Questions on Feature and Label Dimensions in RnCLoss

[teinhonglo]

Hi,

Thank you for your work on the RnCLoss implementation! I have a question regarding the feature and label dimensions in the provided example:

# features: [bs, 2, feat_dim]
features = ...
# labels: [2bs, label_dim]
labels = ...

Specifically, I’d like to clarify two points:

1. What does the 2 in the feature dimensions [bs, 2, feat_dim] represent? Does it indicate a pair of features for comparison (e.g., two embeddings per sample)?
2. For the label dimensions [2bs, label_dim], what value should be set for label_dim? Should it be 1, or does it depend on a specific aspect of the task?

A brief explanation would greatly help me understand how to properly prepare the input features and labels.

Thank you in advance for your time and assistance!

Best regards,
Tien-Hong Lo