huanghoujing / alignedreid-re-production-pytorch Goto Github PK

您好，我按照 Deep Mutual Learning复现 classification 的 mutual learning，但是一直没效果。偶然看到您论文里这部分实验，想问一下单独只做 classification 的 mutual learning 有什么要注意的？

how do you set the dataset size?
you directly resize to 224x224? I find the ratio is strange and the image is strange

do crop the image then resize to 224x224?I find do it ,it will happen the head and feet of person disappear.

how do you do data extend?

@huanghoujing Hi, I reallly appreciate your work and want to use it to get the paper'result on cuhk03, can you help descript how to modify your code to get this result?

Hi, i am trying to reproduce your results and i want to know how many gpus are used for training?

看到你的测试结果里有
GL + LL + TWGD
GL + LL + TWLD
GL + LL + TWGALD
也就是测试时可以选择用global distance或者local distance或者两者结合。请问这个是在哪里设置？或许需要改哪里的代码？

好像设置了local_dist_own_hard_sample和llw就会用TWGALD吧？那么只用local distance进行test的话如何设置呢?

谢谢~

Dear @huanghoujing,
Do you have any plan to train your models (i.e., ResNet50) on MARS data set?
Since the MARS data set has ~ 1M training images, therefore we expect that the performance of models will be increase.

RT.
My experiment environment is Python3.6 and Pytorch0.4.0.
Thanks.

Hi Houjing,

I used this code to train with CUHK03 dataset based on the new protocol as you wrote in the readme, it performed good! But now I want to split CUHK03 dataset to training/test set based on what paper wrote, I read the code and found the information of spliting has been saved in the file of "re_ranking_train_test_split.pkl", could you please give me some advices to generate this information file for spliting? Thanks a lot!!!

你好，请教下，就是加入local分支训练后，相对于只用triplet loss训练原始Resnet50性能只提升一个点吗？

发现一个训练数据集id数量和训练时ids_per_batch的设定的问题：
train_set_ids % ids_per_batch = 1的时候会导致整个epoch最后一个batch只有一个id，此时triloss时会没有负样本，故而会报错

Can you send me a source code，My qq is：1304020120,thank you

I want to create separate folders for gallery and query images. Which part of code should i modify? Also what difference it will make on the time taken to give output?

我按照README的指示跑了一个基于combined的数据集的mutual learning的训练，最后得到了一个stdout***.txt，一个ckpt.pth和一个tensorboard文件夹，但是找不到.pth.tar模型文件。请问程序是没有保存模型文件么？还是说模型文件在ckpt.pth里，需要一些操作来提取出来？
谢谢。我是PyTorch初学者。。

I have downloaded the pre-trained weights but i am not able to run it properly as it gives error that the weights.pth file not found.

Dear Author:
This code is written elegantly. If it could be compatible with pytorch 0.2 or 0.3, it would be better.

Python version Re-ranking (Originally from re_ranking)

1. I have no idea to use it , how to feed data type to use re_ranking
   I can not understand it, why is num_query ,not feature
   code API :
   q_g_dist: query-gallery distance matrix, numpy array, shape [num_query, num_gallery]
   q_q_dist: query-query distance matrix, numpy array, shape [num_query, num_query]
   g_g_dist: gallery-gallery distance matrix, numpy array, shape [num_gallery, num_gallery]

2. it is for training when training or it is for test when test?

thanks , answer can use English or 中文

我看您的代码， 好像是用的Local， 原文用的global……有点疑惑

我按照readme运行时，出现问题Runtime error：cuda runtime error(30)：unknown error at /pytorch/aten/src/THC/THCTensorRandom.cu:25，请问有同学遇到吗？

