huanghoujing / person-reid-triplet-loss-baseline Goto Github PK

@huanghoujing

TripletLoss 里面根据margin传入与否使用了MarginRankingLoss/SoftMarginLoss()。我看了几个Triplet-Loss的代码，大家都避开直接使用TripletMarginLoss和TripletMarginWithDistanceLoss，这到底是为什么呀。

你好,
非常感谢实现了这么好的代码，最近基本将整个源码整体读了一遍。
在我复现的结果中，各项测试指标与你所提供的值基本吻合，但是在与其他方法对比的过程中，发现也有很多类似用到batch hard triplet loss的方法，但是发现此方法得到的结果超过了大多数的同类方法，因此对这一点有点疑惑。
按我理解，这个方法不同的可能在于last_conv_stride，在数据增强上只做了水平翻转，网络模型也很简洁，只提取了简单的全局特征，想探讨一下您认为是什么因素导致性能优于其他同类方法呢，是学习率设置、epoch等等的这些训练技巧吗？

Hi,
I noticed that there are only 15913 images in the test set. However, the original test set contains 19,732 images (http://www.liangzheng.org/Project/project_reid.html). Is this on purpose?

Thanks!

Hi! Great work!
I am trying to understand your model and implemented it in keras myself. There are few areas which I am uncertain of so could u give me some enlightenments?

1. For the base model Resnet-50, other than changing the stride to 1 and removing the top layer, is there any other changes you have made on it? Besides, I notice that for your model, unlike the paper, you have
   i) x = F.avg_pool2d(x, x.size()[2:])
   ii) x = x.view(x.size(0), -1)
   after coming out from Resnet. May I know what are some of the reasons you chose to do these two layers, instead of following the paper, which suggest 2 dense layer and a batchnorm in between? In your case, what would be the output shape of your last flatten layer?

2. It seems that you did not use the default dataset split provided by market1501. What are some of potential advantages using your split? In addition, why there are only 100 persons in the validation set? Shouldnt it be the same number of classes in the training set?

3. I noticed that you also did preprocessing before passing images to the model, however, the values for your image mean ([0.486, 0.459, 0.408]) are different from that provided in imagenet.preprocess ([0.485, 0.456, 0.406]) specified in keras source code. May I know what is the reason for the change?

4. In short, my keras implementation did not perform well on market1501, for the map and top1 top5 are quite low. Is it possible to give me some suggestions (anything) so that I can try to improve my model?

Thank you and hope to hear from you soon!

Dear @huanghoujing,
Have you ever trained other ConvNets (e.g., ResNet101, ResNet152, ResNext, Inception-ResNet-V2, DenseNet) as the backbone of your code?

I find that you did not put your paper on README, so how can i refer your model in my paper

Hi @huanghoujing ,
I am training the Mkaert1501 dataset only, but the performance is worse than your table's.
Are the results in the table training on all the datasets combined, or I have done something wrong.
I do as the following steps:
1
python script/dataset/transform_market1501.py
--zip_file ~/Dataset/market1501/Market-1501-v15.09.15.zip
--save_dir ~/Dataset/market1501
2
python script/experiment/train.py
-d '(0,)'
--only_test false
--dataset DATASET_NAME
--last_conv_stride STRIDE
--normalize_feature false
--trainset_part TRAINVAL_OR_TRAIN
--exp_dir EXPERIMENT_DIRECTORY
--steps_per_log 10
--epochs_per_val 5

How to train with different pre-trained models.
please reply.

你好，请问一下，我在运行visula_heatmap.py文件时，一直显示No such file or directory: '../Market-1501/pytorch/test_data'。但是我这个文件夹是存在的，这是为什么呢？

Respected Author,
I am getting a very weird result from loading your pre-trained model and trying to run the evaluation, gave the following command as described in your readme.

python script/experiment/train.py \
-d '(0,)' \
--only_test true \
--dataset market1501 \
--last_conv_stride 1 \
--normalize_feature false \
--exp_dir ~/results \
--model_weight_file ~/weights/market1501_stride1/model_weight.pth

I am getting map of 2.42 and CMC rank1 of 11.
Any pointers would really be appreciated. Thanks in advance!

I want to save my im_paths and features,but, i donnot konw how to do it

In loss.py[62]: dist_mat[is_pos].contiguous().view(N, -1), 1, keepdim=True),
I think it should be " dist_mat[is_pos].contiguous().view(-1, 1), 1, keepdim=True)"
Because dist_mat[is_pos].shape[0] may not as the same as N if there exist the same ID in a batch.

