Question about the process of calculating top1, 3, 5 accuracy in nice_stand.py

[gyuseok0917]

with torch.no_grad():
    for i, (teeg, tlabel) in enumerate(self.test_dataloader):
        teeg = Variable(teeg.type(self.Tensor))
        tlabel = Variable(tlabel.type(self.LongTensor))
        all_center = Variable(all_center.type(self.Tensor))            

        tfea = self.Proj_eeg(self.Enc_eeg(teeg))
        tfea = tfea / tfea.norm(dim=1, keepdim=True)
        similarity = (100.0 * tfea @ all_center.t()).softmax(dim=-1)  # no use 100?
        _, indices = similarity.topk(5)

        tt_label = tlabel.view(-1, 1)
        total += tlabel.size(0)
        top1 += (tt_label == indices[:, :1]).sum().item()
        top3 += (tt_label == indices[:, :3]).sum().item()
        top5 += (tt_label == indices).sum().item()

Hello. I have a question while referencing your code. I saw the code for obtaining top1,3,5 accuracy in nice_stand.py, and I understand that t_label represents the actual integer label of the data. Then, after taking the inner product of eeg and image embedding, it is expressed as a probability distribution through softmax, and then the top k with the highest probability are selected and indices, which are the index information of these, are obtained. t_label is expressed as tt_label by converting the shape and then selecting k indices in order according to the highest probability according to k = 1, 3, 5 and then comparing them with tt_label. The problem here is that tt_label represents the actual integer label of the corresponding data, and indices represent the index information of the position with the highest probability, but the two values ​​are not comparable, so how can the top-1,3,5 accuracy be obtained?