Digit Recognition CNN from Scratch

A handwritten digit recognition system built from scratch using Python. This project demonstrates high-performance optimization by integrating C extensions, Fortran, and OpenBLAS with a custom Python CNN implementation.

Project Notebook

🎯 Key Features

• From-scratch implementation using NumPy and SciPy.
• Performance optimizations:
  • C extensions for core functions (Softmax, ReLU, Loss, Weight/Bias updates).
  • Fortran f2py extension for vector operations.
  • OpenBLAS integration for accelerated matrix multiplication.
• Custom compiled modules loaded via ctypes.
• Complete training pipeline included in lab.ipynb.

🚀 Getting Started

Prerequisites

• Python 3.8+
• NumPy
• SciPy
• Jupyter Notebook

Installation

1. Clone the repository:

   git clone https://github.com/77AXEL/Digit-Recognizer
   cd Digit-Recognizer

2. Extract the dataset:

   unzip data.zip

   Note: This step is required before training.

3. Install dependencies:

   pip install numpy scipy jupyter

4. (Optional) Rebuild binaries: If you need to recompile for your specific environment:

   cd lib
   make

   Requires gfortran and g++ (MinGW on Windows).

Usage

Open and run the main notebook to train and test the model:

jupyter notebook lab.ipynb

🔧 Technical Details

This project explores low-level neural network optimization. Instead of relying on frameworks like PyTorch or TensorFlow, it implements the forward and backward passes manually.

• Architecture: Standard CNN with alternating Convolutional/Pooling layers and Fully Connected layers.
• Optimization: Computationally intensive tasks (loops, dense math) are offloaded to compiled C/Fortran code to remove Python interpreter bottlenecks.

📝 License

This project is licensed under the MIT License.