An unofficial pytorch implementation of FastFlow: Unsupervised Anomaly Detection and Localization via 2D Normalizing Flows (Jiawei Yu et al.).
As the paper doesn't give all implementation details, it's kinda difficult to reproduce its result. A very close AUROC is achieved in this repo. But there are still some confusions and a lot of guesses:
Really appreciate the inspiring discussion with the community. Feel free to comment, raise new issues or PRs.
pip install -r requirements.txtWe use MVTec-AD to verify the performance.
The dataset is organized in the following structure:
mvtec-ad
|- bottle
| |- train
| |- test
| |- ground_truth
|- cable
| |- train
| |- test
| |- ground_truth
...
Take ResNet18 as example
# train
python main.py -cfg configs/resnet18.yaml --data path/to/mvtec-ad -cat [category]
# a folder named _fastflow_experiment_checkpoints will be created automatically to save checkpoints
# eval
python main.py -cfg configs/resnet18.yaml --data path/to/mvtec-ad -cat [category] --eval -ckpt _fastflow_experiment_checkpoints/exp[index]/[epoch#].ptAs the training process is not stable, I paste both the performance of the last (500th) epoch and the best epoch.
| AUROC (last/best) | wide-resnet-50 | resnet18 | DeiT | CaiT |
|---|---|---|---|---|
| bottle | 0.987/0.989 | 0.975/0.979 | 0.931/0.959 | 0.926/0.976 |
| cable | 0.950/0.978 | 0.942/0.962 | 0.976/0.979 | 0.975/0.981 |
| capsule | 0.987/0.989 | 0.979/0.985 | 0.982/0.988 | 0.987/0.990 |
| carpet | 0.988/0.989 | 0.986/0.986 | 0.991/0.994 | 0.981/0.993 |
| grid | 0.991/0.993 | 0.973/0.985 | 0.965/0.980 | 0.968/0.970 |
| hazel nut | 0.957/0.984 | 0.922/0.963 | 0.982/0.990 | 0.981/0.992 |
| leather | 0.995/0.996 | 0.991/0.996 | 0.991/0.994 | 0.994/0.996 |
| metal nut | 0.968/0.986 | 0.950/0.966 | 0.980/0.988 | 0.977/0.984 |
| pill | 0.968/0.977 | 0.955/0.968 | 0.977/0.989 | 0.984/0.990 |
| screw | 0.969/0.987 | 0.952/0.957 | 0.990/0.990 | 0.991/0.993 |
| tile | 0.955/0.971 | 0.916/0.951 | 0.966/0.966 | 0.946/0.972 |
| toothbrush | 0.985/0.986 | 0.967/0.978 | 0.983/0.988 | 0.989/0.992 |
| transistor | 0.956/0.975 | 0.970/0.975 | 0.959/0.970 | 0.967/0.969 |
| wood | 0.948/0.964 | 0.894/0.954 | 0.960/0.963 | 0.950/0.959 |
| zipper | 0.980/0.987 | 0.969/0.979 | 0.966/0.974 | 0.972/0.984 |
| MEAN | 0.972/0.983 | 0.956/0.972 | 0.973/0.981 | 0.973/0.983 |
| Paper | 0.981 | 0.972 | 0.981 | 0.985 |