当在同一个GPU上运行train_ml.py（-d ((0,),(0,))）时出现如下错误，到底是那里的问题呢，谢谢～～

Ep 1, 37.78s, gp 23.85%, gm 17.38%, gd_ap 13.3094, gd_an 12.2454, gL 1.4902, gdmL 1.1416, loss 2.6318
Exception in thread Thread-6:
Traceback (most recent call last):
File "/usr/lib/python2.7/threading.py", line 801, in __bootstrap_inner
self.run()
File "/usr/lib/python2.7/threading.py", line 754, in run
self.__target(*self.__args, **self.__kwargs)
File "script/experiment/train_ml.py", line 475, in thread_target
normalize_feature=cfg.normalize_feature)
File "./aligned_reid/model/loss.py", line 215, in global_loss
dist_mat, labels, return_inds=True)
File "./aligned_reid/model/loss.py", line 163, in hard_example_mining
dist_mat[is_neg].contiguous().view(N, -1), 1, keepdim=True) #1
RuntimeError: invalid argument 2: size '[32 x -1]' is invalid for input with 1095010584 elements at /pytorch/torch/lib/TH/THStorage.c:37

When first running the code for testing, I can get the correct result. But after that, I get incorrect result running the code repeatly over and over.
Moreover, the results are different every testing, why ?
I was confused, and want to know how to get the correct result.

Environment:
Python: 3.6
torch: 0.4.0

The code was modified a little with the re_ranking method:

re_r_global_q_g_dist = re_ranking(global_q_g_dist, global_q_q_dist, global_g_g_dist)
re_r_global_local_q_g_dist = re_ranking(re_r_global_q_g_dist, local_q_q_dist, local_g_g_dist)

大佬好！
我把代码移植到python3.6+pytorch0.4下面。在训练的时候，loss.py里面报错。
当一个batch块里面出现相同的ID的图片的时候这边is_pos 计算的就不是对角矩阵了。
is_pos = labels.expand(N, N).eq(labels.expand(N, N).t())
is_neg = labels.expand(N, N).ne(labels.expand(N, N).t())
然后导致后面代码：
dist_ap, relative_p_inds = torch.max(
dist_mat[is_pos_test].contiguous().view(N, -1), 1, keepdim=True)#报错
错误提示：
RuntimeError: invalid argument 2: size '[16 x -1]' is invalid for input with 18 elements at ..\src\TH\THStorage.c:37

请问这里我是否可以直接定义一个batch大小的对角阵给is_pos 。还是is_pos 代码里面就可以出现非对角矩阵。
当我将is_pos 改成一直是对角矩阵的时候。只有全局loss，loss的值开始就非常的低，请问可能是什么问题啊。

Have you tried pml, gdml, and ldml seperately? Looks like you only used gdml for your current results on Market1501.
Recently I'm having trouble reproducing the results for mutual classification loss (with my own code):(
Thx.

Hi @huanghoujing,

When training on market1501 using resnet50, the results in open-reid is 85.1 for triplet and 81.4 for softmax. But your baseline result (GL-NF) is 81.53%. Do you use softmax training as the baseline?

See https://cysu.github.io/open-reid/examples/benchmarks.html

Great work!
Best,

作者您好！问一个比较zz的问题，看您的代码train_ml.py里第132行，有关于图像归一化的参数设置，如下：
``

    self.scale_im = True
    self.im_mean = [0.486, 0.459, 0.408]
    self.im_std = [0.229, 0.224, 0.225]

