【图像检索】research-ms-loss复现

骑猪看日落 2023-01-22 04:00 152阅读 0赞

代码参考：[https://github.com/msight-tech/research-ms-loss][https_github.com_msight-tech_research-ms-loss]

论文参考：[Wang\_Multi-Similarity\_Loss\_With\_General\_Pair\_Weighting\_for\_Deep\_Metric\_Learning\_CVPR\_2019\_paper.pdf][Wang_Multi-Similarity_Loss_With_General_Pair_Weighting_for_Deep_Metric_Learning_CVPR_2019_paper.pdf]

# 1.环境准备 #

ubuntu16.04
    cuda10.1
    
    torch==1.7.1
    torchvision==0.8.2
    numpy==1.19.5
    yacs==0.1.8
    Pillow==8.1.0

cd reseach-ms-loss
    python setup.py develop build

# 2.准备预训练模型 #

1）[bn\_inception-52deb4733.pth][bn_inception-52deb4733.pth]

2)   [resnet50-19c8e357.pth][]

3）将其放到

~/.cache/torch/checkpoints/目录下

4）修改ret\_benchmark/config/model\_path.py中的目录：

from yacs.config import CfgNode as CN
    
    MODEL_PATH = {
        'bninception': "~/.cache/torch/checkpoints/bn_inception-52deb4733.pth",
        'resnet50': "~/.cache/torch/checkpoints/resnet50-19c8e357.pth",
    }
    
    MODEL_PATH = CN(MODEL_PATH)

# 3.数据下载与准备 #

1）下载：[https://blog.csdn.net/zengyujianjianghu/article/details/98323836][https_blog.csdn.net_zengyujianjianghu_article_details_98323836]

2）准备：

①修改split\_cub\_for\_ms\_loss.py中的，CUB\_ROOT路径。

②运行（生成train.txt与test.txt文件，之前我误以为是ms loss对数据划分有特殊要求）：

python scripts/split_cub_for_ms_loss.py

3）修改configs/example.yaml中的目录：

DATA:
      TRAIN_IMG_SOURCE: /root/datasets/CUB_200_2011/CUB_200_2011/train.txt
      TEST_IMG_SOURCE: /root/datasets/CUB_200_2011/CUB_200_2011/test.txt

# 4.运行训练 #

bast scripts/run_cub.sh

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70][]

显存大概使用6G。

如果出现了显存不足，最好是调节下，config/example.yaml中的test batch size，为128等更小的。

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 1][]

和readme中对比，结果应该是可以复现，比较一致的。

[https_github.com_msight-tech_research-ms-loss]: https://github.com/msight-tech/research-ms-loss
[Wang_Multi-Similarity_Loss_With_General_Pair_Weighting_for_Deep_Metric_Learning_CVPR_2019_paper.pdf]: https://openaccess.thecvf.com/content_CVPR_2019/papers/Wang_Multi-Similarity_Loss_With_General_Pair_Weighting_for_Deep_Metric_Learning_CVPR_2019_paper.pdf
[bn_inception-52deb4733.pth]: http://data.lip6.fr/cadene/pretrainedmodels/bn_inception-52deb4733.pth
[resnet50-19c8e357.pth]: https://download.pytorch.org/models/resnet50-19c8e357.pth
[https_blog.csdn.net_zengyujianjianghu_article_details_98323836]: https://blog.csdn.net/zengyujianjianghu/article/details/98323836
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70]: /images/20221020/9810430722814b7898dcc100b02b87b6.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 1]: /images/20221020/bad016f563be4a7a834410c5cc4b2bdc.png