Pokemon Card Neural Network Classifier

A Java-based neural network implementation for classifying Pokemon card images. This project implements a custom neural network architecture with support for multi-threaded training and data augmentation.

DISCLAIMER

This project was my first attempt ever using Neural Networks. Be kind please

Features

• Custom neural network implementation with configurable layers
• Multi-threaded batch training for improved performance
• Data augmentation capabilities for training data
• Support for serialization and deserialization of trained models
• Progress tracking and accuracy reporting during training
• Image preprocessing and resizing utilities

Project Structure

• Main.java - Main entry point and training loop implementation
• NeuralNetwork.java - Core neural network implementation
• Layer.java - Neural network layer implementation
• DataLoader.java - Data loading and batch management
• DataScraper.java - Data collection utilities
• PokemonScraper.java - Pokemon-specific data collection
• DataAugment.py - Python script for data augmentation
• ImageResizer.java - Image preprocessing utilities

Requirements

• Java Runtime Environment (JRE)
• Python (for data augmentation)
• Pokemon card image dataset

Usage

1. Prepare your dataset:

   • Place Pokemon card images in the Data/Pokemon/ResizedCardImages/CEL directory
   • Place corresponding labels in Data/Pokemon/CardNames/CEL.txt

2. Configure training parameters in Main.java:

   public static float learnRate = 0.5f;
   public static float dataSetCoverage = 1f;
   public static int epochNumbers = 1000;
   public static int batchSize = 50;
   public static int threadNumber = 4;

3. Run the training:

   javac *.java
   java Main

Model Architecture

The default network architecture consists of:

• Input layer: 1104 nodes
• Hidden layer 1: 786 nodes
• Hidden layer 2: 400 nodes
• Output layer: 50 nodes

Training Process

1. The network loads or initializes a new model
2. Training data is loaded and preprocessed
3. Training proceeds in epochs with multi-threaded batch processing
4. After each epoch:
   • Model is saved
   • Test accuracy is reported
   • Progress is displayed

Data Augmentation

The project includes a Python script (DataAugment.py) for augmenting the training dataset with various transformations to improve model robustness.

Model Persistence

Trained models are saved with the .ser extension and can be loaded for continued training or inference.