I want to use my own photos to inference, how to set up?
Can I migrate to my own test set with a pre-trained model? Or do I have to train my training set first?
thanks

I am using python 3 to run the program but at line 59 im_paths = [ospj(im_dir, n) for n in partitions['trainval_im_names']] typeError: Can't mix strings and bytes in path components will occur.
Is there a way to change this in python 3?

Hi, there is an obvious error about "val" set for the combined dataset. In script/experiment/train.py, you set part='val' at Line 165 and use that to create val_set at Line 324. However, in tri_loss/dataset/__init__.py, there is assert part in ['trainval'], "Only trainval part of the combined dataset is available now." at Line 40. So, this is a conflict and if I fix with part='trainval' at Line 165 in train.py, there will be much more problems.

你好，请问在训练过程中会出现d_ap以及d_an逐渐增大，但是d_ap依然是小于d_an的，出现这样的情况是为什么呢？

have you pre-train on imagenet?
i wanna to train other cnn-net on your baseline and i didn't find pre-train script.

I find that you didn't add any conv layer after Resnet50. So, you are trainning Resnet50, which is surprising to my, because normal net just fine-tune Resnet50, and use the after layer to discriminate, like PCB

你好，我想请教一下：关于一个epoch中的step数目 。假如--ids_per_batch=32，--ims_per_id=4，那么就意味着一个batch的大小是128，参与训练的数据大小是12936，那么一个epoch是不是应该包含101或者102个step吗？为什么我输出了一下一个epoch中step的个数只有23个呢？是我的理解有问题，还是我哪里弄错了呢？谢谢！

when training is over, when I test on dataset, it's wrong, it turn out No valid query

你好，我想请问一下，triplet-loss使用 的是三元组数据，那在代码的数据集以batch_size大小的分批次加载数据的时候，也是要加载三元组数据吗？

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.

@huanghoujing ,Hello, if I don't want to use rerank, please tell me how should I set parameters？

Hello sir,
Thanks for your terrific implementation of triplet-loss-based person re-id method! I have a small question about the evaluation metric. As is pointed out by your document, the evaluation code is borrowed from Open Re-ID project(https://github.com/Cysu/open-reid).
In the open re-id project, it seems that some images(to be specific, those with id -1) in the original "bounding_box_test" folder are ignored while calculating CMC/MAP. I found when these images are added, the CMC1 drops from 87%+ to around 80%. I wonder whether the evaluation part of most person re-id papers also ignore these distractors with label -1.
I would be very appreciated for your reply.
Best wishes!

I can use the provided model for evaluation and get reasonable results, but meet errors when loading the model trained by myself for evaluation. The following is the script I used to test using my own trained model:

python script/experiment/train.py
-d '(0,)'
--only_test true
--dataset duke
--last_conv_stride 1
--normalize_feature false
--exp_dir /export/reid_datasets/transformed_collection/trained_models/duke_stride1_evaluation
--model_weight_file /export/reid_datasets/transformed_collection/self_train/market_weight_c1000_t1_x0/ckpt.pth

The below is the error log:

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

How does this happens? Thanks!

Hello, your code is very good.
I have a question. The dataloader returns mirrored image, but I don't see you use the mirrored image in triplet loss baseline. What's the function of the mirrored?

Hi, as we can see from the paper 'In Defense of the Triplet Loss for Person Re-Identification', rank1 on Market1501 without test augmentation is 82.51%, while yours is 86.43%, I doubt that which settings make your network better, network architecture? batch size ? sampling for a batch?

Hello, I find your evaluate score is high than open-reid on market1501 and dukemtmc. You got 75% mAP which open-reid just 67% before ReRank. Why? because the set of query&gallery ?

@huanghoujing first of all, great work!

I was wondering if you have any thoughts on training this on-vehicle re-id. I'm about to give it a try, will report results.

When I tried testing the pre-trained model on market1501 dataset using python3 the given command
"python script/experiment/train.py
-d '(0,)'
--only_test true
--dataset DATASET_NAME
--last_conv_stride STRIDE
--normalize_feature false
--exp_dir EXPERIMENT_DIRECTORY
--model_weight_file THE_DOWNLOADED_MODEL_WEIGHT_FILE"
I am getting the error
"except Exception, msg:
^
SyntaxError: invalid syntax"
and when testing it with python 2.7, I am getting the following error "train.py: error: unrecognized arguments:"
Any help would be appreciated !!
Thanks !

when I run test.
I meet error like:

THCudaCheck FAIL file=/pytorch/aten/src/THC/THCTensorRandom.cu line=25 error=30 : unknown error
Traceback (most recent call last):
File "train.py", line 561, in
main()
File "train.py", line 347, in main
model_w = DataParallel(model)
File "/home/moonuke/.local/lib/python2.7/site-packages/torch/nn/parallel/data_parallel.py", line 102, in init
_check_balance(self.device_ids)
File "/home/moonuke/.local/lib/python2.7/site-packages/torch/nn/parallel/data_parallel.py", line 17, in _check_balance
dev_props = [torch.cuda.get_device_properties(i) for i in device_ids]
File "/home/moonuke/.local/lib/python2.7/site-packages/torch/cuda/init.py", line 290, in get_device_properties
init() # will define _get_device_properties and _CudaDeviceProperties
File "/home/moonuke/.local/lib/python2.7/site-packages/torch/cuda/init.py", line 143, in init
_lazy_init()
File "/home/moonuke/.local/lib/python2.7/site-packages/torch/cuda/init.py", line 161, in _lazy_init
torch._C._cuda_init()
RuntimeError: cuda runtime error (30) : unknown error at /pytorch/aten/src/THC/THCTensorRandom.cu:25

it seems like when doing TensorRandom it happens?
how to deal with it ?
very thx

Thank you for sharing the code!

In the paper, a R1 accuracy of 89.63 is reported on the CUHK03 dataset. But in your implementation, only ~60.0. How does this happen. Thanks.

Hi @huanghoujing ,
I try to train model on Market1501, but evaluation result is worse than yours.
Your score is mAP=75.29 and rank-1=89.04 on Market1501-S1, but my score is mAP=64.90 and rank-1=81.59.

My training parameters are as following (refer to your description on github):
python3 script/experiment/train.py
-d '(0,)'
--only_test false
--dataset market1501
--last_conv_stride 1
--normalize_feature false
--exp_dir out
--steps_per_log 10
--epochs_per_val 5

other parameters are default,
--trainset_part trainval
--lr_decay_type exp
--exp_decay_at_epoch 151
--total_epochs 300

I train model on Python 2.7 and Pytorch 0.3, and pip install with requirements. But I don't know why evaluation result is worse than yours. Could you give me some hints? Thanks.

RuntimeError: shape '[128, -1]' is invalid for input of size 130
ps:The batch_size = 128

Hi,
I tried to go through your code but couldn't find any fully connected layer. In your base resnet50

I use triple loss between data of two modalities to reduce the distance between different modalities of the same class and increase the distance between different modalities of different class. But when I use batch_all loss, the valid set loss has not changed; now using hard_loss, the valid set loss still has not changed. What is the reason? I found some answers that triplet is difficult to converge. What do you do to deal with triplet loss convergence?

when I try to run python script/dataset/transform_duke.py
It occured:
File "./tri_loss/utils/utils.py", line 291
except Exception, msg:
^
SyntaxError: invalid syntax

Hi. I use this code to train my own dataset, and when I train this model many times and then test them separately, some of them have high rank values but low mAP values. What is the reason? Thank.

Hi

I train and test the Market1501 with the default parameters but only got 66% for CMC1 and 43.3% for mAP.

The precision in training is about 99.99%. Do you have any idea what might be wrong?

Cheers

I've been trying to write up code that would allow me to extract features of a person, given by jpg, extract the features based on the model and then calculate the distance to other features extracted from other person shot.

Any pointers on this would be most appreciated.

大佬你好，关于market1501数据集，paper中TriNet的mAP大概是69.1%，我在stride1的情况下训练出来大概是71%，差2%感觉有点多，自己也仔细的对照了一下区别。

下面的3个问题有点啰嗦，其实我就是想问一下：如果想达到69.1%（而不是71%）的mAP，需要调哪些args？如果你时间比较紧，可以忽略下面的问题，回答这个就足够了。十分期待你的回答，谢谢！

1.训练集共有751个人，按照P=32，算下来就是751/32=24个iteration=1epoch，但是为什么paper中他搞了25k个iteration啊，算下来就是1000+个epoch左右了，这和300个epoch比起来也太夸张了吧。是有什么地方我没算对么？
2.数据增强。我看原文中是"four corner crops and one center crop, as well as a horizontally flipped copy of each"，代码中train_set是通过create_dataset函数创建的，但是create_dataset函数中没有关于数据增强的部分，请问数据增强是在什么地方做的？
3.我看model创建的时候pretrain默认是True，所以不管用trainval还是train，初始模型都是ImageNet上的ResNet50的预训练模型，对吧?