when train_backbone，got a bad performance!
Closed this issue · 4 comments
PHaiJun commented
why??
INFO:main:instance_auroc: 0.682
INFO:main:full_pixel_auroc: 0.661
INFO:main:anomaly_pixel_auroc: 0.007
INFO:main:
INFO:utils:instance_auroc: 0.682
INFO:utils:full_pixel_auroc: 0.661
INFO:utils:anomaly_pixel_auroc: 0.007
DonaldRR commented
The FREEZING backbone is one of the key in this work. The pretrained backbone models the "normal" features distribution. When it's parameters are changing over time, such distribution will be changing too.
DonaldRR commented
The essence of the training is not always Decreasing the Loss. The freezing backbone makes the features from backbone are meaningful,