Why is it necessary to transpose in indices = linear_sum_assignment(C.transpose(0, 1))? This transposition inverses the rows and columns, causing indices[1] not to correspond to cur_embds.

[KaihongLi]

def match_embds(self, ref_embds, cur_embds):
    #  embeds (q, b, c)
    ref_embds, cur_embds = ref_embds.detach()[:, 0, :], cur_embds.detach()[:, 0, :]
    ref_embds = ref_embds / (ref_embds.norm(dim=1)[:, None] + 1e-6)
    cur_embds = cur_embds / (cur_embds.norm(dim=1)[:, None] + 1e-6)
    cos_sim = torch.mm(ref_embds, cur_embds.transpose(0, 1))
    C = 1 - cos_sim

    C = C.cpu()
    C = torch.where(torch.isnan(C), torch.full_like(C, 0), C)

    indices = linear_sum_assignment(C.transpose(0, 1))
    indices = indices[1]
    return indices