这个im_mean和im_std是如何计算得到的呢？是将三个数据集的所有图片的像素值加权求均值和方差吗？

使用train.py训练后保存的ckpt文件进行测试的时候提示

Keys not found in source state_dict: 
	 base.layer3.0.bn1.weight
	 base.layer1.0.bn1.bias
	 base.layer2.1.bn2.running_mean
	 base.layer2.2.conv1.weight
	 base.layer3.3.bn3.running_mean
	 base.layer2.1.bn3.weight
	 base.layer2.0.bn3.running_var
	 base.layer2.3.bn3.running_mean
	 base.layer1.2.bn1.bias
	 base.layer4.2.bn3.running_mean
	 base.layer3.1.conv2.weight
	 base.layer2.0.downsample.1.running_var
	 base.layer3.4.bn3.running_var
	 base.layer3.5.bn2.running_mean
	 base.layer3.2.bn2.running_mean
	 base.layer3.5.bn1.running_var
	 base.layer3.5.bn1.running_mean
	 base.layer1.0.bn2.bias
	 base.layer1.2.bn3.running_var
	 base.layer3.5.bn3.running_var
	 base.layer2.3.bn3.weight
	 base.layer1.0.downsample.1.weight
	 base.layer1.0.conv1.weight
	 base.layer1.1.bn2.bias
	 base.layer3.5.bn2.running_var
	 base.layer1.2.bn2.bias
	 base.layer4.0.bn1.running_var
	 base.layer2.0.bn1.running_mean
	 base.layer2.2.bn3.running_var
	 base.layer1.0.bn3.weight
	 base.layer2.2.bn2.bias
	 base.layer3.3.bn1.bias
	 base.layer3.4.bn3.bias
	 base.layer4.1.conv1.weight
	 base.layer3.1.bn2.bias
	 base.layer4.0.downsample.0.weight
	 base.layer4.0.downsample.1.weight
	 base.layer3.2.bn1.bias
	 base.layer2.1.bn1.weight
	 base.layer1.2.bn1.running_mean
	 base.layer3.0.bn3.running_mean
	 base.layer3.4.conv3.weight
	 base.layer2.1.bn3.bias
	 base.layer3.2.bn1.running_mean
	 base.layer2.2.bn1.bias
	 base.layer2.2.bn3.running_mean
	 base.layer3.1.bn1.running_mean
	 base.layer4.1.bn1.bias
	 base.layer3.2.bn3.running_mean
	 base.layer2.2.bn1.weight
	 base.layer1.1.bn3.bias
	 base.layer2.1.conv3.weight
	 base.layer4.0.downsample.1.running_mean
	 base.layer4.1.conv2.weight
	 base.layer3.2.bn3.running_var
	 base.layer3.1.bn1.bias
	 base.layer3.4.bn1.running_var
	 base.layer4.0.bn1.weight
	 base.layer2.0.bn2.running_mean
	 base.layer3.1.bn2.running_mean
	 base.layer3.3.bn3.running_var
	 base.layer4.1.bn3.running_var
	 base.layer2.2.bn2.running_mean
	 base.layer3.0.bn3.running_var
	 base.layer4.0.bn2.running_mean
	 base.layer4.1.bn2.running_var
	 base.layer2.3.conv2.weight
	 base.layer3.0.bn1.running_mean
	 base.layer1.1.bn1.bias
	 base.layer4.0.bn3.weight
	 base.layer4.2.bn2.weight
	 base.layer2.0.conv2.weight
	 base.layer1.0.bn3.running_mean
	 base.layer1.0.downsample.1.running_mean
	 base.layer2.1.bn1.running_var
	 base.layer3.4.bn2.running_mean
	 base.layer3.1.bn3.weight
	 base.layer3.0.downsample.1.weight
	 base.layer2.3.bn2.running_var
	 base.layer3.1.bn1.running_var
	 base.layer3.0.bn2.weight
	 base.layer4.2.bn3.weight
	 base.layer3.3.conv3.weight
	 base.layer2.2.bn3.weight
	 base.layer4.2.bn1.bias
	 base.layer3.5.bn3.weight
