hirokiadachi / anomaly-detection-on-mvtec

Anomaly Detection on MVTec dataset

This repo. is the implementation of ICLR 2020 paper, "ITERATIVE ENERGY-BASED PROJECTION ON A NORMAL DATA MANIFOLD FOR ANOMALY LOCALIZATION", on Pytorch.

Paper link: https://arxiv.org/abs/2002.03734

The network architecture and training details are same as "Improving Unsupervised Defect Segmentation by Applying Structural Similarity To Autoencoders".

training phase

python3 train.py --data_root ./mvtec_anomaly_detection --category grid --gpu 0 --epochs 20000 --model_type 'ae'

• data_root: Please refer to the path up to the MVTec dataset directory.
• category: Please select the category that you want to train.
• model_type: Please choose the model type from ae or vae.

evaluation phase (anomaly detection)

python3 evaluation.py --gpu 0 --model_path ./checkpoint/grid/model --data_root ./mvtec_anomaly_detection --category grid --model_type 'ae' --use_grad

• model_path: Please refer to the trained model path.
• data_root: Please refer to the path up to the MVTec dataset directory.
• category: Please select the category that you want to evaluate.
• model_type: Please choose the model type from ae or vae.
• use_grad: If you activate this argument, start to iterative update using model gradients.

Dataset

This implementation presupporses to use MVTec AD, so if you'd like to excuse this, you should get MVTec AD bellow link.
MVTec AD: https://www.mvtec.com/company/research/datasets/mvtec-ad

Categories

Textures

• carpet
• grid
• leather
• tile
• wood

Objects

• bottle
• cable
• capsule
• hazelnut
• metalnut
• pill
• screw
• toothbrash
• transistor
• zipper

data structures

mvtec_anomaly_detection
 |--bottle
 |  |--ground_truth
 |  |  |--broken_large
 |  |  |  |--000_mask.png
 |  |  |  |     :
 |  |  |--broken_small
 |  |  |  |--000_mask.png
 |  |  |  |     :
 |  |  |--contamination
 |  |  |  |--000_mask.png
 |  |  |  |     :
 |  |--test
 |  |  |--broken_large
 |  |  |  |--000.png
 |  |  |  |     :
 |  |  |--broken_small
 |  |  |  |--000.png
 |  |  |  |     :
 |  |  |--contamination
 |  |  |  |--000.png
 |  |  |  |     :
 |  |  |--good
 |  |  |  |--000.png
 |  |  |  |     :
 |  |--train
 |  |  |--good
 |  |  |  |--000.png
 |  |  |  |     :
 |  |--license.txt
 |  |--reaadme.txt
 |--cable
 |    :
 |--license.txt
 |--mvtec_anomaly_detection.tar
 |--readme.txt

Docker file

This dockerfile contains the necessary modules to run this code. Firstly, please build a docker image with Dockerfile, and then run a docker container.

cd docker_file
docker build -t [REPOSITORY]:[TAG] .   Build docker image
docker run --runtime=nvidia -it --rm -p [port number]:8888 -v [local directory path]:/root --name [container name] --ipc=host [REPOSITORY]:[TAG]

If you'd like to use Jupyter lab, you execute set_jupyter_lab.sh after ran a docker container, please.

chmod 777 set_jupyter_lab.sh
./set_jupyter_lab.sh

Requirements

Pytorch: over 1.7.x
scikit-image
sklearn
tqdm
PIL