search

cvlab-yonsei / MNAD

An official implementation of "Learning Memory-guided Normality for Anomaly Detection" (CVPR 2020) in PyTorch.
336 stars 80 forks source link

About the number of epoch for training on ShanghaiTech dataset. #11

Closed linshuheng6 closed 3 years ago

linshuheng6 commented 4 years ago

Hi, Thank you for your great word! I try to reproduce the performance on the ShanghaiTech through the source code. The training settings are totally following the Implementation details. However, I only get AUC 67.8. The questions are below:

  1. why can not I get AUC 72 on Shanghai Tech? My pytorch version is 1.1.0.
  2. The training data in ShanghaiTech are more than that in Ped2 or Avenue, but the training epoch num is set as 10. Is it enough for convergence?
hyunjp commented 3 years ago

Hi, thanks for the attention to our work.

As the ShanghaiTech is large dataset, the number of iterations for one epoch is relatively bigger than other datasets. We adjusted the number of epochs according to the number of iterations.

We cannot surely figure out what’s the problem of the low performance, but there are some variance of performances due to the stochastic learning. You may have to learn the model for multiple times.

Thank you.

zugexiaodui commented 2 years ago

I got 67.5% AUC on ShanghaiTech.