search

facebookresearch / dino

PyTorch code for Vision Transformers training with the Self-Supervised learning method DINO
Apache License 2.0
6.35k stars 907 forks source link

Semi-supervised extension #153

Closed clane9 closed 2 years ago

clane9 commented 3 years ago

Thank you for releasing this code! DINO seems like it would naturally extend to the semi-supervised setting by adding a classification loss for any samples with labels. Have you tried something like this? Does it work? Thanks in advance!

mathildecaron31 commented 2 years ago

Hi @clane9

Thanks for your interest in this work. That is actually a good remark and we quickly tried some semi-supervised extension of the DINO loss. However, we found that different data augmentations were optimal for the supervised / SSL part of the loss and so the overall pipeline was a bit complicated.

On another side, I'd recommend our ICCV21 paper PAWS that extends SwAV to semi-supervised learning.