Birds, Beets, Battlestar Galactica

Welcome to the project site for Adina & James’ bird classifier submission. This gitbook serves as documentation for our participation in the Spring 2023 CSE 455 bird classification competition hosted on Kaggle.

Summary

Problem

Train a model that can classify 555 types of birds with high accuracy, given a relatively small train set.

Dataset

We were provided a training dataset consisting of 30 thousand bird images divided into 555 separate classes. We were provided names for each class and a .csv file mapping the training images to their classes.

We were also provided a test dataset consisting of 10 thousand images. We were not provided labels for the test dataset.

Code

We were provided starter code by our professor Joseph Redmon. You can see our complete code here. The changes we made will be highlighted in the experiment pages.

Techniques

We trained two versions of RESNET as our model. RESNET comes pre-trained on IMAGENET but we fine tuned the model for classifying birds. Our process consisted of essentially of trial and error, tweaking the hyperparameters and image transforms on the training set. Once we stopped seeing an improvement in RESNET18, we switched to RESNET34.