	 base.layer1.2.conv3.weight
	 base.layer3.2.conv3.weight
	 base.layer2.3.bn2.running_mean
	 base.bn1.running_mean
	 base.layer3.3.bn2.weight
	 base.layer3.2.bn2.running_var
	 base.layer2.1.bn2.running_var
	 base.layer1.2.bn2.running_var
	 base.layer1.0.downsample.0.weight
	 base.layer4.2.bn2.running_var
	 base.layer2.1.bn2.weight
	 base.layer2.1.conv2.weight
	 base.layer3.2.bn2.weight
	 base.layer3.0.bn1.running_var
	 base.layer3.3.bn1.running_mean
	 base.layer3.4.bn1.weight
	 base.layer1.1.bn3.running_mean
	 base.layer3.0.bn3.weight
	 base.layer2.2.bn1.running_var
	 base.layer3.5.bn3.running_mean
	 base.layer1.1.bn1.running_mean
	 base.layer2.1.conv1.weight
	 base.conv1.weight
	 base.layer4.0.bn1.bias
	 base.layer3.2.bn2.bias
	 base.layer3.5.conv2.weight
	 base.layer2.0.bn2.running_var
	 base.layer2.3.bn2.bias
	 base.layer2.3.bn3.running_var
	 base.layer2.0.bn1.running_var
	 base.layer2.0.downsample.1.running_mean
	 base.layer4.0.bn3.running_mean
	 base.layer1.0.bn1.running_mean
	 base.layer1.2.bn2.running_mean
	 base.layer4.0.bn2.bias
	 base.layer1.2.bn2.weight
	 base.layer3.1.conv1.weight
	 base.layer3.3.bn3.bias
	 base.layer2.2.conv3.weight
	 base.bn1.bias
	 base.layer3.3.bn2.bias
	 base.layer2.3.bn1.weight
	 base.layer3.0.downsample.1.running_var
	 base.layer3.3.bn1.weight
	 base.layer1.0.conv2.weight
	 base.layer4.1.bn1.weight
	 base.layer3.3.bn3.weight
	 base.layer1.2.conv2.weight
	 base.layer1.0.bn2.weight
	 base.layer2.2.bn1.running_mean
	 base.layer1.0.bn3.bias
	 base.layer3.2.conv1.weight
	 base.layer4.2.bn3.bias
	 base.layer1.1.bn3.running_var
	 base.layer3.4.conv2.weight
	 base.layer1.0.bn3.running_var
	 base.layer4.0.conv3.weight
	 base.layer3.5.bn1.weight
	 base.layer3.1.bn1.weight
	 base.layer3.1.bn2.running_var
	 base.layer4.2.bn2.bias
	 base.layer4.0.downsample.1.bias
	 base.layer4.0.conv1.weight
	 base.layer3.4.bn2.running_var
	 local_conv.weight
	 base.layer3.0.bn2.bias
	 base.layer3.0.conv1.weight
	 base.layer2.2.bn2.running_var
	 base.layer4.2.conv1.weight
	 base.layer3.5.bn2.weight
	 base.layer1.1.bn1.weight
	 base.layer3.0.conv2.weight
	 base.layer4.2.conv2.weight
	 base.layer3.0.downsample.0.weight
	 base.layer3.5.bn2.bias
	 base.layer2.3.bn1.running_var
	 base.layer3.5.conv3.weight
	 base.layer2.0.bn3.running_mean
	 base.layer2.1.bn1.running_mean
	 base.layer1.0.conv3.weight
	 base.layer3.0.downsample.1.bias
	 base.layer4.1.bn3.running_mean
	 base.layer3.4.bn1.bias
	 local_bn.weight
	 base.layer3.0.bn2.running_mean
	 base.layer2.1.bn2.bias
	 base.layer2.3.bn1.running_mean
	 base.layer4.2.bn3.running_var
	 base.layer1.1.bn2.weight
	 base.layer1.1.bn2.running_mean
	 base.layer3.1.conv3.weight
	 base.layer3.1.bn3.running_mean
	 base.layer1.2.bn3.bias
	 base.layer1.1.bn1.running_var
