layumi / person_reid_baseline_pytorch Goto Github PK

I run the project with my trained model whose parameters include --train_all --batchsize 32 no color jitter.
when I evaluate the model,I got rank1 for 87% but map for 2.96%.
My train loss image is listed.

I find you use bottleneck instead of conv 1*1 and classfier .What effect will this operation have?

I found this line 'f = outputs.data.cpu()' very slow

Hi,
I download the pretrainded model offered by the author from Googledrive

But it went error as follow
RuntimeError: Error(s) in loading state_dict for ft_net:
Missing key(s) in state_dict: "model.fc.weight", "model.fc.bias", "classifier.add_block.0.weight", "classifier.add_block.0.bias", "classifier.add_block.1.weight", "classifier.add_block.1.bias", "classifier.add_block.1.running_mean", "classifier.add_block.1.running_var", "classifier.classifier.0.weight", "classifier.classifier.0.bias".
Unexpected key(s) in state_dict: "model.fc.0.weight", "model.fc.0.bias", "model.fc.1.weight", "model.fc.1.bias", "model.fc.1.running_mean", "model.fc.1.running_var", "classifier.0.weight", "classifier.0.bias".

How can I solve this problem? Any suggestion would be appreciable.
Thanks,
Zhang

@layumi 您好，有没有C++的开源的源码呀？

This line causes error when using Windows Plateform
Kindly changed it to filename = os.path.basename(path)

When I try to use Trained DenseNet-121 model to get features, I got this error: KeyError: 'unexpected key "model.fc.0.weight" in state_dict'.

Could you please tell me which part is wrong?

Many thanks!

It seems all the junk images are also used during the re-ranking process. But the problem is that for each query image, some of the junk images actually come from the same camera and may affect the retrieval performance if these junk images are used for re-ranking.

I want to know:

1. Do the junk images really participate in the re-ranking process?
2. If 1 is true, is it a widely accepted procedure in other published papers?

Thanks so much for providing such an excellent codebase.

I'm sorry if this is too simple a question. The model is trained as a classification problem (input were images and output was ID numbers). If the model was previously outputting ID numbers during training, how would it output retrieved images during testing?

请问，在PCB实现的特征提取部分，是融合了全局特征和局部特征吗？

When I try to train PCB model, I got an error of 'out of memory'.
python train.py --gpu_ids 0 --PCB --batchsize 64 --name PCB
When I try to train PCB model with multi GPUs, the error still occurs.
python train.py --gpu_ids 0,1 --PCB --batchsize 64 --name PCB
Thus I have 3 relevant questions.

Q1. Can the model with parameters ResNet50 + PCB + BatchSize64 is able to run on a single GPU with memory 12GB?
Q2. If Q1 is negative, can multi GPUs solve the problem of 'out of memory' by straightly setting gpu_ids '0,1'.
Q3. If Q2 is negative, how can I solve the problem.

Thank you!

I just download the Market1501 dataset and run the following command
python prepare.py
python train.py --gpu_ids 0 --name ft_ResNet50 --train_all --batchsize 32 --data_dir Market-1501-v15.09.15/pytorch/

It runs without error and the loss is decreasing with each epoch, however, the acc is abnormal with 0 or 1, just as follow

Is there something wrong?
Thanks
Zhang

Hi!
Thanks for the baseline.
The evaluation baseline works for the Single Query settings. Can you provide insights on how to modify the baseline and get the same results for multi query settings.

Thank you

Hello,
I want to test the code on DukeMTMC-reID dataset and need to improve the performance. In the README.md, it said:

Futhermore, you also can test our code on DukeMTMC-reID Dataset. Our baseline code is not such high on DukeMTMC-reID Rank@1=64.23%, mAP=43.92%. Hyperparameters are need to be tuned.

Which hyperparameters should I adjust? Any informations is helpful.
Thank you and waiting for your reply.

python train.py --gpu_ids 0 --name ft_ResNet50 --train_all --batchsize 32 --data_dir your_data_path

the error is:

net output size:
Traceback (most recent call last):
File "/reid_baseline/model.py", line 100, in <module>
print(output.shape)

i want try resnet50

I tried this code on CUHK03-NP dataset, but it didn't work well, and only arrived rank 1: 61.6 , mAp: 56.7.

This is only PCB added , not include RPP, right?

