The code is written in Python 3.10, the required packages are given in requirements.txt.
To run the code clone the repository to the current directory of your machine:
git clone https:/jana370/DeepSAD
It is recommended to set up a virtual environment to run the code:
# pip install virtualenv
cd <path of DeepSAD directory>
python -m virtualenv env
.\env\Scripts\activate
Then the required packages can be installed:
pip install -r requirements.txt
After that the code can be run.
DeepSAD.py is the Deep SAD implementation,
make_graphics.py is the code used for creating the figure for the tests including labeled normal data,
make_graphic_pollution.py is the code used for creating the figure for the tests with mislabeled data.
python DeepSAD.py
python make_graphis.py
python make_graphic_pollution.py
For the Deep SAD implementation different options in the command line can be used:
-d or --dataset: choose the dataset which will be used; either "mnist", "fmnist", or "cifar10" can be used;
default is "mnist"
-m or --mode: choose the type of loss function, which will be used for Deep SAD;
"standard" will treat labeled normal data the same as unlabeled data and use the weight only for
labeled anomalies;
"standard_normal" will use the weight for both labeled normal data and labeled anomalies;
"extended" will use the weight for the labeled normal data and the second weight for the labeled
anomalies;
default is "standard"
-w or --weight: choose the weight that will be used in the loss function; Note, that this only defines the weight
for the labeled normal data if the "extended" mode is used;
default is 3
-sw or --second_weight: choose the second weight that will be used for the labeled anomalies if the "extended"
mode is used;
default is 4
-cn or --category_normal: choose category which will be used as the normal class, the following categories
are defined for each dataset:
MNIST: 0: 0, 6, 8, and 9; 1: 1, 4, and 7; 2: 2, 3, and 5;
F-MNIST: 0: T_shirt, Pullover, Coat, and Shirt; 1: Trouser, and Dress;
2: Sandal, Sneaker, Bag, and Ankleboot;
CIFAR-10: 0: plane, car, ship, and truck; 1: bird, and frog;
2: cat, deer, dog, and horse;
default is 0
-ca or --category_anomaly: choose category which will be used as the anomaly class, the following categories are
defined for each dataset:
MNIST: 0: 0, 6, 8, and 9; 1: 1, 4, and 7; 2: 2, 3, and 5;
F-MNIST: 0: T_shirt, Pullover, Coat, and Shirt; 1: Trouser, and Dress;
2: Sandal, Sneaker, Bag, and Ankleboot;
CIFAR-10: 0: plane, car, ship, and truck; 1: bird, and frog;
2: cat, deer, dog, and horse;
default is 1
-ra or --ratio_anomaly: choose the ratio of labeled anomalies that will be used; Note, that the value should
be between 0 and 1;
default is 0.05
-rn or --ratio_normal: choose the ratio of labeled normal data that will be used; Note, that the value should be
between 0 and 1;
default is 0.0
-rpu or --ratio_pollution_unlabeled: choose the ratio of pollution in the unlabeled data; Note, that the value
should be between 0 and 1;
default is 0.1
-rpl or --ratio_pollution_labeled: choose the ratio of pollution in the labeled anomalies; Note, that the value
should be between 0 and 1;
default is 0.0
This means, that Deep SAD using MNIST, the "standard" mode, with the weight 3, the 0 category as normal class, the 1 category as anomaly class, a labeled
anomaly ratio of 0.05, no labeled normal data, a pollution of 0.1 in the unlabeled data, and no pollution in the labeled anomalies, can be run by using:
python DeepSAD.py
Deep SAD using CIFAR-10, the "extended" mode, with the weight 2 for labeled normal data, and a second weight 4 for labeled anomalies,
the 1 category as normal class, the 2 category as anomaly class, a labeled anomaly ratio of 0.01, a labeled normal data ratio of 0.1,
a pollution of 0.1 in the unlabeled data, and a pollution of 0.01 in the labeled anomalies, can for example be run by using:
python DeepSAD.py -d "cifar10" -m "extended" -w 2 -sw 4 -cn 1 -ca 2 -ra 0.01 -rn 0.1 -rpu 0.1 -rpl 0.01