Implemention of Dimensionality reduction by learning an invariant mapping in CVPR2006
- In the abstract of MOCO, The DrLIM was introduced, which utilizes the perspective of contrastive learning.
- MNIST
- train: 3000 + a (Augmentation of -6, -3, and 3 degree rotation plus 6 pixel dropout.) of each class
- test: 400 of each class
- training set
- ex) img1, img2, label. img1, img2 is random shuffle of "4" and "9". and label is write under the description.
- Using class of "4" and "9"
- label 0: same class image
- label 1: different class image