byungjae89 / mahalanobisad-pytorch

PyTorch implementation of "Modeling the Distribution of Normal Data in Pre-Trained Deep Features for Anomaly Detection"

Modeling the Distribution of Normal Data in Pre-Trained Deep Features for Anomaly Detection in PyTorch

PyTorch implementation of Modeling the Distribution of Normal Data in Pre-Trained Deep Features for Anomaly Detection.

This paper presents an anomaly detection approach that consists of fitting a multivariate Gaussian to normal data in the pre-trained deep feature representations, using Mahalanobis distance as anomaly score.
It is a simple yet effective approach and achieves SOTA on MVTec AD dataset.

Prerequisites

• python 3.6+
• PyTorch 1.5+
• efficientnet_pytorch == 0.6.3
• sklearn, matplotlib

Install prerequisites with:

pip install -r requirements.txt

If you already download MVTec AD dataset, move a file to data/mvtec_anomaly_detection.tar.xz.
If you don't have a dataset file, it will be automatically downloaded during the code running.

Usage

To test this implementation code on MVTec AD dataset:

cd src
python main.py

After running the code above, you can see the ROCAUC results in src/result/roc_curve_{model_name}.png

Results

Below is the implementation result of the test set ROCAUC on the MVTec AD dataset.

	Paper	Implementation
bottle	-	100.0
cable	-	94.2
capsule	-	92.3
carpet	-	98.1
grid	-	94.6
hazelnut	-	98.6
leather	-	100.0
metal_nut	-	94.3
pill	-	83.4
screw	-	78.1
tile	-	98.6
toothbrush	-	96.7
transistor	-	96.1
wood	-	98.5
zipper	-	97.7
Average	94.8	94.7

ROC Curve

mahalanobisad-pytorch's Issues

How to calculate the threshold value？

Thank you very much for your excellent work！
In the evaluation part of your code, only the auroc part is calculated. If it is used in inference, how to calculate the threshold of anomaly detection?

when running efficientnet-b0, error occurs

Loaded pretrained weights for efficientnet-b0
| feature extraction | train | bottle |: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 7/7 [00:32<00:00, 4.58s/it]
Traceback (most recent call last):
File "main.py", line 150, in
main()
File "main.py", line 66, in main
mean = torch.mean(torch.cat(train_output, 0).squeeze(), dim=0).cpu().detach().numpy()
RuntimeError: There were no tensor arguments to this function (e.g., you passed an empty list of Tensors), but no fallback function is registered for schema aten::_cat. This usually means that this function requires a non-empty list of Tensors. Available functions are [CPU, CUDA, QuantizedCPU, Autograd, Profiler, Tracer, Autocast]

The results are different from yours

@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