UserWarning: invalid index of a 0-dim tensor. This will be an error in PyTorch 0.5. Use tensor.item() to convert a 0-dim tensor to a Python number
train_loss += loss.data[0]

so transforming the 'train_loss += loss.data[0] ' into 'train_loss+=loss.item() ' in the 'train.py:200' can fix the warning on the new pytorch version.

I try to use other convnets, however, I meet to size problems:
When I use densenet121, I change the crop size to 224 * 224 but it occurs that
"alueError: empty range for randrange() (0,-79, -79)"
If I don't change the size to 224*224 and keep it as "256,128" it will occur
“RuntimeError: Given input size: (1024x8x4). Calculated output size: (1024x2x-2). Output size is too small at /opt/conda/conda-bld/pytorch_1512386481460/work/torch/lib/THCUNN/generic/SpatialAveragePooling.cu:63”
How can I change the size for training other nets?

wty@z:~/soft_wty/Person_reID_baseline_pytorch-master$ python3 train.py --gpu_ids 0 --name ft_ResNet50 --train_all --batchsize 32 --data_dir /home/wty/soft_wty/dataset/market1501 --erasing_p 0.5
Downloading: "https://download.pytorch.org/models/densenet121-a639ec97.pth" to /home/wty/.torch/models/densenet121-a639ec97.pth
Traceback (most recent call last):
File "/home/wty/soft_wty/local_install/lib/python3.6/urllib/request.py", line 1318, in do_open
encode_chunked=req.has_header('Transfer-encoding'))
File "/home/wty/soft_wty/local_install/lib/python3.6/http/client.py", line 1239, in request
self._send_request(method, url, body, headers, encode_chunked)
File "/home/wty/soft_wty/local_install/lib/python3.6/http/client.py", line 1285, in _send_request
self.endheaders(body, encode_chunked=encode_chunked)
File "/home/wty/soft_wty/local_install/lib/python3.6/http/client.py", line 1234, in endheaders
self._send_output(message_body, encode_chunked=encode_chunked)
File "/home/wty/soft_wty/local_install/lib/python3.6/http/client.py", line 1026, in _send_output
self.send(msg)
File "/home/wty/soft_wty/local_install/lib/python3.6/http/client.py", line 964, in send
self.connect()
File "/home/wty/soft_wty/local_install/lib/python3.6/http/client.py", line 1400, in connect
server_hostname=server_hostname)
File "/home/wty/soft_wty/local_install/lib/python3.6/ssl.py", line 407, in wrap_socket
_context=self, _session=session)
File "/home/wty/soft_wty/local_install/lib/python3.6/ssl.py", line 814, in init
self.do_handshake()
File "/home/wty/soft_wty/local_install/lib/python3.6/ssl.py", line 1068, in do_handshake
self._sslobj.do_handshake()
File "/home/wty/soft_wty/local_install/lib/python3.6/ssl.py", line 689, in do_handshake
self._sslobj.do_handshake()
ssl.SSLError: [SSL: CERTIFICATE_VERIFY_FAILED] certificate verify failed (_ssl.c:833)
During handling of the above exception, another exception occurred:

Traceback (most recent call last):
File "train.py", line 20, in
from model import ft_net, ft_net_dense, PCB
File "/home/software_mount/wty/Person_reID_baseline_pytorch-master/model.py", line 200, in
net = ft_net_dense(751)
File "/home/software_mount/wty/Person_reID_baseline_pytorch-master/model.py", line 83, in init
model_ft = models.densenet121(pretrained=True)
File "/home/wty/soft_wty/local_install/lib/python3.6/site-packages/torchvision-0.2.1-py3.6.egg/torchvision/models/densenet.py", line 35, in densenet121
File "/home/wty/soft_wty/local_install/lib/python3.6/site-packages/torch/utils/model_zoo.py", line 65, in load_url
_download_url_to_file(url, cached_file, hash_prefix, progress=progress)
File "/home/wty/soft_wty/local_install/lib/python3.6/site-packages/torch/utils/model_zoo.py", line 70, in _download_url_to_file
u = urlopen(url)
File "/home/wty/soft_wty/local_install/lib/python3.6/urllib/request.py", line 223, in urlopen
return opener.open(url, data, timeout)
File "/home/wty/soft_wty/local_install/lib/python3.6/urllib/request.py", line 526, in open
response = self._open(req, data)
File "/home/wty/soft_wty/local_install/lib/python3.6/urllib/request.py", line 544, in _open
'_open', req)
File "/home/wty/soft_wty/local_install/lib/python3.6/urllib/request.py", line 504, in _call_chain
result = func(*args)
File "/home/wty/soft_wty/local_install/lib/python3.6/urllib/request.py", line 1361, in https_open
context=self._context, check_hostname=self._check_hostname)
File "/home/wty/soft_wty/local_install/lib/python3.6/urllib/request.py", line 1320, in do_open
raise URLError(err)
urllib.error.URLError: <urlopen error [SSL: CERTIFICATE_VERIFY_FAILED] certificate verify failed (_ssl.c:833)>

