在mini  batch中，我要返回一个（batch_size, dim）尺寸的张量，然后计算相同类别之间的平均内积，和异类之间的平均内积。

这里其实有一点可以考虑，咱们到底是将内积作为目标函数来优化好，还是将距离作为目标函数来优化好呢？

如果是内积的话，同类的好说，同类之间的内积越接近1越好，异类之间的内积，是接近0好，还是-1好呢？这确实是一个问题。

在我下面的代码中，我是将异类的内积设计为-1为目标值了，注意，送到Pull_Push_Features_loss 中的特征参数已经是经过二范数规范后的了。

class Pull_Push_Features_loss(nn.Module): #特征拉近推远损失 
    def __init__(self):
        super(Pull_Push_Features_loss,self).__init__()

    def forward(self,features, batch_size, person_num):#features是经过renorm(2,0,1e-5).mul(1e5)后的特征，尺寸为(batch_size,dim)，dim为特征的长度，person_num为每个label具有的图片数量，用这种损失函数，在训练的时候，送入网络的数据都是N*K形式的组合。
        ID_num = batch_size // person_num #计算一共有多少个label
        loss_all = 0

        pull_distance_all = 0#计算相同label之间的平均内积，目标为1最好，所以用了1-distance
        for id_index in range(ID_num):
            features_temp_list = features[id_index * person_num:(id_index + 1) * person_num ]
            loss_temp = 0
            distance = torch.mm(features_temp_list, features_temp_list.t())
            distance = 1 - distance
            pull_distance_all = pull_distance_all + distance.sum()
        pull_distance_avg = pull_distance_all / ( ID_num * ( person_num * (person_num-1) ) )


        dif_ID_features = features[0:1]
        for i in range(1,ID_num):
            dif_ID_features = torch.cat( (dif_ID_features, features[i*person_num:i*person_num + 1] ), 0 )
        mat_similary = torch.mm(dif_ID_features, dif_ID_features.t())
        push_similary_all = torch.sum(mat_similary) - ID_num
        push_similary_avg = push_similary_all / ( ID_num * ID_num - ID_num )

        loss = pull_distance_avg + push_similary_avg
        return loss

我想了想，其实也没那么多事，就是我用了余弦距离作为衡量标准而已