• About the Project
• Prerequisites
• Installation
• Usage
• Results
• Authors
• Acknowledgments

KerSmoSVM deals with classification tasks through the usage of Support Vector Machines (SVMs).
In particular the SVM is implemented with kernel functions, and the fitting algorithm called Sequential Minimal Optimization by John Platt that optimizes the Lagrangian which arises from the maximization of the margin in SVMs.
The project proposes the usage of such classifier to distinguish between state of ease and state of motor imagery activity from a personal dataset.
This particular implementation takes in account also the usage of Gram Matrices to speed up computation, and it avoids calls to the kernel function.

To get a local copy up and running follow these simple steps.

The project provides a Pipfile file that can be managed with pipenv.
pipenv installation is strongly encouraged in order to avoid dependency/reproducibility problems.

• pipenv

pip install pipenv

1. Clone the repo

git clone https://github.com/CrisSherban/KerSmoSVM

2. Enter in the project directory and install Python dependencies

cd KerSmoSVM
pipenv install

Here's a brief description of the .py files:

• kernel.py: Class that represents a kernel, thus provides a kernel function
• svm.py: SVM class, provides fit() and predict() methods.
• main.py: Verifies that the SVM class works, loads the dataset and shows the result with various kernels.
• test_svm.py: Contains a simple unit test on the SVM class by fitting and predicting over the MNIST dataset.
• tools.dataset_tools.py: Provides the loading, preprocessing, and splitting of the personal EEG dataset.
• tools.validation_tools.py: Provides functions to predict using several SVMs, and cross validate.

To run the main.py script just run from the project directory:

python main.py

Along fitting several SVMs and show results, it will also show how a sample of the dataset looks like after bandpass and standardization.

As we can appreciate, the RBF kernel in this situation with default parameters is the one that performs best:

• Serban Cristian Tudosie

CDMO-2 © Course held by Professor Vittorio Maniezzo