Dissertation Learning in Non-Stationary Environments - Code and Thesis

This repository contains the source code and the final pdf for the dissertation 'Learning in Non-Stationary Environments'. Standard language is python and mentioned otherwise. The files are usually self-explaining for example, demo or train algorithms, utils, plot methods etc. The folder structure roughly represents the chapters of the thesis:

├── sda                     # Code for chapter 3 & 4: Geometric and Subspace Domain Adaptation    
│   ├── gda.py              # Python Demo for Geometric Domain Adaptation
│   ├── so.py               # Python Demo for Subspace Override Algorithm
│   ├── nso.py              # Python Demo for Nyström SO
│   ├── matlab              # Matlab Code 
│   │   ├── study_all.m     # Reproduces all performance experiments (takes a while!)   
│   │   ├── study_<name>.m  # Reproduces performance experiments on dataset <name>   
├── dsda                    # Code for chapter 5: Deep Spectral Domain Adaptation
│   ├── train_asan.py       # Demo for Adversarial Spectral Adaptation Network
│   ├── train_dsn.py        # Demo for Deep Spectral Network
│   ├── study_asan.py       # Reproduces ASAN performance experiments on selected dataset (takes a while!)   
│   ├── study_dsn.py        # Reproduces DSN performance experiments on selected dataset (takes a while!)   
├── rrslvq                  # Code for chapter 6: Non-Stationary Online Prototype Learning
│   ├── demo.py             # Demo for Reactive Robust Soft Learning Vector Quantization (RRSLVQ)
│   ├── study               # Folder of study scripts
└── └── └── study_<type>.m  # Reproduces experiments of type <type>   
└── thesis_learning_in_non_stationary_enviroments.pdf       # Dissertation as Pdf  

Installation

Use the requirements.txt and pip_requirements.txt to install dependencies. The recommended enviroment is conda in combindation with pip.

git clone https://github.com/ChristophRaab/thesis.git 
cd thesis
conda create -n thesis python=3.8
conda activate thesis
conda install --file requirements.txt
pip install -r pip_requirements.txt # If conda requirements fail

Execution

All python scripts can be started from top-level directory or in the file directory.
All matlab scripts must be started in the file directory.

Datasets

All datasets are included except the Visda dataset required for reproducing the results in the appendix. The Visda dataset can be found here.
To use the Visda datasets copy the extracted folders train, validation, test to:

    ├── dsda                    # Code for chapter 4: Deep Spectral Domain Adaptation
    │   ├── dataset             # Domain Adaptation dataset

Licenses

Code MIT
Thesis CC BY-SA 4.0