search

marco-rudolph / differnet

This is the official repository to the WACV 2021 paper "Same Same But DifferNet: Semi-Supervised Defect Detection with Normalizing Flows" by Marco Rudolph, Bastian Wandt and Bodo Rosenhahn.
215 stars 67 forks source link

test_loss is super large, is that normal？ #17

Closed renqiqi closed 3 years ago

renqiqi commented 3 years ago

when I trained the net on my own dataset, the test_loss is super large, I do not know is right or not. Can you give me some advice?

Compute loss and scores on test set: Epoch:1 test_loss:30636857344.0000 AUROC: last:0.9935 max:0.9935 epoch_max:1

marco-rudolph commented 3 years ago

As long as this is only because the anomalies in the test set have a high loss, which is good because then they are considered unlikely, it doesn't matter. These scores can distort the mean loss very much. In your case it seems to work anyway, when I look at the AUROC value. Perhaps it would be better to measure the loss separately for the anomalies and normal data. On the non-anomalies the loss should be low.

renqiqi commented 3 years ago

As long as this is only because the anomalies in the test set have a high loss, which is good because then they are considered unlikely, it doesn't matter. These scores can distort the mean loss very much. In your case it seems to work anyway, when I look at the AUROC value. Perhaps it would be better to measure the loss separately for the anomalies and normal data. On the non-anomalies the loss should be low.

thanks for your reply.