Remembering Normality

Unofficial implementation of the paper : "Remembering Normality: Memory-guided Knowledge Distillation for Unsupervised Anomaly Detection" Article

We designed the code following the guidelines of the official Paper and its Supplementary materials

You will need Python 3.10+ and the packages specified in requirements.txt.

Install packages with:

$ pip install -r requirements.txt

To run the code, please download the MVTEC AD dataset and add the path in the config.yaml file Link to download the dataset : https://www.mvtec.com/company/research/datasets/mvtec-ad

To run train and test the model :
With STPM backbone (Article)

python trainRM_ST.py

With RD backbone (Article)

python trainRM_RD.py

To modify the object categories or hyperparameters, you can modify the config.yaml file.

data_path (STR): The path to the dataset
dataset (STR): The dataset used (mvtec, visa, mvtec3d or eyecandies)
backbone (STR): The name of the model backbone (either resnet18 or wide_resnet50_2)
obj (STR): The object category
phase (STR): Either train or test
save_path (STR): The path to save the model weights
visu (BOOL): Wheter to calculate the localization score and save images of segmentation
training_data:
- epoch : number of epoch for training
- batch_size : the minibatch size
- lr : the learning rate img_size
- img_size : Size for the image resizing
- crop_size : Size for the image cropping (if equal to img_size, no cropping)
- embed_dim : The parameter L from the paper (number of keys and values)
- n_embed : The parameter N from the paper, the number of normal sample for normality embedding learning
- lambda1 : The parameter lambda1 from the paper (ponderation of the lossNM)
- lambda2 : The parameter lambda2 from the paper (ponderation of the lossORTH)

To download the datasets, please refer to the following links:
mvtec ad
visa
mvtec3d-ad
eyecandies

Feel free to ask for any improvements need in the code :)

This project is licensed under the MIT License.