
phanav/simclr-presentation
GitHub 中提及
XG293/SupConLoss
pytorch
GitHub 中提及
eminorhan/baby-vision
pytorch
GitHub 中提及
roydenwa/cell-centroid-former
tf
GitHub 中提及
bjader/QSSL
pytorch
GitHub 中提及
lightly-ai/lightly
pytorch
GitHub 中提及
mariaauslander/capstone_fall20_irrigation
tf
GitHub 中提及
mdiephuis/simclr
pytorch
GitHub 中提及
beresandras/contrastive-classification-keras
tf
GitHub 中提及
sayakpaul/SimCLR-in-TensorFlow-2
tf
GitHub 中提及
filipbasara0/matryoshka-representation-learning
pytorch
GitHub 中提及
radiradev/contrastive-neutrino
pytorch
GitHub 中提及
HobbitLong/SupContrast
pytorch
GitHub 中提及
Westlake-AI/openmixup
pytorch
GitHub 中提及
clabrugere/pytorch-scarf
pytorch
GitHub 中提及
ushmabhatt/Contrastive-Learning
tf
GitHub 中提及
SaeedShurrab/SimSiam-pytorch
pytorch
GitHub 中提及
vturrisi/solo-learn
pytorch
GitHub 中提及
guerbet-ai/wsp-contrastive
pytorch
GitHub 中提及
wilile26811249/SimCLR
pytorch
GitHub 中提及
sthalles/SimCLR
pytorch
ut-vision/simhand
pytorch
GitHub 中提及
davidczy/supcon_gamma
pytorch
GitHub 中提及
nzw0301/simclr
pytorch
zpf0117b/CLMB
pytorch
GitHub 中提及
edureisMD/ConVIRT-pytorch
pytorch
GitHub 中提及
talipucar/PyFlow_SimCLR
pytorch
GitHub 中提及
liuch37/image-processing
pytorch
GitHub 中提及
BalajiAI/SimCLR
pytorch
GitHub 中提及
Daniel-H-99/SimCLR
pytorch
GitHub 中提及
leftthomas/SimCLR
pytorch
GitHub 中提及
facebookresearch/vissl
pytorch
GitHub 中提及
rmastand/jetclr_ad
pytorch
GitHub 中提及
forcesh/SupContrast
pytorch
GitHub 中提及
anonymousmlsubmission/datadistributiontransfer
pytorch
GitHub 中提及
AndrewAtanov/simclr-pytorch
pytorch
GitHub 中提及
Mind23-2/MindCode-76
mindspore
GitHub 中提及
alibaba/EasyCV
pytorch
serre-lab/prj_selfsup
tf
GitHub 中提及
caesarea38/doclangid
pytorch
GitHub 中提及
ninatu/mood_challenge
pytorch
GitHub 中提及
cyrilvallez/image-manipulation-detection
pytorch
GitHub 中提及
p3i0t/simclr-cifar10
pytorch
GitHub 中提及
63days/SimCLR
pytorch
GitHub 中提及
yoojungsun0/Psych239
pytorch
GitHub 中提及
salarim/Semi-Supervised-CL
pytorch
GitHub 中提及
edreismd/convirt-pytorch
pytorch
GitHub 中提及
jwcalder/GraphLearningLayer
pytorch
GitHub 中提及
hemakaraveenhansika/convirt
pytorch
GitHub 中提及
ta9ryuWalrus/simclr
tf
GitHub 中提及
john-mlr/CIFAR-10-SimCLR
pytorch
GitHub 中提及
parkinkon1/simclr
tf
GitHub 中提及
spijkervet/simclr
pytorch
GitHub 中提及
chingyaoc/DCL
pytorch
GitHub 中提及
htdt/self-supervised
pytorch
GitHub 中提及
sunjesse/SimCLR_pytorch
pytorch
GitHub 中提及
optimization-ai/sogclr
pytorch
GitHub 中提及
thunderInfy/simclr
pytorch
GitHub 中提及
kawshik8/DL-project
pytorch
GitHub 中提及
krafton-ai/mini-batch-cl
pytorch
GitHub 中提及
ae-foster/pytorch-simclr
pytorch
GitHub 中提及
rdbch/tutorial_simclr
pytorch
GitHub 中提及
ssumin6/SimCLR
pytorch
GitHub 中提及
rdbch/simclr
pytorch
GitHub 中提及
Liut2016/ecg-supcontrast
pytorch
GitHub 中提及
rahimentezari/datadistributiontransferlearning
pytorch
GitHub 中提及
applecrumble123/SimCLR_pytorch
pytorch
GitHub 中提及
asd08573064/SimCLR
tf
GitHub 中提及
open-mmlab/mmselfsup
pytorch
GitHub 中提及
danielzgsilva/MOT
pytorch
GitHub 中提及
mandiehyewon/goodviews_ecg
pytorch
GitHub 中提及
reppy4620/SimCLR4Paint
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| contrastive-learning-on-imagenet-1k | ResNet50 | ImageNet Top-1 Accuracy: 69.3 |
| image-classification-on-places205 | SimCLR | Top 1 Accuracy: 53.3 |
| object-recognition-on-shape-bias | SimCLR (ResNet-50x4) | shape bias: 40.7 |
| object-recognition-on-shape-bias | SimCLR (ResNet-50x1) | shape bias: 38.9 |
| object-recognition-on-shape-bias | SimCLR (ResNet-50x2) | shape bias: 41.7 |
| person-re-identification-on-sysu-30k | SimCLR (self-supervised) | Rank-1: 10.9 |
| self-supervised-image-classification-on | SimCLR (ResNet-50) | Number of Params: 24M Top 1 Accuracy: 69.3% Top 5 Accuracy: 89.0% |
| self-supervised-image-classification-on | SimCLR (ResNet-50 4x) | Number of Params: 375M Top 1 Accuracy: 76.5% Top 5 Accuracy: 93.2% |
| self-supervised-image-classification-on | SimCLR (ResNet-50 2x) | Number of Params: 94M Top 1 Accuracy: 74.2% Top 5 Accuracy: 92.0% |
| self-supervised-image-classification-on-1 | SimCLR (Resnet-50) | Top 1 Accuracy: 77.2% |
| self-supervised-person-re-identification-on | SimCLR | Rank-1: 10.9 |
| semi-supervised-image-classification-on-1 | SimCLR (ResNet-50 4×) | Top 1 Accuracy: 63.0% Top 5 Accuracy: 85.8% |
| semi-supervised-image-classification-on-1 | SimCLR (ResNet-50 2×) | Top 1 Accuracy: 58.5% Top 5 Accuracy: 83.0% |
| semi-supervised-image-classification-on-1 | SimCLR (ResNet-50) | Top 1 Accuracy: 48.3% Top 5 Accuracy: 75.5% |
| semi-supervised-image-classification-on-2 | SimCLR (ResNet-50) | Top 5 Accuracy: 87.8% |
| semi-supervised-image-classification-on-2 | SimCLR (ResNet-50 4×) | Top 5 Accuracy: 92.6% |
| semi-supervised-image-classification-on-2 | SimCLR (ResNet-50 2×) | Top 5 Accuracy: 91.2% |