GAP baseline about renet

[dahyun-kang]

Hi

Sorry for late, I hope you've already figured it out.

The GAP baseline is as follows:

    def gap_baseline(self, spt, qry):
        spt = spt.squeeze(0)

        # shifting channel activations by the channel mean
        spt = self.normalize_feature(spt)
        qry = self.normalize_feature(qry)

        if self.args.shot > 1:
            spt = spt.view(self.args.shot, self.args.way, *spt.shape[1:])
            spt = spt.mean(dim=0)

        spt = spt.unsqueeze(0).repeat(qry.shape[0], 1, 1)
        qry_pooled = qry.mean(dim=[-1, -2])

        # cosine similarity (l2 normalization involved)
        similarity_matrix = F.cosine_similarity(spt, qry.unsqueeze(1), dim=-1)

        if self.training:
            return similarity_matrix / self.args.temperature, self.fc(qry_pooled)
        else:
            return similarity_matrix / self.args.temperature

I hope it works. Have a good weekend! 🙂

Best, Dahyun.