My confusion about your evaluation Logic

[ZuoZh3n]

Hi, author
I read through how you train a few shot learning model and how you eval it. I found something that I cannot understand.

In training phase, we got a Wmap to map the samples_embeding to labels_embedding. Wmap make the samples_embedding simliar to labels_embedding correct ?

Then In the evaluation phrase, we want to calculate the similarity score between test samples and candidate labels, why should we mapping the labels_embedding via Wmap? I think we only need to transform the sentence_embedding on Wmap and then we directly compare with the label_embedding without Wmap transformation ?