kaira.losses.multimodal.CMCLoss

Inheritance diagram for CMCLoss

class kaira.losses.multimodal.CMCLoss(lambda_cmc=1.0)

Bases: BaseLoss

Cross-Modal Consistency Loss Module.

This module implements a loss to ensure consistency across modalities.

Methods

__init__	Initialize the CMCLoss module.
forward	Forward pass through the CMCLoss module.

__init__(lambda_cmc=1.0)

Initialize the CMCLoss module.

Parameters:

lambda_cmc (float) – Weight for the CMC loss. Default is 1.0.

forward(x1: Tensor, x2: Tensor, proj1: BaseLoss, proj2: BaseLoss) → Tensor

Forward pass through the CMCLoss module.

Parameters:

• x1 (torch.Tensor) – Features from the first modality.

• x2 (torch.Tensor) – Features from the second modality.

• proj1 (BaseLoss) – Projection head for the first modality.

• proj2 (BaseLoss) – Projection head for the second modality.

Returns:

The cross-modal consistency loss.

Return type:

torch.Tensor