Deep Learning

Intel Image Classifier

GPU-accelerated transfer-learning pipeline with Optuna tuning and W&B logging.

2025

OVERVIEW

The Project

A high-performance image-classification pipeline on Intel's scene dataset (6 terrain classes: buildings, forest, glacier, mountain, sea, street). It leverages PyTorch Lightning, transfer learning, and experiment tracking with Weights & Biases for scalable, reproducible training.

Objectives

Classify scene images into 6 terrain classes via transfer learning.

Optimize backbone architecture and learning rate with Optuna.

Log and visualize training metrics; deploy a modular, scalable pipeline.

Tools & Technologies

PyTorch LightningOptunaWeights & BiasesResNet18 / EfficientNet-B0

METHODOLOGY

The Approach

Download the Intel dataset via Kaggle API; organize into train/test/predict.

Build a Lightning DataModule to abstract loading.

Automate model tuning with Optuna across trials; log to W&B.

Select the best model, retrain, and evaluate on unseen data.

OUTCOME

Results & Learnings

Best model: EfficientNet-B0 with ~92% test accuracy after Optuna tuning.

Dashboards provided actionable visualizations of performance trends.

Key Learnings

Optuna-tuned hyperparameters lifted accuracy in just a few trials.
Good infrastructure (Lightning + W&B) makes training reproducible and simple.

Related Case Studies