pip3 list show:
cycler (0.10.0)
kiwisolver (1.0.1)
matplotlib (2.2.2)
numpy (1.15.0)
Pillow (5.2.0)
pip (9.0.3)
pyparsing (2.2.0)
python-dateutil (2.7.3)
pytz (2018.5)
PyYAML (3.13)
setuptools (39.0.1)
six (1.11.0)
torch (0.4.0)
torchvision (0.2.1)
tqdm (3.7.0)

python3 -V show:
Python 3.6.5

hi @layumi , I try to test inceptionv3, I change the forward and model.
It appears the following problems
Traceback (most recent call last): File "inceptionv3_train.py", line 346, in <module> num_epochs=500) File ''inceptionv3_train.py", line 215, in train_model outputs = model(inputs) File "/lib/python3.5/site-packages/torch/nn/modules/module.py", line 357, in __call__ result = self.forward(*input, **kwargs) File "inceptionv3_train.py", line 195, in forward x = self.classifier(x) File "/lib/python3.5/site-packages/torch/nn/modules/module.py", line 357, in __call__ result = self.forward(*input, **kwargs) File "inceptionv3_train.py", line 55, in forward x = self.add_block(x) File "/lib/python3.5/site-packages/torch/nn/modules/module.py", line 357, in __call__ result = self.forward(*input, **kwargs) File "/lib/python3.5/site-packages/torch/nn/modules/container.py", line 67, in forward input = module(input) File "/lib/python3.5/site-packages/torch/nn/modules/module.py", line 357, in __call__ result = self.forward(*input, **kwargs) File "/lib/python3.5/site-packages/torch/nn/modules/linear.py", line 55, in forward return F.linear(input, self.weight, self.bias) File "/lib/python3.5/site-packages/torch/nn/functional.py", line 837, in linear output = input.matmul(weight.t()) File "/lib/python3.5/site-packages/torch/autograd/variable.py", line 386, in matmul return torch.matmul(self, other) File "/lib/python3.5/site-packages/torch/functional.py", line 192, in matmul output = torch.mm(tensor1, tensor2) RuntimeError: size mismatch at /pytorch/torch/lib/THC/generic/THCTensorMathBlas.cu:247

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

When I use PCB to train the model ,it always occurs error "out of memory",my GPU is 1080.How much memory do I need to train the model?or change the batchsize?

And when I train the Duke dataset, the train loss and the train accurate is quite small .But the final rank-1 is about 64%as you account.Does it mean overfitting ? what should I do if I want to increase the rank-1 in Duke dataset?

Thanks a lot!

Hello,
thank you for the release code.

My sever cuda is 7.5,could be trained the model?

thank you .

Hello sir,
First thanks a lot for the great implementation, it is really a strong baseline!
Then, after studying your code, I have a doubt about the evaluation metric. You wrote in evaluate.py that:

for i in range(ngood):
        d_recall = 1.0/ngood
        precision = (i+1)*1.0/(rows_good[i]+1)
        if rows_good[i]!=0:
            old_precision = i*1.0/rows_good[i]
        else:
            old_precision=1.0
        ap = ap + d_recall*(old_precision + precision)/2

Note in the last line, there is (old_precision+precison)/2. In my understanding of AP calculation, there shouldn't be such "old_precision" term.
For example, if there are 4 GT results retrieved at position 1,2,4,7, then AP should be (1/1+2/2+3/4+4/7)/4, right? In your implementation, the result seems not to be equal to that value(should be a little bit lower in fact). So what's the intuition behind such AP calculation mechanism and is it a custom practice in re-id evaluation?
I will be really appreciated for you answer.
Best Wishes!

I found that your code doesn't include the RPP part.I did it myself , but it didn't perform that good, can you tell me how to set the learning rate,epoch nums and so on in Step3 and Step4.

Dear @layumi,
Thank you for your nice work. I have a question. What is the correct size of input image, If one want to use your released pre-trained model (i.e., for forward pass & feature extraction of input image)? I mean (256,128) or (288,144)?

请问这个代码计算损失函数时，是同时考虑了分类损失和验证损失吗？

when i run the train.py ,it work this result,

you problem can download the densenet121-a639ec97.pth automatically ,but the Great Wall may stop download.so,i download this file in hand ,and put it into /home/mrzhu/.torch/models,like this.

en.would i do for this problem,may you give a good idea .thank you very much.

Thanks for share this wonderful code! I download your code, only changed the data dir, and followed all your instruction and hyperparameters; also the same version of pytorch and torchvision, but I tried several times, only reached rank@1=~0.75 and mAP=~0.6... What would you think the reason is? Thx again, good luck with your work!

Why your images reshape to 384192 instead of 384128 which was proposed in paper.

Hi, @layumi.Thanks for the great code. I meet some problems when using the PCB.
python train.py --PCB --batchsize 10--name PCB-10. only change the batchsize to 10, When using PCB to train, It report as below:

Traceback (most recent call last):
  File "train.py", line 273, in <module>
    model = PCB(len(class_names))
  File "/home/yhangbin/code/github/Person_reID_baseline_pytorch/model.py", line 140, in __init__
    setattr(self, name, ClassBlock(2048, class_num, True, False, 256))
TypeError: __init__() takes from 3 to 4 positional arguments but 6 were given

I changed ClassBlock(2048, class_num, True, False, 256) to ClassBlock(2048, class_num, 256) this error was solved. But when training at epoch 1 , I got such errors as below in the val phase

train Loss: 3.7999 Acc: 0.0192
Traceback (most recent call last):
  File "train.py", line 332, in <module>
    model = train_model(model, criterion, optimizer_ft, exp_lr_scheduler,num_epochs=10)
  File "train.py", line 176, in train_model
    outputs = model(inputs)
  File "/home/yhangbin/anaconda3/lib/python3.6/site-packages/torch/nn/modules/module.py", line 357, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/yhangbin/code/github/Person_reID_baseline_pytorch/model.py", line 162, in forward
    predict[i] = c(part[i])
  File "/home/yhangbin/anaconda3/lib/python3.6/site-packages/torch/nn/modules/module.py", line 357, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/yhangbin/code/github/Person_reID_baseline_pytorch/model.py", line 47, in forward
    x = self.add_block(x)
  File "/home/yhangbin/anaconda3/lib/python3.6/site-packages/torch/nn/modules/module.py", line 357, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/yhangbin/anaconda3/lib/python3.6/site-packages/torch/nn/modules/container.py", line 67, in forward
    input = module(input)
  File "/home/yhangbin/anaconda3/lib/python3.6/site-packages/torch/nn/modules/module.py", line 357, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/yhangbin/anaconda3/lib/python3.6/site-packages/torch/nn/modules/batchnorm.py", line 37, in forward
    self.training, self.momentum, self.eps)
  File "/home/yhangbin/anaconda3/lib/python3.6/site-packages/torch/nn/functional.py", line 1013, in batch_norm
    return f(input, weight, bias)
RuntimeError: CHECK_ARG(tensor->nDimension >= 2 && tensor->nDimension <= 5) failed at torch/csrc/cudnn/BatchNorm.cpp:13

I don't know how to solve this problem，thanks for your help.

python: symbol lookup error: /home/sn/anaconda3/envs/pytorch0.3/lib/python3.6/site-packages/torch/_C.cpython-36m-x86_64-linux-gnu.so: undefined symbol: cudnnSetConvolutionGroupCount

I use python 3.6.2, pytorch0.3, numpy 1.14.
I got the error when I run " python test.py --gpu_ids 0 --name /home/sn/person-reid/Person_reID_baseline_pytorch/model/ft_ResNet50 --test_dir /home/sn/person-reid/Market-1501/pytorch --which_epoch 59".
Do you know how to solve this problem?

