douzi0248 / re-id Goto Github PK

Hi, I use your framework in my paper. How do I cite your work?

`0.03
train Loss: 1.1448 Acc: 0.9010

Epoch 29/219

0.03
train Loss: 1.1467 Acc: 0.9064

Epoch 30/219

0.03
train Loss: 1.1458 Acc: 0.9059

Epoch 31/219

0.03
train Loss: 1.1320 Acc: 0.9053

Epoch 32/219

0.03
train Loss: 1.1399 Acc: 0.9075

Epoch 33/219

0.03
train Loss: 1.1289 Acc: 0.9057

Epoch 34/219

0.03
train Loss: 1.1493 Acc: 0.9033

Epoch 35/219

`

How to decide learning rate decrease in the 60 and 130 epoch? Thx.

感谢你分享的代码，有个问题想和你讨论下。
我看到你推理的代码是
outputs_1,outputs_2,outputs_3,,,,,,= model(input_img)
outputs = torch.cat((outputs_1,outputs_2,outputs_3),1)
这里 outputs_1,outputs_2,outputs_3 在训练的时候经过分类器分别进行了单独的监督。但是为什么你认为在推理的时候3个特征concat起来效果就好呢？类似的尝试我也试过，好像concat起来并没有什么显著效果。请问你这里的出发点或者依据是什么？

感谢回答，不是很明白为什么model里return那么多。另外还有一个问题，怎么样计算map和rank1 rank5指标

I have run the training and then run the test like this "python test.py", but it complains as following:

Traceback (most recent call last):
File "test.py", line 169, in
gallery_feature = extract_feature(model,dataloaders['gallery'])
File "test.py", line 116, in extract_feature
ff = ff+f
RuntimeError: The size of tensor a (1536) must match the size of tensor b (2253) at non-singleton dimension 1

Firstly,thank you for you sharing the nice code!
And can you provide a re-ranking demo?because after i add the re-ranking code,the result is
Rank@1:0.000000 Rank@5:0.000000 Rank@10:0.000000 mAP:0.000465

如题
如果您已发布，可否发出链接参考一下呢？感谢

hello,where is the pretrained model "r50_ibn_a.pth"?

这个路径设置有啥不对么，我的路径是读到market1501那里。。。然后加上训练集和测试集的名字。。

After I modified batchsize from 32 to 64, the training outputs following errors.

Traceback (most recent call last):
File "train_11.py", line 333, in
num_epochs=320)
File "train_11.py", line 189, in train_model
loss5 = triplet(q1, labels)[0]
File "/data1/sanjie.lp/projects/person_reid/Re-ID-douzi0248/triplet_loss.py", line 110, in call
dist_mat, labels)
File "/data1/sanjie.lp/projects/person_reid/Re-ID-douzi0248/triplet_loss.py", line 66, in hard_example_mining
dist_mat[is_pos].contiguous().view(N, -1), 1, keepdim=True)
RuntimeError: invalid argument 2: size '[64 x -1]' is invalid for input with 288 elements at /pytorch/aten/src/TH/THStorage.cpp:80

请问您有没有试过同样的网络在MSMT17数据集上训练，在MSMT测试集上Rank1能达到多少？
我用您的网络在MSMT17上训练，batchsize=128，最终在MSMT测试集上Rank1只能达到55%，mAP只达到40%左右，相对于其他论文还存在明显差距，请问跟调参有很大的关系吗？ @douzi0248

up请教一下，我看你数据集的路径设置的是：data_dir = '/home/pt/下载/Market/pytorch' 然后是不是pytorch文件夹里面有train和val文件夹？我是这样设置的，但是跑出来好像出错了。。。 错误是RuntimeError: Found 0 files in subfolders of:

Hi, i have tried your code and found that was really good, and i am wondering whether you have published any relevant papar? I want to further understand your ideas in model training. Thanks

首先谢谢昨天的指点，已经可以运行了，但是请教一下运行evaluate_gpu.py测试rank准确度，需要准备什么文件么？我看好像要一个pytorch_result.mat这样的文件，这个是怎么生成的呢？

你好，我在使用这个结构训练时loss非常容易变为nan，尤其是当特征图分割的数量超过3时，loss必定变为nan。
后来发现是程序里直接把不同高度的特征concat成一维向量导致的，直接使用卷积变成1维就不会出现nan，但效果感觉并不是很好。请问你有出现过这个情况吗。

File "train.py", line 230, in train_model
loss5 = triplet(p, labels)[0]
File "/home/zzy/anaconda3/envs/baseline/lib/python3.6/site-packages/torch/nn/modules/module.py", line 547, in call
result = self.forward(*input, **kwargs)
File "/home/zzy/Downloads/Person_reID_baseline_pytorch/triplet_loss.py", line 143, in forward
dist = torch.pow(inputs, 2).sum(dim=1, keepdim=True).expand(n, n)
RuntimeError: expand(torch.cuda.FloatTensor{[16, 1, 16, 8]}, size=[16, 16]): the number of sizes provided (2) must be greater or equal to the number of dimensions in the tensor (4)

你好，在运行你的代码的时候，tripelt loss这里老是报错，请问有什么解决方法吗

test.py
里面的特征提取部分都有问题的。
而且没有提供map和rank1的计算