search

paulhager / MMCL-Tabular-Imaging

82 stars 13 forks source link

LAAF questions #16

Closed EvelyneCalista closed 4 months ago

EvelyneCalista commented 4 months ago

Hi Hager,

Thanks for your great work. I would like to ask for training the Laaf:

  1. do the corruption augmentation and one hot encoder also applied for the label?
  2. Do the loss function for Laaf is using "supcon_loss_custom.py" this py file? I am getting confused with the other two py files "supcon_loss_clip_binary.py" and "supcon_loss_clip.py"

Thank you.

paulhager commented 4 months ago

Hi Evelyne,

  1. Yes, the label is treated as a normal feature that can be corrupted. I did not experiment with leaving it out of the corruption strategy, this might be worth taking a look at!
  2. You can use either the supcon loss (yes, its supcon_loss_custom, sorry for the confusion) or a normal NTXent loss, both work. In the paper we showed that the effect is often additive, so supcon loss + LaaF should return best results.