After training the resnet50 model and run test.py to generate the feature mat file,
I run the python file evaluate.py and got the result as below:
top1:0.802850 top5:0.911817 top10:0.945962 mAP:0.627301
while when I run evaluate_rerank.py, I got the bad result:
top1:0.000297 top5:0.000594 top10:0.001485 mAP:0.000627

Do you konw what is the matter with the bad result?
Looking forward to your reply.

Hi @layumi
I have read your evaluation code
Did you use the camera indexes to constraint the query and the gallery in the same cameras?
Thank you so much for your reply.

I was using trained model to extract feature by command
python test.py --gpu_ids 1 --name ft_ResNet50 --test_dir Market-1501-v15.09.15/pytorch/ --which_epoch 59
However, there exists the following error:
RuntimeError: cuda runtime error (2) : out of memory at /opt/conda/conda-bld/pytorch_1524586445097/work/aten/src/THC/generic/THCStorage.cu:58

How can I solve the problem? Any suggestion would be appreciable.

Thanks,
Zhang

问题1：文章中说的是用随机值填充，代码给的均值是统计得到的吗？
问题2：对应这种global represention用这种方法可能有效，但是针对part-based方法，反倒丢失了细节，不知是否可行？

python train.py --gpu_ids 0 --name ft_ResNet50 --train_all --batchsize 32
error occurs as below:

File "C:\Program Files (x86)\Microsoft Visual Studio\Shared\Anaconda3_64\envs\pytorch\lib\multiprocessing\spawn.py", line 144, in get_preparation_data
_check_not_importing_main()
File "C:\Program Files (x86)\Microsoft Visual Studio\Shared\Anaconda3_64\envs\pytorch\lib\multiprocessing\spawn.py", line 137, in _check_not_importing_main
is not going to be frozen to produce an executable.''')
RuntimeError:
An attempt has been made to start a new process before the
current process has finished its bootstrapping phase.

    This probably means that you are not using fork to start your
    child processes and you have forgotten to use the proper idiom
    in the main module:

        if __name__ == '__main__':
            freeze_support()
            ...

    The "freeze_support()" line can be omitted if the program
    is not going to be frozen to produce an executable.
ForkingPickler(file, protocol).dump(obj)

BrokenPipeError: [Errno 32] Broken pipe

Hi @layumi, I found you normalize the image pixel with fixed mean value and standard deviation for each channel respectively. I wonder how to determine these three pairs of values. Are they only suitable for Market-1501?

Thanks a lot! ;)

Hi, @layumi
What's the metric learing method in your code?
Thank you for much for your reply.

哲东大佬，问一下 market测试时候 single query 和multi query分别代表啥，你的测试代码是 single query么

Have you met this problem.when I am running train.py

after about ten minutes.....

Do you have any idea about this?
Thanks a lot.

-------test-----------
64
THCudaCheck FAIL file=/opt/conda/conda-bld/pytorch_1512386481460/work/torch/lib/THC/generic/THCStorage.cu line=58 error=2 : out of memory

Hello ZZD!

I change this line

CUDA_VISIBLE_DEVICES=6,7 python train.py --PCB --batchsize 60 --name PCB-64 --train_all

I use multi gpu, so add some code:
if torch.cuda.device_count() > 1 and use_gpu:
  model_wraped = nn.DataParallel(model).cuda()
  model = model_wraped

but error in forward:
RuntimeError: cuda runtime error (77) : an illegal memory access was encountered at /opt/conda/conda-bld/pytorch_1513368888240/work/torch/lib/THC/THCTensorCopy.cu:204

Hi，thanks for your sharing.I have tried to write SVDNet based on this baseline.However, the loss is not convergent.Do you have the SVDNet source code written on Pytorch?

Hi, @layumi. Thanks for the great code. I tried your code on Market-1501 and it worked very well. Now I have trained the model for DukeMTMC-reID but when running test.py I got such errors as below

RuntimeError: inconsistent tensor size, expected tensor [751 x 512] and src [702 x 512] to have the same number of elements, but got 384512 and 359424 elements respectively
RuntimeError: While copying the parameter named classifier.0.weight, whose dimensions in the model are torch.Size([751, 512]) and whose dimensions in the checkpoint are torch.Size([702, 512]).

when I run train.py using resnet50,after running for sometime, I meet a problen of ValueError: expected 2D or 3D input (got 1D input),in class ClassBlock(nn.Module): def forward(self, x): x = self.add_block(x).