ArinRoths/2026-24679-image-autogluon-predictor-snacks-hw2pt2

Sweet vs. Savory Image Classifier

Model Description

This model classifies food images as sweet or savory. I trained it using AutoGluon MultiModal on a classmate's Homework 1 image dataset.

Dataset and Task

Dataset: https://huggingface.co/datasets/kwongnon/2026-24679-image-dataset

The task is binary image classification. The dataset has 385 training images, 7 validation images, and 8 test images.

AutoML Search

I ran a random search with an ASHA scheduler, 4 trials, and a fixed 20-minute time budget. The search space included ResNet-18, ResNet-34, and ResNet-50, with learning rates of 0.0001 and 0.0005.

I set the maximum epochs to 10 and early-stopping patience to 2. The four sampled trials tested ResNet-18 and ResNet-34 at both learning rates. ResNet-50 was included in the search space but was not sampled.

Selected Model and Results

AutoGluon selected ResNet-18 with a learning rate of 0.0005, maximum epochs of 10, and early-stopping patience of 2. The selected trial completed 6 training iterations.

Validation accuracy: 85.71% Test accuracy: 75.00%

Limitations

The validation and test sets are very small, so the results may not represent performance on a larger variety of food images. The AutoML search was also limited to four trials and did not test every architecture in the search space.

AI use disclosure

ChatGPT was used to help write this metacard based on my code and also helped to understand how to add additional architecture types and how to format the code properly for this part and also how to do early stopping