【行人重识别】fast-reid复现（20210119-v1.0.0）

朱雀 2023-01-08 05:29 247阅读 0赞

参考代码：

[https://github.com/JDAI-CV/fast-reid/tree/v1.0.0][https_github.com_JDAI-CV_fast-reid_tree_v1.0.0]

# 0.环境 #

ubuntu16.04
    cuda9.0
    python3.6
    torch==1.1.0
    torchvision==0.3.0
    Cython
    yacs
    tensorboard
    future
    termcolor
    sklearn
    tqdm
    opencv-python==4.1.0.25
    matplotlib
    scikit-image
    numpy==1.16.4
    faiss-gpu==1.6.3

安装apex（不要直接通过pip安装）：

git clone https://www.github.com/nvidia/apex
    cd apex
    # python setup.py install --cuda_ext --cpp_ext 
    pip install -v --no-cache-dir ./

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70][]

# 1.准备数据 #

参考[https://blog.csdn.net/qq\_35975447/article/details/106664593][https_blog.csdn.net_qq_35975447_article_details_106664593]

数据目录结构如下：

fast-reid
    	datasets
    		Market-1501-v15.09.15
    			bounding_box_train
    			bounding_box_test
    			query

# 2.修改与训练 #

修改的部分与[https://blog.csdn.net/qq\_35975447/article/details/112482765][https_blog.csdn.net_qq_35975447_article_details_112482765]一样，直接就可以训练起来：

CUDA_VISIBLE_DEVICES="0,1" python ./tools/train_net.py --config-file='./configs/Market1501/sbs_R101-ibn.yml'

![20210119120057310.png][]

# 3.知识蒸馏训练 #

此处准备的是Market1501数据集，将yaml文件改为：

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 1][]

1）训练教师模型：

CUDA_VISIBLE_DEVICES='0,1' python ./projects/FastDistill/train_net.py --config-file ./projects/FastDistill/configs/sbs_r101ibn.yml  --num-gpus 2

![20210119121409756.png][]

无Non-local， image size 也只有 128x256的结果：

![20210119144450522.png][]

2）单独训练r34：

CUDA_VISIBLE_DEVICES='0' python ./projects/FastDistill/train_net.py --config-file ./projects/FastDistill/configs/sbs_r34.yml  --num-gpus 1

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 2][]

3）loss，r101作为教师模型训练学生模型：

CUDA_VISIBLE_DEVICES='0,1' python ./projects/FastDistill/train_net.py --config-file ./projects/FastDistill/configs/kd-sbs_r101ibn-sbs_r34.yml  --num-gpus 2

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 3][]

4）loss，以r34作为预训练模型训练，r101作为教师模型训练学生模型：

CUDA_VISIBLE_DEVICES='0,1' python ./projects/FastDistill/train_net.py --config-file ./projects/FastDistill/configs/kd-sbs_r101ibn-sbs_r34.yml  --num-gpus 2 MODEL.WEIGHTS projects/FastDistill/logs/market1501/r34/model_best.pth

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 4][]

5）loss+overhaul distillation，r101作为教师模型训练学生模型：

CUDA_VISIBLE_DEVICES='0,1' python ./projects/FastDistill/train_net.py --config-file ./projects/FastDistill/configs/kd-sbs_r101ibn-sbs_r34.yml  --num-gpus 2 --dist-url 'tcp://127.0.0.1:49153' MODEL.META_ARCHITECTURE DistillerOverhaul

![20210120092818691.png][]

6）loss+overhaul distillation，以r34作为预训练模型训练，r101作为教师模型训练学生模型：

![20210119224746722.png][]

7）Market1501数据集上对比表格（下面的表格写在知乎了，写表格太麻烦，直接截图过来了）：

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 5][]

<table style="width:500px;"> 
 <thead> 
  <tr> 
   <th>Model</th> 
   <th>Rank@1</th> 
   <th>mAP</th> 
  </tr> 
 </thead> 
 <tbody> 
  <tr> 
   <th>R101_ibn (teacher)</th> 
   <td>95.52%</td> 
   <td>88.75%</td> 
  </tr> 
  <tr> 
   <th>R34 (student)&nbsp;&nbsp; &nbsp;</th> 
   <td>91.95%</td> 
   <td>79.60%</td> 
  </tr> 
  <tr> 
   <th>JS Div</th> 
   <td>94.74%</td> 
   <td>86.85%</td> 
  </tr> 
  <tr> 
   <th>JS Div+R34</th> 
   <td>94.71%</td> 
   <td>86.60%</td> 
  </tr> 
  <tr> 
   <th>JS Div&nbsp;+ Overhaul</th> 
   <td>94.80%</td> 
   <td>87.39%</td> 
  </tr> 
  <tr> 
   <th>JS Div&nbsp;+ Overhaul+R34</th> 
   <td>95.19%</td> 
   <td>87.33%</td> 
  </tr> 
 </tbody> 
</table>

是否加入r34作为预训练模型，结果差异不大。实际上训练知识蒸馏是不需要提前预训练学生模型的。根据知识蒸馏结果显示，知识蒸馏还是很有效的（涨点很多）。

[https_github.com_JDAI-CV_fast-reid_tree_v1.0.0]: https://github.com/JDAI-CV/fast-reid/tree/v1.0.0
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70]: /images/20221119/d238cc28b1c64efc88cffb378f0b8412.png
[https_blog.csdn.net_qq_35975447_article_details_106664593]: https://blog.csdn.net/qq_35975447/article/details/106664593
[https_blog.csdn.net_qq_35975447_article_details_112482765]: https://blog.csdn.net/qq_35975447/article/details/112482765
[20210119120057310.png]: /images/20221119/6a824592d77d43b589e1bc10f6ac44e6.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 1]: /images/20221119/0b51fd07185a470eb2eff7d6f5190909.png
[20210119121409756.png]: /images/20221119/b103e0724ee4452babb772690d26a835.png
[20210119144450522.png]: /images/20221119/5ea8549b3f7a43729fc557166cf9c74d.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 2]: /images/20221119/f215e7b40bfb479ab64933a7720cd8c2.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 3]: /images/20221119/702f7798585d400b868cc400c69690a0.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 4]: /images/20221119/52df4297f55c42a887152ef35e7ab6bf.png
[20210120092818691.png]: /images/20221119/41698e96c3344807b1dc88956434df67.png
[20210119224746722.png]: /images/20221119/ebf1de4de41d4e7d9457659cd6f346a0.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 5]: /images/20221119/69728b2e4c2340e38600834d8ebe8377.png