I see here that X-CLIP uses learnable weight matrices to compute the final scores from the similarity vectors/matrices. However, I am having trouble reconciling this with the equations in Section 3.3 of the paper, which seem to show just a softmax-weighted summation with no learnable parameters.
I'm not sure what I'm missing here. Would you mind clarifying this for me?
I see here that X-CLIP uses learnable weight matrices to compute the final scores from the similarity vectors/matrices. However, I am having trouble reconciling this with the equations in Section 3.3 of the paper, which seem to show just a softmax-weighted summation with no learnable parameters.
I'm not sure what I'm missing here. Would you mind clarifying this for me?