GithubHelp home page GithubHelp logo

blupblupblup / near-ood-doppler-signatures Goto Github PK

View Code? Open in Web Editor NEW
7.0 1.0 3.0 5.03 MB

Companion repository of the "Near OOD detection for low-resolution radar micro-Doppler signatures" paper

License: MIT License

Jupyter Notebook 11.97% Python 88.03%
anomaly-detection deep-learning machine-learning out-of-distribution-detection radar semi-supervised-learning

near-ood-doppler-signatures's Introduction

Table of contents

General description

This is the github repository for the ECML PKDD 2022 paper "Near out-of-distribution detection for low-resolution radar micro-Doppler signatures". An expanded version is available as preprint on arXiv, the conference proceedings edited by Springer Nature being unavailable at the moment. The paper presents AD experiments on simulated micro-Doppler radar signatures of helicopter-like targets. This code enables you to launch similar experiments to the ones presented in the paper, and more generally to conduct unsupervised and semi-supervised AD (as defined in Lukas Ruff paper) experiments with deep and non-deep approaches.

If you found this code useful, please consider citing the paper:

@article{bauw2022near,
  title={Near out-of-distribution detection for low-resolution radar micro-Doppler signatures},
  author={Bauw, Martin and Velasco-Forero, Santiago and Angulo, Jesus and Adnet, Claude and Airiau, Olivier},
  journal={arXiv preprint arXiv:2205.07869},
  year={2022}
}

Code organisation

  • utils.py and training.py contain various functions used for loading data and in the main model training scripts deep_svdd.py, deep_msvdd.py, deep_rpo.py, shallow_pca.py and shallow_tpca.py.

  • networks.py contain several neural networks architectures considered for our paper, only one ("net0") was kept to produce the paper results. The latter led the best results during preliminary comparisons.

  • The AD approached themselves are coded in deep_XXX.py and shallow_XXX.py scripts, which receive experiments hyperparameters from the launchers_XXX.py scripts and run experiments accordingly for the number of random seeds chosen.

  • data/ directory in which to put the data files you can download here. This data was generated using a simple MATLAB simulation code that can be found on this other repository. You can easily modify this simulation code to generate a similar dataset with more classes, or increase the intraclass diversity. The data files you should download are loaded and preprocessed using the simulated_radar_dataset class in utils.py.

How to launch an experiment

  • A simple example launch script is proposed in launcher_example.py, the experiments of the published paper can be reproduced using the launcher_paper_XXX.py scripts.

  • Launchers are divided into several scripts in order to ease the distribution of computations over several GPUs and computing instances. Shallow baselines should be computed on a CPU machine with numerous cores, the shallow AD methods being CPU-bound.

  • Executing launcher_paper_deep_experiments.py takes a little more than two days on a P100 GPU, but one could easily divide this launcher into two or more launcher scripts (for example start the bimodal normality experiments in a dedicated script) to train everything faster in several independent cluster jobs. This is made easy by the storage of results in a .csv file for every launch script: one only needs to gather the lines of all .csv files created by the different jobs in a new .csv file to read all results at once using the dedicated jupyter notebook (see explanations below).

How to find and read the results of an experiment

The results of every experiment (one experiment being one method trained on any number of seeds) are available in the results/ directory. This repository already includes the results directory of the experiments launched using launcher_example.py. Inside results/, experiments results are organized according to the input data format, the AD method used and the launch time.

The figures generated for each seed of a deep AD experiment include training metrics (loss, AUCs, AD score distribution, see example below) and 2D projections of the data points before and after training with PCA and tSNE (2D latent representations are available for Deep SVDD and Deep RPO adaptations). The example results directories do not contain the .pt files with the untrained and trained model weights to avoid putting large files on github, but will be generated if you launch experiments on your machine. The performances of experiments for every seed are stored in a .csv in comparison_graphs/ and in the results folders inside results/. In comparison_graphs/, a jupyter notebook is available to read the .csv created by experiments, and enables the easy computation of the mean performances when methods run on several seeds.

metrics

latent2D

The title of a group of figures indicates various information about the experiment setup: type of loss, SAD ratio (how many SAD labelled anomalies are available during training), etc. Several centroids (black stars) can be seen on latent 2D representations because the current version of the code plots the normal samples centroid, the SSL samples centroid, and the SAD samples centroid as soon as the required samples are available in the training set (i.e. even when the samples are not concentrated around their dedicated centroid during training).

