search

dvlab-research / Parametric-Contrastive-Learning

Parametric Contrastive Learning (ICCV2021) & GPaCo (TPAMI 2023)
https://arxiv.org/abs/2107.12028
MIT License
240 stars 32 forks source link

The difference between GPaCo and PaCo #18

Open madoka109 opened 1 year ago

madoka109 commented 1 year ago

Hello! Thanks for your excellent work! I would like to know what is the difference between GPaCo and PaCo?

jiequancui commented 1 year ago

Hi,

Thanks for your interests in our work.

We systematically examine the necessary of each component (augmentation strategy, formulation of PaCo, the momentum encoder and queue size) in PaCo and simplify Parametric Contrastive Learning (PaCo) to Generalized Contrastive Learning (GPaCo) by removing the momentum encoder. GPaCo outperforms PaCo by a large margin. We verify the generality of GPaCo on various tasks, including long-tailed recognition, recognition on full ImageNet and CIFAR-100 datasets across CNNs to vision transformers, robustness on out-of-distribution data, and semantic segmentation.

Best,