ib031-cifar10

This project was created as the assignment solution for the IB031 subject on FI MU in Brno, Czech Republic.

In this project, we will try to classify CIFAR-10 dataset images, using a simple classifier, by which we mean not neural networks. Our aim is to maximize accuracy. Therefore and with respect to the nature of this dataset, we will use the k-Nearest Neighbours model and Support Vector Machines.

We will try two different approaches to pre-processing, and we will evaluate the results for both models. Eventually, we will compare the models with respect to accuracy (mainly), and also the time needed for predictions and training.

Results

After preprocessing by HOG descriptors and HUE transformation, we conclude that for all models, the HOG transformation is faster and more accurate, with baseline model (Decision Tree) achieving 31% accuracy on the test set, KNN model 55% and the Bagging Ensemble with SVM achieved 59%, however slight overfitting was present.

Usage

Setup

After downloading probably you can find the final notebook in “src” directory. However, you can recreate the available notebooks from the source files using jupytext:

1. Navigate to “src” directory
2. run jupytext --to notebook [script] where script is the name of script
   You want to generate notebook from. At this time it may be one of:
   • CIFAR10.py (contains all the code, still being updated to match src/CIFAR10.ipynb
3. Then You can run it (or here is a shorthand:
   jupytext --to notebook --execute [script])
4. If You are a developer and already have the notebook computed, it (should)
   be enough to just rename it.

Making changes

You can make changes both in notebook or script. But to synchronize
them run jupyter --sync [notebook] or just make [notebook]
(Try typing “make “ and then should iterate over all notebooks)

Adding notebooks

To add new notebook to git tracking:

1. jupytext --set-formats ipynb,py [notebook]
2. Add "jupytext": { "notebook_metadata_filter": "all" } to notebook metadata
   (Edit -> Edit Notebook Metadata)