The results are different from yours

[wuqiangch]

@byungjae89 I have test the code,but I cant get the results.
bottle ROCAUC: 1.000
cable ROCAUC: 0.940
capsule ROCAUC: 0.923
carpet ROCAUC: 0.955
grid ROCAUC: 0.929
hazelnut ROCAUC: 0.987
leather ROCAUC: 1.000
metal_nut ROCAUC: 0.931
pill ROCAUC: 0.834
screw ROCAUC: 0.812
tile ROCAUC: 0.974
toothbrush ROCAUC: 0.958
transistor ROCAUC: 0.959
wood ROCAUC: 0.976
zipper ROCAUC: 0.979
Average ROCAUC: 0.944

[byungjae89]

@wuqiangch
Is there any modification from the original source code?
Since there is no training stage, randomness does not exist and the same result should be reproduced.

[TzxzT]

@byungjae89
Hello：
I just test your code, interestingly, my results are the same as @wuqiangch , except on the class 'transistor'. I get the AUROC 0.951 instead of 0.959 in wuqiangch's reproducation.
But I think this is not an insignificant problem. Do you think there are some changes on the MVTEC AD itself???（I didn't find any randomness in the code)
Here is my evidence：
byungjae89： about 6 months ago
wuqiangch: about Nov,2020 (different on many classes)
mine: Jan, 2021 (just on "transistor")
At the same time, i run the code "PaDiM", and find my results are slightly worse than the author @xiahaifeng1995, just on the class "transistor". It seems the results by xia finished on Nov,2020.
xiahaifeng1995/PaDiM-Anomaly-Detection-Localization-master#6 (comment)
Since MVTEC AD is an important dataset for scientific purpose, I'm concerned about this problem.
Hope your reply, thanks!
sorry for my english :)

[byungjae89]

@wuqiangch @TzxzT

I found the reason that why results were shown differently.

python package I used efficientnet_pytorch has version issue which makes different result by version.
lukemelas/EfficientNet-PyTorch#262

package version	ROCAUC
efficientent_pytorch==0.6.3	0.947
efficientent_pytorch==0.7.0	0.944

This repo was used efficientent_pytorch==0.6.3, so I pushed code changes 913a999 to freeze this package version.

Thanks.

[wuqiangch]

@byungjae89 OK.Thanks.