Why does your model have only one reduction? Reductions between branches in the paper are not shared.

    fg_p1 = self.reduction(zg_p1).squeeze(dim=3).squeeze(dim=2)
    fg_p2 = self.reduction(zg_p2).squeeze(dim=3).squeeze(dim=2)
    fg_p3 = self.reduction(zg_p3).squeeze(dim=3).squeeze(dim=2)
    f0_p2 = self.reduction(z0_p2).squeeze(dim=3).squeeze(dim=2)
    f1_p2 = self.reduction(z1_p2).squeeze(dim=3).squeeze(dim=2)
    f0_p3 = self.reduction(z0_p3).squeeze(dim=3).squeeze(dim=2)
    f1_p3 = self.reduction(z1_p3).squeeze(dim=3).squeeze(dim=2)
    f2_p3 = self.reduction(z2_p3).squeeze(dim=3).squeeze(dim=2)

In network.py you using maxpooling in branches, not avgpooling in the paper, why ?

I believe you have incorrect link to Market1501 dataset

thanks !

triplwtloss is giving dimension error. Please help me to fix this error.

File "main.py", line 50, in train
loss = self.loss(outputs, labels)
File "/usr/local/lib/python3.7/dist-packages/torch/nn/modules/module.py", line 889, in _call_impl
result = self.forward(*input, **kwargs)
File "/content/STREID/loss.py", line 14, in forward
Triplet_Loss = [triplet_loss(output, labels) for output in outputs[1:4]]
File "/content/STREID/loss.py", line 14, in
Triplet_Loss = [triplet_loss(output, labels) for output in outputs[1:4]]
File "/usr/local/lib/python3.7/dist-packages/torch/nn/modules/module.py", line 889, in _call_impl
result = self.forward(*input, **kwargs)
File "/content/STREID/loss.py", line 57, in forward
dist = torch.pow(inputs, 2).sum(dim=1, keepdim=True).expand(n, n)
IndexError: Dimension out of range (expected to be in range of [-1, 0], but got 1)

In the paper, global features from each part for softmax loss are 2048d. In your code , you did reductions to 256d. In network.py, the bottle neck for classification is always 256, although same layers named with '2048'. Is that a mistake or did you do that for a reason?

main.train()
loss = self.loss(outputs, labels)
File "/toolscnn/env_pyt1.1.0_py3.6/lib/python3.6/site-packages/torch/nn/modules/module.py", line 493, in call
result = self.forward(*input, **kwargs)
File "/MGN/loss.py", line 18, in forward
CrossEntropy_Loss = sum(CrossEntropy_Loss) / len(CrossEntropy_Loss)
RuntimeError: CUDA error: device-side assert triggered

In the description of Figure 2, there is a sentence "Notice that the 1 × 1 convolutions for dimension reduction and fully connected layers for identity prediction in each branch DO NOT share weights with each other."

I think your code share the weights.

Will user warnings affect the final results?

I have trained and evaluated using your default setting, results below can be taken as a reference by others.

start evaluate
[With    Re-Ranking] mAP: 0.9293 rank1: 0.9483 rank3: 0.9676 rank5: 0.9736 rank10: 0.9795
[Without Re-Ranking] mAP: 0.8544 rank1: 0.9376 rank3: 0.9700 rank5: 0.9783 rank10: 0.9863

https://github.com/GNAYUOHZ/ReID-MGN/blob/f0251e9e6003ec6f2c3fbc8ce5741d21436c20cf/data.py#L81C9-L81C59

In "data.py", line 81, "return int(file_path.split('/')[-1].split('')[0])" is wrong. If you want to get the id, this line should be changed to "return int(file_path.split('/')[-1].split('')[2].split('\')[-1])".

Hello,thanks for your contributation!
Because I can't download from Google,can you updata your weight on BaiDuYun or send it to me by email?My email is [email protected].
Thank you!

How to test a single image of my own or in the dataset.

The performance on DukeMTMC-reID set is extremely lower than you report.

[With Re-Ranking] mAP: 0.3625 rank1: 0.4654 rank3: 0.5274 rank5: 0.5597 rank10: 0.6154
[Without Re-Ranking] mAP: 0.2323 rank1: 0.3968 rank3: 0.5004 rank5: 0.5498 rank10: 0.6185

Above is the results on Dukeset. I tested it with weights you shared.
However, the performance on MARKET1501 is reproducible with the pre-trained weights.
So I suspect that this is caused by the absence of appropriate weights.
Anyway, could you share the weight file that is trained for DukeMTMC-reID?

Hi! Thinks for your contribution.I find that I have a very terrible result,map is 0.0048 after 350 epochs,could you give me some advice?

Hi when data set is changes Im getting below error. would you please help me to fix ths.

