Latent Outlier Exposure for Anomaly Detection with Contaminated Data (LOE)

The experiments here are an attempt to reproduce the results from https://arxiv.org/abs/2202.08088. and stresstest the NTL https://github.com/boschresearch/NeuTraL-AD. Model. For the original code go to to: https://github.com/boschresearch/LatentOE-AD. This repo contains altered code from the original implementation, with stress-tests on robustness and lowering the amount of data the model is trained on. Results are only obtained on the fmnist / cifar10 dataset. This Repository is not maintained!

The full work can be found on https://github.com/buehlpa/LatentOE-FUAD/blob/main/Specialization_Project2_buehlpa1_MSE_DS.pdf

Comments on the CODE

most of the changes to the original repo are in /Additional_CODE_BULE

/loe_utils # contains the helperfunctions for the saving of pickle files in the original LOE implementation

/ plots # plots of the results

/training_scripts # contains bash script to run multiple runs of the LOE blind etc of ntl

/utils # contains helper functions and model for the Added code from the project work

evaluating_models.ipynb # compares different models from project work and BOSCH LOE inspecting_dataset.ipynb # inspecting the features with dimensionality reduction method and rescaling of the features

for LOE

run : python Launch_Exps.py --config-file config_fmnist.yml --contamination 0.0 --assumed-contamination 0.0 --dataset-name fmnist --trainset_fraction

if you want different loss functions etc. change the config file

• it saves the contaminated results in a different folder, to concatenate them run the function:

concatenate_allresults(MODEL_RESULT_PATH:str,modelname:str='loe_hard',assumed_contamination:float=0.0,n_runs:int=5) from LatentOE-AD/Additional_Code_BULE/loe_utils/helperfunctions.py

for the multi-run bash scripts run:

bash Additional_Code_BULE/training_scripts/run_multiple.sh

Data

• downsample dataset with: Extract_img_features import downsample_dataset on the extraced features

for project work autencoder etc.

run training_scripts/ training_main_*.py

• it automatically saves the pickle files of the results in a folder under RESULTS/

each run gets an allresults_run_{}.pkl file which contains all the results for the different contamination ratios

inspect results

• evaluating_models.ipynb