george2cool36/hw2_image_automl_autogluon

HW2 Neural AutoML — AutoGluon MultiModalPredictor (Signs)

Model Overview

This model is a supervised image classification on a classmate’s dataset (ecopus/sign_identification) using AutoGluon Multimodal. It builds a compact model under a small compute budget and report results with a clear, reproducible pipeline.

Summary

• Backbone: resnet18 (via timm)
• Input resolution: 224×224 (images resized in Colab)
• Train/Val/Test: ~64% / 16% / 20% split (stratified)
• Epochs: 3 (short budget, early-stop not overridden)
• Batch size: 8
• Metric (val): Accuracy + Macro-F1
• Result (test): Accuracy = 0.4286, Macro-F1 = 0.3

Dataset

• Source: ecopus/sign_identification
• Task: Multiclass sign recognition
• Classes: Stop, Yield, SpeedLimit, NoEntry, Crosswalk
• Preprocessing:
  • datasets → decode to PIL
  • Resize to 224×224, RGB
  • Labels normalized to integers/strings for AutoGluon

Training & AutoML Setup

Library: autogluon.multimodal.MultiModalPredictor
Problem type: multiclass
Eval metric: accuracy (Macro-F1 also reported)

AI Tool Disclosure

This notebook used ChatGPT for scaffolding code and documentation. All dataset selection, training, evaluation, and uploads were performed by the student.