This is unofficial of Anomaly Detection via Reverse Distillation from One-Class Embedding
Paper : https://arxiv.org/abs/2201.10703
-
MVtecAD Dataset : https://www.mvtec.com/company/research/datasets/mvtec-ad
-
ViSA Dataset : https://github.com/amazon-science/spot-diff
-
Since I use two difference dataset, so I unify structure of dataset. So, before running training, you need to make a
split_csv
file using make_csv4mvtecad.ipynb ,which is to make csv file containing information about dataset like whatViSA
dataset has.
Single - gpu
python Reversedistillation.py --yaml_config ./configs/mvtec.yaml
Multi - GPU
accelerate config
accelerate launch Reversedistillation.py --yaml_config ./configs/mvtec.yaml
- Before start training using multi-gpu, You need to set accelerate config so that Accelerate can use multi-gpu
Notion : https://www.notion.so/hunim/Reverse-distillation-f7a79fc9d07a4d8d8bb3d4a914465e7a