HSIC输入参数的大小是什么？

[pending1face]

您好，我看到论文里的描述，HSIC的输入似乎应该是batch_size * assign_num * embedding_dim的大小。但是在代码里，由于diffpool_net里使用了torch.sum(h, dim=1)的readout，输入变成了batch_size * embedding_dim。
我打印了nets.superpixels_graph_classification.diffpool.py 里HSIC_weight模块中的feature1。

 def forward(self, lambdap, global_epoch, lambda_decay_rate, lambda_decay_epoch, min_lambda_times, first):
     if True:
         self.all_weights = self.weights
     else:
         self.all_weights = torch.cat((self.weights, self.pre_weight.detach()), dim=0)
     lossb = Variable(torch.FloatTensor([0]).cuda())
     for i in range(self.assign_num-1):
         for j in range(i+1, self.assign_num):
             feature1 = self.all_features[:, i*self.embedding_size : (i+1)*self.embedding_size]
             feature2 = self.all_features[:, j*self.embedding_size : (j+1)*self.embedding_size]
            
            print(feature1)
            #lossb += self.loss_dependence(feature1, feature2, softmax(self.all_weights), self.all_features.size(0)).view(1)
            lossb += self.biased_estimator(feature1, feature2, softmax(self.all_weights)).view(1)
    #lossq = (torch.sum(self.weights*self.weights)-self.n)**2 
    lossp = softmax(self.weights).pow(2).sum()
    lambdap = lambdap * max((lambda_decay_rate ** (global_epoch // lambda_decay_epoch)),
                                 min_lambda_times)
    lossg = 1e+3*lossb / lambdap + lossp
    return lossg, lossb

发现它大小很奇怪。

可以解答一下吗？