jullyz / memae-1

Memorizing Normality to Detect Anomaly: Memory-augmented Deep Autoencoder for Unsupervised Anomaly Detection (MemAE)

Memorizing Normality to Detect Anomaly: Memory-augmented Deep Autoencoder for Unsupervised Anomaly Detection (MemAE)

** This repo. is not official and not perfectly implemented. **

You can see the paper in https://arxiv.org/pdf/1904.02639.pdf.

It is partly implemented since it is a small part of my current project.

Requirements

• Python 3.6.4
• Pytorch >= 1.0.0
• tensorboardX >= 1.6
• tqdm >= 1.6

How to use

Train

To train the model,

python main.py \
  --train
  --num-istances [INT] \
  --num-classes [INT] \
  --num-memories [INT] \
  --addressing ['soft', 'sparse'] \
  --log-dir=$DIR

You can see other configuration in 'main.py'.

Visualize

python main.py --visualize --ckpt [TRAINED_MODEL]

Test

python main.py --test --ckpt [TRAINED_MODEL]