Tic-Tac-Toe Symbol Classifier (CNN)

A Deep Learning project built with PyTorch that uses a Convolutional Neural Network (CNN) to recognize and classify Tic-Tac-Toe board symbols from images.

---

📌 Project Overview

This project classifies 128x128 grayscale images into three categories:

• Blank: An empty square.
• Circle: Representing the 'O' player.
• Nought: Representing the 'X' player.

The custom CNN uses Batch Normalization and Dropout to improve performance and reduce overfitting.

---

🚀 Features

• Custom CNN Architecture: 3 convolutional layers + 3 fully connected layers.
• Data Augmentation: Random vertical flips and rotations to make the model robust.
• High Performance: Achieved ~98% training accuracy and ~94% test accuracy.
• GPU Support: Runs on CUDA-enabled devices for faster training.

---

🛠️ Tech Stack

• Framework: PyTorch
• Data Handling: Torchvision, DataLoader
• Evaluation: Scikit-learn (Confusion Matrix, Classification Report)
• Language: Python

---

📂 Project Structure

.
├── images/
│   ├── train/          # 480 images (Blank, Circle, Nought)
│   └── test/           # 120 images (Blank, Circle, Nought)
├── Tic_Tac_Toe_CNN_v1.ipynb  # Main notebook with training & evaluation
└── README.md

---

🏗️ Model Architecture

CustomCNN consists of:

1. Conv Layer 1: 1 → 16 channels, BatchNorm, ReLU, MaxPool
2. Conv Layer 2: 16 → 32 channels, BatchNorm, ReLU, MaxPool
3. Conv Layer 3: 32 → 64 channels, ReLU, MaxPool
4. Fully Connected Layers: 256 → 128 → 3 units (Output)
5. Regularization: Dropout (0.5)

---

📈 Results

Trained for 15 epochs using AdamW optimizer:

Metric	Training	Testing
Accuracy	98.3%	93.3%
Loss	0.0503	-

Classification Report (Test Set)

              precision    recall  f1-score   support
       blank       1.00      1.00      1.00        40
      circle       0.85      1.00      0.92        40
       nought      1.00      0.82      0.90        40

---

🔄 Data Augmentation

• Random Vertical Flip: Helps the model recognize symbols even if images are flipped vertically.
• Random Rotation (±20°): Makes the network robust to slight rotations in the Tic-Tac-Toe board.

These augmentations reduce overfitting and improve generalization to unseen images.

---

⚙️ How to Run

1. Clone the repo:

git clone https://github.com/your-username/tic-tac-toe-cnn.git
cd tic-tac-toe-cnn

2. Install Dependencies:

pip install torch torchvision numpy scikit-learn

3. Update Paths: Edit the ImageFolder paths in the notebook to point to your local dataset directory.

4. Run the Notebook: Open Tic_Tac_Toe_CNN_v1.ipynb in Jupyter or VS Code and run all cells.

---

📄 License

This project is open-source and available under the MIT License.