MoCo v2

motivation

  • SimCLR提出projection head & data augmentation能够大大提高表现
  • SimCLR训练过程使用过大batch size,需要TPU支持

contribution

  • 在MoCo v1中使用了projection head & data augmentation
  • 通过v1中的queue机制解决SimCLR过大batch size问题(仅concat mini_batch的低维特征向量)

method

  • SimCLR为end2end机制,需要较大batch size来提供negative set
  • MoCo通过queue保存negative key,每轮迭代仅encode mini_batch样本
  • 依旧使用InfoNCE

experiment

MLP head

  • 通过两层MLP(ReLU)代替v1中的fc
  • in contrast to the big leap on ImageNet, the detection gains are smaller

Augmentation

  • 引入blur augmentation(stronger color distortion在实验中没有较明显的效果提升)
  • 单纯引入aug+在检测任务中效果优于MLP,但在线性分类中效果较差
  • linear classification accuracy is not monotonically related to transfer performance in detection

Comparison with SimCLR

Computational cost

  • SimCLR需要更新q、k enconder,而MoCo仅更新q encoder,k encoder通过Momentum update进行更新

   Reprint policy


《MoCo v2》 by Liangyu Cui is licensed under a Creative Commons Attribution 4.0 International License
  TOC