	 base.layer3.4.bn2.weight
	 base.layer1.0.downsample.1.bias
	 base.layer4.2.conv3.weight
	 base.layer3.0.bn3.bias
	 base.layer3.0.conv3.weight
	 base.layer3.4.bn1.running_mean
	 base.layer2.0.bn3.bias
	 base.layer3.3.conv1.weight
	 base.layer2.0.bn3.weight
	 base.layer3.1.bn3.bias
	 base.layer1.0.bn2.running_mean
	 base.layer1.1.conv2.weight
	 base.layer2.0.conv3.weight
	 base.layer1.1.bn3.weight
	 base.layer4.0.conv2.weight
	 base.layer2.1.bn3.running_var
	 base.layer3.3.bn2.running_mean
	 local_bn.bias
	 base.layer2.2.bn2.weight
	 base.layer3.1.bn2.weight
	 base.layer4.1.bn2.weight
	 base.layer4.2.bn1.running_mean
	 base.layer3.5.conv1.weight
	 base.layer4.0.bn2.weight
	 base.layer4.0.bn3.bias
	 base.layer2.3.conv3.weight
	 base.layer3.1.bn3.running_var
	 base.layer2.2.bn3.bias
	 base.layer1.0.bn1.running_var
	 base.layer4.1.bn1.running_var
	 base.layer2.0.downsample.0.weight
	 base.layer4.1.bn3.bias
	 base.layer4.2.bn2.running_mean
	 base.layer2.0.bn2.weight
	 fc.bias
	 base.layer1.2.bn3.weight
	 base.layer1.2.bn1.weight
	 base.bn1.running_var
	 base.layer3.2.bn1.weight
	 base.layer3.0.downsample.1.running_mean
	 base.layer4.1.bn2.running_mean
	 base.layer4.1.bn2.bias
	 base.layer3.5.bn3.bias
	 base.bn1.weight
	 base.layer3.0.bn1.bias
	 base.layer4.1.conv3.weight
	 base.layer3.2.bn3.bias
	 base.layer2.0.bn1.weight
	 base.layer1.0.bn1.weight
	 base.layer2.3.conv1.weight
	 base.layer2.0.conv1.weight
	 base.layer3.4.bn2.bias
	 base.layer3.3.bn2.running_var
	 base.layer2.0.bn1.bias
	 local_conv.bias
	 base.layer3.3.conv2.weight
	 base.layer3.5.bn1.bias
	 base.layer2.0.downsample.1.bias
	 base.layer2.1.bn1.bias
	 base.layer4.1.bn1.running_mean
	 base.layer4.2.bn1.running_var
	 base.layer2.0.downsample.1.weight
	 base.layer3.4.bn3.weight
	 base.layer1.1.bn2.running_var
	 base.layer1.1.conv1.weight
	 local_bn.running_mean
	 base.layer3.4.conv1.weight
	 base.layer4.1.bn3.weight
	 base.layer2.3.bn1.bias
	 base.layer3.4.bn3.running_mean
	 base.layer3.2.conv2.weight
	 base.layer1.2.bn3.running_mean
	 base.layer4.0.downsample.1.running_var
	 base.layer4.0.bn3.running_var
	 base.layer2.3.bn2.weight
	 base.layer4.2.bn1.weight
	 base.layer1.2.bn1.running_var
	 fc.weight
	 base.layer3.2.bn1.running_var
	 base.layer2.0.bn2.bias
	 base.layer3.3.bn1.running_var
	 base.layer2.3.bn3.bias
	 base.layer2.2.conv2.weight
	 base.layer4.0.bn2.running_var
	 base.layer3.0.bn2.running_var
	 base.layer3.2.bn3.weight
	 local_bn.running_var
	 base.layer1.0.bn2.running_var
	 base.layer1.0.downsample.1.running_var
	 base.layer1.2.conv1.weight
	 base.layer2.1.bn3.running_mean
	 base.layer4.0.bn1.running_mean
	 base.layer1.1.conv3.weight
Keys not found in destination state_dict: 
	 scores
	 state_dicts
	 ep

