miniMNIST-c

miniMNIST-c

This project implements a minimal neural network in C for classifying handwritten digits from the MNIST dataset. The entire implementation is ~200 lines of code and uses only the standard C library.

Features

• Two-layer neural network (input → hidden → output)
• ReLU activation function for the hidden layer
• Softmax activation function for the output layer
• Cross-entropy loss function
• Stochastic Gradient Descent (SGD) optimizer

Performance

…

Prerequisites

• GCC compiler
• MNIST dataset files:
  • train-images.idx3-ubyte
  • train-labels.idx1-ubyte

Compilation

gcc -O3 -march=native -ffast-math -o nn nn.c -lm

Usage

1. Place the MNIST dataset files in the data/ directory.

2. Compile the program.

3. Run the executable:

   bash复制

   ./nn

The program will train the neural network on the MNIST dataset and output the accuracy and average loss for each epoch.

Configuration

You can adjust the following parameters in nn.c:

• HIDDEN_SIZE: Number of neurons in the hidden layer
• LEARNING_RATE: Learning rate for SGD
• EPOCHS: Number of training epochs
• BATCH_SIZE: Mini-batch size for training
• TRAIN_SPLIT: Proportion of data used for training (the rest is used for testing)

License

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