xiahaifeng1995 / STPM-Anomaly-Detection-Localization-master

This is an unofficial implementation of the paper “Student-Teacher Feature Pyramid Matching for Unsupervised Anomaly Detection”.
MIT License
45 stars 6 forks source link

关于loss的问题 #3

Open TimZhang001 opened 3 years ago

TimZhang001 commented 3 years ago
  1. 目前的loss计算方法如下 f_loss = 0.5 * (ft_norm - fs_norm)*2 ave_loss = f_loss.sum() / (hw) t_loss += ave_loss

  2. 我感觉这样做有些不合理,只是统计意义的最小,但是针对一些局部细节,并不一定能够做到很好。是否要考虑max value的限制呢