(base) mca@Rnd-Comp-05:~/Downloads/content/ReID-MGN-master$ python main.py --mode train --data_path /home/mca/Downloads/LTCC_ReID

epoch 1
/home/mca/anaconda3/lib/python3.7/site-packages/torch/optim/lr_scheduler.py:136: UserWarning: Detected call of lr_scheduler.step() before optimizer.step(). In PyTorch 1.1.0 and later, you should call them in the opposite order: optimizer.step() before lr_scheduler.step(). Failure to do this will result in PyTorch skipping the first value of the learning rate schedule. See more details at https://pytorch.org/docs/stable/optim.html#how-to-adjust-learning-rate
"https://pytorch.org/docs/stable/optim.html#how-to-adjust-learning-rate", UserWarning)
/home/mca/Downloads/content/ReID-MGN-master/utils/TripletLoss.py:31: UserWarning: This overload of addmm_ is deprecated:
addmm_(Number beta, Number alpha, Tensor mat1, Tensor mat2)
Consider using one of the following signatures instead:
addmm_(Tensor mat1, Tensor mat2, *, Number beta, Number alpha) (Triggered internally at /pytorch/torch/csrc/utils/python_arg_parser.cpp:882.)
dist.addmm_(1, -2, inputs, inputs.t())
total loss:16.37 Triplet_Loss :2.75 CrossEntropy_Loss:6.81 Traceback (most recent call last):
File "main.py", line 155, in
main.train()
File "main.py", line 48, in train
loss = self.loss(outputs, labels)
File "/home/mca/anaconda3/lib/python3.7/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
result = self.forward(*input, **kwargs)
File "/home/mca/Downloads/content/ReID-MGN-master/loss.py", line 14, in forward
Triplet_Loss = [triplet_loss(output, labels) for output in outputs[1:4]]
File "/home/mca/Downloads/content/ReID-MGN-master/loss.py", line 14, in
Triplet_Loss = [triplet_loss(output, labels) for output in outputs[1:4]]
File "/home/mca/anaconda3/lib/python3.7/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
result = self.forward(*input, **kwargs)
File "/home/mca/Downloads/content/ReID-MGN-master/utils/TripletLoss.py", line 38, in forward
dist_an.append(dist[i][mask[i] == 0].min().unsqueeze(0))
RuntimeError: operation does not have an identity.

Hi,
Im trying your method on different datraset, but getting bellow error.
please help me to fix

/content/ReID-MGN-master/ReID-MGN-master/utils/TripletLoss.py:56: UserWarning: This overload of addmm_ is deprecated:
addmm_(Number beta, Number alpha, Tensor mat1, Tensor mat2)
Consider using one of the following signatures instead:
addmm_(Tensor mat1, Tensor mat2, *, Number beta, Number alpha) (Triggered internally at /pytorch/torch/csrc/utils/python_arg_parser.cpp:1005.)
dist.addmm_(1, -2, inputs, inputs.t())
total loss:14.65 Triplet_Loss:2.70 CrossEntropy_Loss:5.98 Traceback (most recent call last):
File "main.py", line 155, in
main.train()
File "main.py", line 48, in train
loss = self.loss(outputs, labels)
File "/usr/local/lib/python3.7/dist-packages/torch/nn/modules/module.py", line 889, in _call_impl
result = self.forward(*input, **kwargs)
File "/content/ReID-MGN-master/ReID-MGN-master/loss.py", line 14, in forward
Triplet_Loss = [triplet_loss(output, labels) for output in outputs[1:4]]
File "/content/ReID-MGN-master/ReID-MGN-master/loss.py", line 14, in
Triplet_Loss = [triplet_loss(output, labels) for output in outputs[1:4]]
File "/usr/local/lib/python3.7/dist-packages/torch/nn/modules/module.py", line 889, in _call_impl
result = self.forward(*input, **kwargs)
File "/content/ReID-MGN-master/ReID-MGN-master/utils/TripletLoss.py", line 63, in forward
dist_an.append(dist[i][mask[i] == 0].min().unsqueeze(0))
RuntimeError: operation does not have an identity.

Thank you for sharing. The total training process lasts for 80 epochs in the paper. Why you should train the network for 400 epochs?

Hello, Can you tell me how to train and test in DukeMTMC datasets? thanks.

Hey, thanks for sharing your code!
I wonder which version of scikit-learn are you using. I found that with version > 0.18，the mAP calculated by average_precision_score can be different from common evalution code of market 1501

I found you extract the features from the same images twice, then you add those features, why do you do that?just because of the trick of computing the norm, or other thing?

@GNAYUOHZ Hi, Is there any way to test video in this code?
Thanks in advance

The paper claimed only need 80 epochs to train. Another reproduction from seathiefwang has the same problem.

The result seems to be much higher than the paper

please note that data.py needs work for out of the box execution on pretrained weights due to the way path setting is done in it. I am grateful for you making and sharing :)

thanks for you share,you say:" you can save features as .mat file for multiple uses",how to do it?

Hi,

Did you get the current best results with the default setting in the opt.py file?

Thanks!

Hi, thanks for your excellent work.

I directly use your code on Market Dataset without any additional changes.

I got only rank1 93.46% mAp 90.68% under 500 epoch, but you got 3% higher.

Can you please give some suggestion about how can I do?

Thank you very much sincerely.

 q_g_dist = np.dot(qf, np.transpose(gf)) 
 q_q_dist = np.dot(qf, np.transpose(qf))  
 g_g_dist = np.dot(gf, np.transpose(gf)) 
 dist = re_ranking(q_g_dist, q_q_dist, g_g_dist)

Thanks for sharing the codes, but I have a quetion when I read it. why the initial distance matrices for re-ranking are the correlation matrices of features, not the Mahalanobis distance ?

in main.py line.113 to line.137

I modified some programs but still can't get out.

I modified this code :

   for i in range(10):
        img_path = gallery_path[index[i]]
        
        print(img_path)

        ax = plt.subplot(1, 11, i + 2)
        ax.axis('off')
        plt.imshow(plt.imread(img_path))
        
        if img_path == gallery_label:
            ax.set_title('%d'%(i+1),color='green')
        else:
            ax.set_title('%d'%(i+1),color='red')
            
    fig.savefig("show.png")
    print('result saved to show.png')