search

Roc-Ng / DeepMIL

Real-world Anomaly Detection in Surveillance Videos CVPR2018 UCF-Crime dataset
117 stars 10 forks source link

loss的计算存在问题 #21

Closed whystopped closed 2 years ago

whystopped commented 2 years ago

loss中的sparsity和smooth项计算了两次 https://github.com/Roc-Ng/DeepMIL/blob/6c639a0e5becc53fbbfc7c0f54e2e0cc0068a4e8/train.py#L27

另外，原论文只对异常视频计算sparsity和smooth项，但代码中对所有视频都计算了，实验感觉影响不大 https://github.com/Roc-Ng/DeepMIL/blob/6c639a0e5becc53fbbfc7c0f54e2e0cc0068a4e8/train.py#L43

Roc-Ng commented 2 years ago

是的，感谢提醒。针对第一条我已经修改。