CNN-Ship-Classifier

A full end-to-end image classification pipeline written entirely in Perl.

---

---

Overview

This repository contains a complete implementation of a Convolutional Neural Network (CNN) in Perl, built to classify satellite/aerial images of ships vs. non-ships.

The entire workflow —dataset loading, preprocessing, tensor conversion, CNN architecture, training loop, learning-rate scheduling, early stopping, metrics, and visualizations— is implemented using modules from the Perl ecosystem:

• AI::MXNet (deep learning)
• PDL (tensors / numerical computing)
• Imager (image loading)
• Chart::Plotly (visualization)
• IPerl (Jupyter kernel for Perl)

Plus a set of reusable custom modules:

• ImageFolderDataset.pm
• ImageFolderDatasetSubset.pm
• DataLoader.pm
• ConfusionMatrixPlot.pm

This project serves as a practical demonstration of modern machine learning in Perl — no Python required.

---

Features

✔ Mini-VGG style CNN architecture
✔ SoftmaxCrossEntropy loss
✔ Adam optimizer
✔ Learning Rate Scheduler
✔ Early Stopping
✔ Batch-level F1-score
✔ Macro F1 per epoch
✔ Confusion Matrix visualization
✔ Plotly training curves
✔ Full interactive IPerl notebook

---

Installation

Requirements

• Perl 5.30+
• PDL
• AI::MXNet >= 1.6 (from source)
• Imager
• Chart::Plotly
• IPerl (optional, for the notebook)

Install IPerl

cpanm IPerl
iperl notebook