cqylunlun / GLASS

[ECCV 2024] Official Implementation and Dataset Release for <A Unified Anomaly Synthesis Strategy with Gradient Ascent for Industrial Anomaly Detection and Localization>
MIT License
156 stars 19 forks source link

为何运行该代码,srew类别性能不及论文中性能,仅有91.8? #16

Closed GuoRunTang closed 1 month ago

GuoRunTang commented 2 months ago

为何运行该代码,srew类别性能不及论文中性能,仅有91.8?

cqylunlun commented 2 months ago

您好,我们已经进行了多次验证,Screw类别均未出现明显下降。除此以外,问题 #13 的提问者也给出了Screw类别上合理的实验结果。请您重新检查实验配置,并考虑以下几种可能的解决方案(重要性从高到低):

  1. 每次训练开始前,清空路径./results/models,避免ckpt_best_*.pth模型文件冗余;
  2. 每次测试开始前,确保路径./results/models/backbone_0/mvtec_screw下只有一个ckpt_best_*.pth
  3. 参照 #3 ,下载Foreground Mask并将--fg设置为1 ;
  4. 测试时按照代码,默认选择ckpt_best_*.pth
  5. 确保程序运行完所有640个epoch ;
  6. 检查数据集完整性。