使用您提供的ckpt文件不会出现以上提示，请问这个可能是哪里的问题呢,望指教，谢谢~

我的环境：
Python 3.6.5 :: Anaconda, Inc.
altgraph==0.15
backports.weakref==1.0rc1
bleach==1.5.0
certifi==2018.4.16
future==0.16.0
h5py==2.6.0
html5lib==0.9999999
macholib==1.9
Markdown==2.2.0
numpy==1.11.3
opencv-python==3.2.0.7
pefile==2017.11.5
Pillow==5.1.0
protobuf==3.5.2.post1
PyInstaller==3.3.1
PyYAML==3.12
pyzmq==17.0.0
scikit-learn==0.18.1
scipy==0.18.1
six==1.11.0
tensorboardX==0.8
tensorflow==1.2.0
torch==0.3.1
torchvision==0.2.0
Werkzeug==0.14.1

请问，我如何可以进行测试，直观地查看我的网络训练情况？

please ask you where is model_weight.pth when training is over, I can,t find the final model_weight.pth , only see one ckpt.pth . whereas to test ckpt.pth , the result is very slow.

Hello! I am a student who is using your code calculating the CMC. But recently I'm stuck with your code when calculating the CMC in aligned_reid/utils/metrics.py :
matches = (gallery_ids[indices] == query_ids[:, np.newaxis])
I wonder what this line returns and what the gallert_ids[indices] and the query_ids[:,np.newaxis] respectively
looks like. I would appreciate it if you could give me an simple example to explain this line! thank you very much!

Thanks for the reproduction code!

I trained a model with triploss (use train.py), combined train data (market1501, duke and cuhk03), and the other parameters kept same as the settings to train only on market1501. I found that both the mAP and rank1 are ~1% less than the result of trained only on market1501. Should I tune the parameters MORE CAREFULLY to get a better performance?

btw, I have reproduced the result only trained on market1501. (got 87.14% rank-1 and 71.36% mAP w/o R.R.)

Thank you!

@huanghoujing you seem to be actively maintaining both projects and both work very well. I'm wondering what are the main differences between both projects?

Have you ever tried that trainning on one dataset, and test on another dataset?
The result I tried is abjective.

作者您好，我看了re-rank.py的代码，有一点很困惑：当寻找prob image的K近邻时，为何在prob images 和 gallary images中一起查找，不应该只在gallary images查找吗？prob images 和 gallary images合并以后，查找到的K近邻，可能一部分来自prob images。请作者解释一下，谢谢~

@huanghoujing hi, I run the train.py but in the 19 lines :“from aligned_reid.dataset import create_dataset”，but i can't find this file in "dataset" .

Hi @huanghoujing ,

I find that the distance between local features performs twice normalization.
1: https://github.com/huanghoujing/AlignedReID-Re-Production-Pytorch/blob/master/aligned_reid/tri_loss/model/Loss.py#L169
2: https://github.com/huanghoujing/AlignedReID-Re-Production-Pytorch/blob/master/aligned_reid/tri_loss/model/Loss.py#L75

After twice normalize the final distance is less than 1. I wonder whether this is a problem?

Best,

root@omnisky:/raid/wtz/AlignedReID-Re-Production-Pytorch# python script/dataset/transform_cuhk03.py \

--zip_file ~/Dataset/cuhk03/cuhk03_release.zip
--train_test_partition_file ~/Dataset/cuhk03/re_ranking_train_test_split.pkl
--save_dir ~/Dataset/cuhk03
Extracting zip file
Saving images 1400/1467
Traceback (most recent call last):
File "script/dataset/transform_cuhk03.py", line 152, in
transform(zip_file, train_test_partition_file, save_dir)
File "script/dataset/transform_cuhk03.py", line 84, in transform
raise RuntimeError('Train/test partition file should be provided.')
RuntimeError: Train/test partition file should be provided.

