ANOGAN

AnoGAN(anomaly GAN) GANを用いた異常検知

Unsupervised Anomaly Detection with Generative Adversarial Networks to Guide Marker Discovery https://arxiv.org/abs/1703.05921

result

generate image

code

以下コードの説明です

main.py

usage: main.py [-h] [--datapath DATAPATH] [--epoch EPOCH]
               [--batchsize BATCHSIZE] [--mode MODE] [--imgsize IMGSIZE]
               [--channels CHANNELS] [--zdims ZDIMS] [--testpath TESTPATH]
               [--label_idx LABEL_IDX] [--img_idx IMG_IDX]

train AnoGAN

optional arguments:
  -h, --help            show this help message and exit
  --datapath DATAPATH, -d DATAPATH
  --epoch EPOCH, -e EPOCH
  --batchsize BATCHSIZE, -b BATCHSIZE
  --mode MODE, -m MODE  train, test
  --imgsize IMGSIZE
  --channels CHANNELS
  --zdims ZDIMS
  --testpath TESTPATH, -p TESTPATH
  --label_idx LABEL_IDX
  --img_idx IMG_IDX

dcgan.py

DCGANを構築し，学習を行うクラス mainから呼び出し

load.py

MNISTや画像データ・CSVデータを読み込むスクリプト mainから呼び出し

model.py

mainから呼び出し

model

Dsicriminator

Generator

Requirement

Software
- python3.6.3
- tensorflow==1.7.0
- keras==2.1.5
- numpy==1.14.0
- matplotlib==2.2.2
- opencv-python==3.4.1.15

reference

https://github.com/tkwoo/anogan-keras https://www.renom.jp/ja/notebooks/tutorial/generative-model/anoGAN/notebook.html

shinGangan / ANOGAN