Skip-WGANomaly: Unsupervised Image Anomaly Detection Based in WGAN

This repository contains the framework for implementation and training of Skip-WGANomaly jointly with the implementation of Skip-GANomaly for comparison and testing purposes.

• Skip-WGANomaly
  • 1. Contents
  • 2. Requirements
  • 3. Execution
    • 3.1. Training in CIFAR-10
    • 3.2. Tuning test on CIFAR-10
    • 3.3. Train on Custom Dataset
    • 3.4. Visualization and evaluation
  • 4. Results

The file environment.txtincludes the installed packages of the environment used for the project. The model was implemented using PyTorch (v1.11, Python 3.8.10 and CUDA 11.6). Experiments of the paper were performed using an NVIDIA GeForce RTX 2080 GPU.

Example: training all cifar classes as anomaly in w_skipganomaly

# CIFAR
bash experiments/run_cifar_wskip.sh

Example: run a parametrized training for airplane as anomaly in Skip-WGANomaly

python train.py --dataset cifar10 --abnormal_class 'airplane' --niter 10 --display --w_con 50 --w_lat 1 --w_adv 1 --model "w_skipganomaly"

All arguments are detailed in options.py. For displaying all possible arguments, run:

python train.py -h

Example: run a RayTune test on CIFAR-10 with Skip-WGANomaly.

python train_tune.py --dataset cifar10 --abnormal_class 'bird' --tune True  --display --model "w_skipganomaly"

Parameters to be tested for tuning have to be edited and personalized inside train_tune.py.

To train the model on a custom dataset, the dataset should be copied into ./data directory, and should have the following directory & file structure:

Custom Dataset
├── test
│   ├── 0.normal
│   │   └── normal_tst_img_0.png
│   │   └── normal_tst_img_1.png
│   │   ...
│   │   └── normal_tst_img_n.png
│   ├── 1.abnormal
│   │   └── abnormal_tst_img_0.png
│   │   └── abnormal_tst_img_1.png
│   │   ...
│   │   └── abnormal_tst_img_m.png
├── train
│   ├── 0.normal
│   │   └── normal_tst_img_0.png
│   │   └── normal_tst_img_1.png
│   │   ...
│   │   └── normal_tst_img_t.png

Then model training is the same as the training explained above.

python train.py                     \
    --dataset <name-of-the-data>    \
    --isize <image-size>            \
    --niter <number-of-epochs>      \
    --display                       # optional if you want to visualize

In the file histogram.csv, the labels (normal, abnormal) and anomaly scores of the last experiment are saved and can be retrieved for plotting.

When --display enabled, use Visdom for metrics and images visualization. Run in a different window:

python -m visdom.server

By default, port 8097 of localhost will hold the service.

The results of Skip-WGANomaly for CIFAR-10 obtained by us, with max 15 epochs: