popastefan10 / deepad

A PyTorch implementation of the paper https://arxiv.org/abs/1811.06861. It presents a reconstruction-based approach to anomaly detection with focus on surface defects.

DeepAD

A PyTorch implementation of "Anomaly Detection using Deep Learning based Image Completion". This paper was published at the 17th IEEE International Conference on Machine Learning and Applications (ICMLA). You can find it on arXiv.

Run locally

First step is to install the dependencies. On my machine I have used a conda environment, but the project can be run with venv or without any environment at all.

# Use only one option from below

# Install packages within a Conda environment
$ conda create -n deep-ad -f environment.yml

# Install packages within a virtual environment
$ pip install virtualenv
$ virtualenv deep-ad
$ source deep-ad/bin/activate
(deep-ad) $ pip install -r requirements.txt

# Install packages globally (not recommended)
$ pip install -r requirements.txt

In order to be able to use modules from src/deep_ad inside notebooks we need to install the project. For development purposes use --editable/-e.

python -m pip install .
# OR
python -m pip install -e .

Required packages were added to environment.yml and requirements.txt files with these commands:

conda env export --from-history | findstr /v "^prefix" > environment.yml
pip list --format=freeze > requirements.txt

The findstr function is a Windows equivalent for grep, so on Linux use this instead:

conda env export --from-history | grep -v "^prefix" > environment.yml

Environment file

The minimum required keys inside the .env are the following:

running_env=HOME
# In this directory you should have at least the raw DAGM dataset
dagm_dir="C:\\path-to-datasets-directory"