In Section 3.4 I studied the accuracy of
- Generalized Learning Vector Quantization (GLVQ),
- Generalized Matrix Learning Vector Quantization (GMLVQ), and
- Generalized Tangent Learning Vector Quantization (GTLVQ)
with different numbers of prototypes on the following datasets:
- Circle;
- Spiral;
- MNIST;
- Cifar10;
- Indian Pine.
The objective of these experiments was to study the interpretability and accuracy performances and to compare the methods. The thesis can be found here: will be updated as soon as the thesis published.
The experiments are based on the internal anysma package. The provided
snippet of the package is mostly undocumented. To get an idea how it works
and how the package can be used to create LVQ models that can be trained and
evaluated on a GPU see the available files in the ./examples directory.
To run the scripts you have to install anysma. Clone the script and go to the directory by calling:
git clone https://github.com/saralajew/thesis_GTLVQ_experiments.git
cd thesis_GTLVQ_experiments
Before you install the package it is recommended to create a virtual environment via pipenv or to use a docker container. Since I have not tested the package with other versions of the required package I made the version requirements strict. Important notes related to this are:
- use Python 3.6;
- if you wanna use Tensorflow-GPU, install CUDA Version 9.0. and CUDNN 7.1.4;
Internally, we use anysma with different Tensorflow 1 versions without any problems. Therefore, feel free to try other Tensorflow 1 versions. In this case, you can use other Python and CUDA versions.
If the above requirements are met install the package by:
pip install -e .
Now, you can start a script from the ./experiments directory by simpling
executing:
python ./experiments/<path_to_the_script>
See the experiments folder for further descriptions related to the scripts.