Skip to main content
. 2024 May 16;14(5):590. doi: 10.3390/biom14050590
Algorithm 1 Feature self augmentation process
  •   1:

    # I Input Image

  •   2:

    # T Input Text

  •   3:

    # F ← Image_Encoder()

  •   4:

    # T ← Feature_Extractor()

  •   5:

    # A ← Adaptor()

  •   6:

    # N ← Feature_Filter()

  •   7:

    pretrain_init(F)

  •   8:

    for each x in data_loader do

  •   9:

    # Asymmetry constraint

  • 10:

    # extract feature representations of different modes

  • 11:

        I_f = image_encoder(I)

  • 12:

        T_f = text_encoder(T)

  • 13:

    # Loss function

  • 14:

        loss_cl = cross_entropy_loss (I_f, T_f)

  • 15:

        loss_IRC = cross_entropy_loss (I_f, I_f) - β cross_entorpy_loss (T_f, T_f)

  • 16:

        loss = loss_(cl) +loss_(IRC)

  • 17:

        F ← F.detach()

  • 18:

        update(T, D)

  • 19:

    end for