nprab428 / fashion-mnist-classifier

fashion-mnist-classifier

These are the results of a fun Kaggle competition I competed in for my machine learning class. The assignment was to correctly classify images from the Fashion-MNIST dataset, which is comprised of various articles of clothing.

Methodology

We used convolution neural networks to build the classifier. While there are an abundance of ways to implement a CNN, I wanted to keep my model simple and go with some of the tried-and-true architectures we read about in class. The Stanford Computer Vision staff recommends a "CONV-RELU" layer stacking architecture, with multiple POOL layers. I decided to implement something similar, and as they stress in the literature, not reinvent the wheel.

Training the model

The training dataset contained 60,000 samples with 10 different classes. I used PyTorch to implement my architecture and then focused on fine-tuning the "learning rate" and "momentum" parameters. After some empirical testing, I found their optimal values by observing gradual minimization in the loss function over 16 epochs. See my full script in fashionMNIST.ipynb.

Results

Over the 10,000 test samples, I achieved 88% accuracy with my model. This score was high enough to put me in the top 10 of the class!

I only let my model run for 20 epochs, but if I had more time (or more computing power e.g. access to a CUDA device), I think I could have improved my model even further.