【行人重识别】复现deep-person-reid中特征可视化

清疚 2022-12-08 04:51 646阅读 0赞

## 0.环境 ##

ubuntu16.04
    python3.6
    cuda9.0
    torch==1.1.0
    torchvision==0.3.0
    numpy
    Cython
    h5py
    Pillow
    six
    scipy
    opencv-python
    matplotlib
    tb-nightly
    future
    yacs
    gdown
    flake8
    yapf
    isort==4.3.21
    imageio

下载：

git clone https://github.com/KaiyangZhou/deep-person-reid.git

环境配置(创建虚拟环境+安装依赖包+安装torchreid)：

cd deep-person-reid
    pip install -i https://pypi.tuna.tsinghua.edu.cn/simple virtualenv
    virtualenv env
    source ./env/bin/activate
    
    pip  --default-time=500 install -i https://pypi.tuna.tsinghua.edu.cn/simple -r requirements.txt
    pip  --default-time=500 install -i https://pypi.tuna.tsinghua.edu.cn/simple torch==1.1.0
    pip  --default-time=500 install -i https://pypi.tuna.tsinghua.edu.cn/simple torchvision==0.3.0
    
    
    python setup.py develop

## 1.准备数据集与预训练模型 ##

## （1）数据 ##

我这里使用的是Market1501数据，所以需要先下载数据解压后，新建datasets目录，并将目录修改为market1501。

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70][]

## （2）预训练模型 ##

如果能够自动下载就自动下载，不行的话，可以从[我的链接][Link 1] 下载。后面放到/root/.cache/torch/checkpoints/目录下。

## 2.运行 ##

python tools/visualize_actmap.py --root /root/deep-person-reid/datasets/ -d market1501 -m osnet_x1_0 --weights PATH_TO_PRETRAINED_WEIGHTS --save-dir log/visactmap_osnet_x1_0_market1501

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 1][]

## 3.结果 ##

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 2][]

通过分析模型输出的特征维度，发现为\[batchsize, 512, 16, 8\] 。所以我们自己想要利用这段代码的话，就需要输出最后一个pooling层的特征用于可视化。

对两处的shape进行可视化：

# compute activation maps
    print("output before shape :{}".format(outputs.shape))
    outputs = (outputs**2).sum(1)
    print("outputs shape :{}".format(outputs.shape))
    b, h, w = outputs.size()

结果：

output before shape :torch.Size([8, 512, 16, 8])
    outputs shape :torch.Size([8, 16, 8])

根据输出分析，可知是pool层前relu后的特征。

## 4.可能bug解决 ##

raise RuntimeError('DataLoader worker (pid(s) {}) exited unexpectedly'.format(pids_str))
    RuntimeError: DataLoader worker (pid(s) 28315) exited unexpectedly

找到对应代码：改小batchsize

![20200917195020612.png][]

如果还有问题，参考[torch error:DataLoader worker (pid(s) 26462) exited unexpectedly][torch error_DataLoader worker _pid_s_ 26462_ exited unexpectedly]。

## 参考 ##

[1.deep-person-reid（girhub代码）][1.deep-person-reid_girhub]

[2.deep-person-reid环境配置][2.deep-person-reid]

[3.可视化激活map][3._map]

[4.torch error:DataLoader worker (pid(s) 26462) exited unexpectedly][torch error_DataLoader worker _pid_s_ 26462_ exited unexpectedly]

[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70]: /images/20221123/37877cf2aaa8488586a528e7d227c112.png
[Link 1]: https://download.csdn.net/download/qq_35975447/12854057
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 1]: /images/20221123/4037a782bbb1436198f1c2fb7a36430f.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 2]: /images/20221123/f95f5c3859244bc28dfc0fa44c6baf63.png
[20200917195020612.png]: /images/20221123/fec7877aac8d41e685ea1ed7efb1246b.png
[torch error_DataLoader worker _pid_s_ 26462_ exited unexpectedly]: https://blog.csdn.net/qq_35975447/article/details/107287614
[1.deep-person-reid_girhub]: https://github.com/KaiyangZhou/deep-person-reid
[2.deep-person-reid]: https://kaiyangzhou.github.io/deep-person-reid/
[3._map]: https://kaiyangzhou.github.io/deep-person-reid/user_guide.html#visualize-activation-maps