daodaofr/hypergraph_reid

Code for CVPR 2020 paper Learning Multi-Granular Hypergraphs for Video-Based Person Re-Identification

hypergraph_reid's Issues

Cannot get train data.

请问定义时间粒度的代码在哪，BatchedGraphSAGEDynamicRangeMean1中比较相关的只看到了 t=3，而不是论文中所说的(1,3,5)

RuntimeError: cublas runtime error

Hello, an error occured when I ran "sh run_hypergraphsage_part.sh". It seems to be caused by the version of cudatoolkit.
I'm using python3.6.13, torch0.4.0 and cuda9.0. Could you share your environment? Thank you!

请问这个模型是否没有用到mi_loss，代码里没有看到

论文里提到的互信息损失，似乎在开源代码里没有提到。

BatchedGraphSAGEDynamicRangeMean1类中的实现只有一种时间尺度？

作者您好！对您的工作十分感兴趣，但根据源码里ResNet50GRAPHPOOLPARTHyper和BatchedGraphSAGEDynamicRangeMean1这两个类的相关实现，似乎只使用了一个时间粒度（3）？谢谢！

ValueError: not enough values to uppack (expected 6, got 5)

Excuse me. When I tested the computer memory was not enough, I changed the loader mode of the test data set to the random mode, and the computer reported an error: ValueError: not enough values to unpack (expected 6, got 5).Looking foward for your answer.Thank you very much!

注意力超图特征聚合的代码

在代码中只看到了图的池化相关的代码，例如 f1 = F.avg_pool1d(f1, int(t*self.p1))，将节点级的特征汇总到每个超图的图级表示。
而没有找到论文中所提及的 注意力超图特征聚合，计算节点的注意力，麻烦看到回复一下，谢谢！

KeyError: 'state_dict'

loss converges to 4.4, map only reaches 60.0%, rank1 only reaches 73.3%, learning rate is set to 0.000075, and the step size is 100.

I use GeForce RTX 2080 with 7G memory, loss converges to 4.4, map only reaches 60.0%, rank1 only reaches 73.3%, batchsize is set to 8,learning rate is set to 0.000075, and the step size is 100.
Because there was not enough memory during the test, only a part of it was loaded. Do you think it is the parameter setting or too little test load data? I adjusted it several times, but the effect was not satisfactory.
Thank you ！

How to create "mars_query.txt" and "mars_gallery.txt"?

Hello, I want to save the results of ReID with the function "save_results". However, it seems that Mars folder doesn't include the two txt files needed. Could you tell me how to create them? Thank you!

code availability: is there any update?

any update on the code?

About trained models

Hi, thanks for your code first. The shared model is based on MARS. Can you offer the trained model on iLIDS-VID?

where is the resnet50-19c8e357.pth

when I want to train the model ,terminal shows that pretrained model resnet50-19c8e357.pth is no such file.
hope to get reply.
thanks!

评估每个超边的重要性的代码

对您的这篇论文非常感兴趣，想问您一下，您在定义layers1的时候 只用到了BatchedGraphSAGEDynamicRangeMean1 这个类，其他剩下的类您注释了BatchedGraphSAGEDynamicMean1，BatchedGraphSAGEMean1，BatchedGraphSAGEMean1Temporal，BatchedGAT_cat1 请问这几个类有用吗，还有想请问您一下 ，
for i in range(int(N/p)):
idx_start = max(0, i-t)
idx_end = min(i+t+1, int(N/p))
tmp_x = x[:,idx_startp:idx_endp,]
dis = NearestConvolution.cos_dis(tmp_x) 这个是计算特征之间的相似性
if i==0:
tk = min(dis.shape[2], self.kn)
#print(tk)
_, idx = torch.topk(dis, tk, dim=2) 是包含前K个近邻的邻居集
k_nearest = torch.stack([torch.stack([tmp_x[j, idx[j, i]] for i in range(p*(idx_end-idx_start))], dim=0) for j in range(b)], dim=0) #(b, xp, kn, d)
#print(k_nearest)
k_nearest_list.append(k_nearest[:,p(i-idx_start):p*(i-idx_start+1),])
k_nearest = torch.cat(k_nearest_list, dim=1) #(b,N, kn, d)
x_neib = k_nearest[:,:,1:,].contiguous() 我们将除节点 v i 以外的超边中的所有节点特征进行平均操作,作为超边的特征.
x_neib = x_neib.mean(dim=2)
h_k = torch.cat((self.W_x(x), self.W_neib(x_neib)), 2)

    h_k = F.normalize(h_k, dim=2, p=2)
    h_k = F.relu(h_k)
    #print(h_k.shape)
    if self.use_bn:
        #self.bn = nn.BatchNorm1d(h_k.size(1))
        h_k = self.bn(h_k.permute(0,2,1).contiguous())
        #print(h_k.shape)
        h_k = h_k.permute(0, 2, 1)
        #print(h_k.shape)

    return h_k

请问评估每个超边的重要性的代码 在哪里啊 谢谢您的回复

About the mutual information

Hi! As shown in table1, the mutual information module improves the performance well. But I can't find the implementation of this module in your code. Since the general mutual information estimators can only get the lower bound of the real mutual information, I have no idea how do you minimize the mutual information between different parts. Do you use the methods like CLUB to model the upper bound of the mutual information? Looking forward to your reply, thx!

Any trained weights/checkpoints provided?

Is there any chance that trained weights/checkpoints be provided?

请问，什么时候开源代码？十分感谢。

How to evaluate on other datasets and save results?

Hello, I want to check the performance of this method on my dataset. However, it only provides the way to train. May I ask how to evaluate with the trained model, and how to save the results?
Thank you!