Multi-Crop on supervised loss

facebookresearch / swav

PyTorch implementation of SwAV https//arxiv.org/abs/2006.09882

Other

2.01k stars 280 forks source link

Multi-Crop on supervised loss #57

Closed zhangdan8962 closed 3 years ago

zhangdan8962 commented 3 years ago

Hi,

I am wondering how did you apply multi-crop on supervised setting?

Suppose we have 2x224 crops for each image, did you just calculate the cross entropy for each crop and then took the average as the loss for each sample?

mathildecaron31 commented 3 years ago

Hi @zhangdan8962 , yes this is exactly what I did ! In this case it is similar to "repeated augmentation" actually (https://arxiv.org/abs/1901.09335)