Minecraft Biome Classifier 🌍

Convolutional Neural Network (CNN) model trained to classify 29 different Minecraft biomes from images.

🧠 Overview

This project uses a fine-tuned ResNet-18 model trained with PyTorch to identify Minecraft biomes such as forests, deserts, and oceans.
Developed as part of a hackathon challenge where participants built machine learning models to recognize biomes from in-game screenshots.

• Achieved 0.8454 private leaderboard accuracy
• Implemented transfer learning with frozen early ResNet layers
• Used data augmentation: horizontal flips and random rotations
• Experimented with multiple models — CNN performed best

🧩 Model

• Architecture: ResNet-18 (transfer learning)
• Framework: PyTorch / Torchvision
• Loss Function: Cross-Entropy
• Optimizer: Adam (lr = 1e-4)
• Early Stopping: delta = 1e-4
• Augmentations: Resize, Normalize, Random Horizontal Flip, Random Rotation

📂 Folder Structure

• assets/ – Hackathon logo and visuals
• models/ – Jupyter notebook & trained model (.pth)
• output/ – Sample and generated prediction CSVs
• userkits/ – Custom dataset & utils for PyTorch

Note: Other ML notebooks (rf.ipynb, ridge.ipynb, xgboost.ipynb) are excluded from the main workflow.

⚙️ Installation/Usage

1. Clone the repository:

git clone https://github.com/MokeyCodes/minecraft-biome-classifier.git
cd minecraft-biome-classifier

2. Install Dependencies:

pip install -r requirements.txt

3. Run the Jupyter notebook

jupyter notebook models/pytorch.ipynb

📊 Results

Metric	Score
Public Leaderboard	0.8058
Private Leaderboard	0.8454

🙌 Acknowledgements

Done during the Blockography AI Hackathon hosted by ACM AI at University of California, San Diego.