MoCo
Momentum Contrast for Unsupervised Visual Representation Learning
Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, Ross Girshick · CVPR 2020
Build a consistent contrastive dictionary with a momentum encoder and cross-batch negative queue.
What you will implement
Query-only InfoNCE, parameter EMA and a circular dictionary, with 12 actual synthetic SGD steps; no image-training or paper benchmark reproduction.
- 85. MoCo Momentum Encoder and Negative Dictionary
Combine query-only contrastive gradients, a momentum key encoder and a cross-batch circular dictionary.
Three core functions and a supplied 12-step synthetic linear-encoder SGD lab; no image encoder, distributed BN or paper benchmark reproduction.