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:

• Unmatched model A.d. parameter
• Unmentioned but critical LayerNorm

Really appreciate the inspiring discussion with the community. Feel free to comment, raise new issues or PRs.

pip install -r requirements.txt

We 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#].pt

As the training process is not stable, I paste both the performance of the last (500th) epoch and the best epoch.