Micro-Doppler dataset

This micro-Doppler signatures dataset, to be placed in the data/ directory, was generated thanks to a MATLAB simulation, the code of which can be found in this other repository. Four classes are used to create normal and anomalous modes, each class being defined by a helicopter-like target with a specific number of rotating blades, the latter being responsible for specific spectrum modulation patterns.

Inter-class diversity: the number of blades changes across classes, changing the modulation pattern

four classes

Intra-class diversity: diversity example with preprocessed samples from class four, the one with the most blades, i.e. with the most complex modulation pattern. One can notice the varying modulation width, stemming from different values of blade length & RPMs, as well as the varying central Doppler shift. The lateral background is set to zero in order to make the neural network input format closer to MNIST and Fashion-MNIST samples, since our network architecture is inspired by this paper. The idea of creating MNIST-like benchmarks has been of interest in different scientific communities such as biomedical images and 3D modeling.

Potential follow-ups to the paper experiments

Easy follow-ups to our experiments include:

  • conducting the same experiments on a similar dataset where samples have a lower Doppler resolution, e.g. use 32 Doppler bins instead of 64 (which is a relevant radar experiment, since some pulse Doppler radars have few pulses to characterize targets, cf. our paper);
  • conducting the same experiments on a similar dataset where samples have higher intra-class diversity, which could be generated thanks to a stronger noise in the samples generating process as well as wider uniform distributions from which to pick the targets parameters like the blade size or blades RPMs;
  • adding one or more AD methods to the comparison proposed in our paper (such as deep AD using an auto-encoder, where the reconstruction error serves as the AD score, as was already done here or here, or use an auto-encoder to replace the PCA for dimensionality reduction before using shallow AD);
  • consider several pollution ratios to check whether all Deep AD approaches suffer from it similarly.

Requirements

Create a virtual environment and use the following pip command, it should at least work with Python 3.7 and 3.8:

pip3 install seaborn scikit-learn numpy geomstats tqdm matplotlib scipy pandas torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu113

License

MIT

near-ood-doppler-signatures's People

Contributors

blupblupblup avatar

Stargazers

 avatar  avatar  avatar  avatar  avatar  avatar  avatar

Watchers

 avatar

Recommend Projects

  • React photo React

    A declarative, efficient, and flexible JavaScript library for building user interfaces.

  • Vue.js photo Vue.js

    ๐Ÿ–– Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.

  • Typescript photo Typescript

    TypeScript is a superset of JavaScript that compiles to clean JavaScript output.

  • TensorFlow photo TensorFlow

    An Open Source Machine Learning Framework for Everyone

  • Django photo Django

    The Web framework for perfectionists with deadlines.

  • D3 photo D3

    Bring data to life with SVG, Canvas and HTML. ๐Ÿ“Š๐Ÿ“ˆ๐ŸŽ‰

Recommend Topics

  • javascript

    JavaScript (JS) is a lightweight interpreted programming language with first-class functions.

  • web

    Some thing interesting about web. New door for the world.

  • server

    A server is a program made to process requests and deliver data to clients.

  • Machine learning

    Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.

  • Game

    Some thing interesting about game, make everyone happy.

Recommend Org

  • Facebook photo Facebook

    We are working to build community through open source technology. NB: members must have two-factor auth.

  • Microsoft photo Microsoft

    Open source projects and samples from Microsoft.

  • Google photo Google

    Google โค๏ธ Open Source for everyone.

  • D3 photo D3

    Data-Driven Documents codes.