Fashion-MNIST Machine Learning Project (93.6% accuracy)

Description

This project explores three machine learning techniques to classify the Fashion-MNIST dataset, achieving an accuracy of 93.6% on unseen data. The techniques explored are Multilayer Perceptrons (MLP), Convolutional Neural Networks (CNN), and Visual Transformers (ViT). For more information on the optimal architectures we found and a discussion of our results, check out our report.

How to Run the Project

To clone the repository, type:

git clone https://github.com/leonpetrinos/Fashion-MNIST.git

The dataset is provided in the root directory as a zip file. Unzip before doing any other command. Navigate to the /project directory and here are some sample commands to execute:

MLP

python main.py --data ../dataset --nn_type mlp --lr 2e-4 --max_iters 30

CNN

python main.py --data ../dataset --nn_type cnn --lr 1e-3 --max_iters 200

ViT

python main.py --data ../dataset --nn_type transformer --lr 2e-4 --max_iters 5

Additional Arguments

• --device <device_name>: For CNN and MLP, specify the device (cpu by default, mps for MacOS).
• --use_pca: Allows the use of Principal Component Analysis (dimensionality reduction technique) and only works for the MLP.
• --pca_d <number_of_principal_components>: Number of principal components for PCA.
• --nn_batch_size <batch_size>: Modify the batch size for stochastic gradient descent.
• --test: Train on the whole training set and output predictions on the test set. Without this argument, the model trains on 80% of the data (20% for validation) and prints validation accuracies and F1 scores.

Requirements

Install the necessary packages using:

pip install torch torchinfo matplotlib numpy

If case of errors even when all the packages are installed, run:

pip install opencv-python==4.6.0.66

Collaborators

Name	GitHub
Leon Petrinos	leonpetrinos
Andrea Trugenberger	AndreaTrugenberger
Youssef Chelaifa	chelaifa