`
Traceback (most recent call last):

File "script/experiment/train.py", line 632, in
main()

File "script/experiment/train.py", line 316, in main
train_set = create_dataset(**cfg.train_set_kwargs)

File "./aligned_reid/dataset/init.py", line 54, in create_dataset
partitions = load_pickle(partition_file)

File "./aligned_reid/utils/utils.py", line 24, in load_pickle
assert osp.exists(path)

AssertionError
`
Hi, I am getting this error when Im trying to run the code to test on the market-1501 dataset. Im new to Pytorch, so please help. Im using PyTorch 3 and python 2.7 as directed in the code. Thanks

Thank you for sharing the code!

when I write the codes as follow ,

if osp.exists(train_test_partition_file):
            train_test_partition = load_pickle(train_test_partition_file)
        else:
            raise RuntimeError("Train/test partition file should be provided.'") 
trainval_pids=set()
query_pids = set()
gallery_pids = set()
for im_type in ['detected','labeled']:
            trainval_im_names = train_test_partition[im_type]['train_im_names']
            query_im_names = train_test_partition[im_type]['query_im_names']
            gallery_im_names = train_test_partition[im_type]['gallery_im_names']
            
            trainval_pids.update(set([parse_im_name(n,'id') for n in trainval_im_names]))
            
            query_pids.update(set([parse_im_name(n,'id') for n in query_im_names]))
            gallery_pids.update(set([parse_im_name(n,'id') for n in gallery_im_names]))
           
        
print (trainval_pids & gallery_pids)
assert query_pids <= gallery_pids
assert trainval_pids.isdisjoint(gallery_pids)

It causes "AssertionError".
I find that there is an intersection between trainval and gallery. (set([1201, 1389]))
The training/testing partition file is “re_ranking_train_test_split.pkl” downloaded from the given link.
Looking forward for your reply, thank you.

Hi @huanghoujing,
How many milliseconds did you costed in training time for per mini-batch?
Thanks.

作者你好，请问在duke数据集上测试，参数设置和market1501有什么不同吗？为什么mAP会这么低？

作者你好,！
我发现在open-reid中公布的market1501上triplet loss的rank1有85%左右，我运行源码用一块GPU，batch_size为128，初始学习率没调整，150轮之后rank1只有79%多一点，然后我又把初始学习率减半了，batch_size还是128，结果还是79%多，过不了80%，和公布结果差不少，用softmax loss结果都有80.5%的rank1准确率，比公布的结果只低一丁点。但是到triplet loss这反而比softmax loss还低。我看您研究过open-reid中的triplet loss，发表过很多issues，所以想请教一下您。另外我用的Python2.7.3，pytorch0.2，open-reid0.2.0。不光在market1501上，在cuhk03上也是，比公布的结果低8个点，但是比。
我结合分类损失和triplet loss在market1501上也出现了效果下降的现象，不知道作者您解决了吗？
感谢！

hi, I always get this error “RuntimeError: cuda runtime error (2) : out of memory at /opt/conda/conda-bld/pytorch_1503961620703/work/torch/lib/THC/generic/THCStorage.cu:66” . is the batchsize too big? How can I change it?

Hi @huanghoujing,

Look like you have implemented the Triplet Local Loss. So, if you did, you forget to check the TODO list of Triplet Local Loss.

When I run the program, there is a mistake ；ModuleNotFoundError: No module named 'aligned_reid'，What is the reason for it ？ I've just been in contact with Python and hope that the question can be answered,thank you

在您的代码中找到两个单词，怎么翻译都感觉不符您意：
query
gallery
query-gallery
query-query
gallery-gallery
可否告知在您的代码里，这两个单词所代表的含义？
感激不尽！

@huanghoujing Hello, I use the train_ml.py to train, when training is over , please tell me how to test and how to save model as model.pth.tar? Thank you!

• 您好，按照你所写的，我们得出logs文件之后，运用tensorboard得出的网址是https://9408:6006，请问是因为这个网址是错误的打不开还是怎么回事呢？

我有点不是很理解，为什么您的代码实现的时候，图像的输入是256x128的尺寸呀，原论文不是224x224吗；顺便问一下改成你这样的是有什么优点吗，以及图像的不同尺寸对最终效果有影响吗。希望您能帮我解惑，谢谢啦！