Neural network

A neural network for classifying the images from the Fashion-MNIST dataset with over 90% accuracy. Written from scratch using C++ without any ML libraries.

Features

• Customizable feed-forward architecture
• Training using mini-batch gradient descent and backpropagation
• Softmax output layer and categorical cross entropy loss
• Glorot weight initialization
• Input standardization
• Dropout and L1 and L2 regularization for better generalization
• Adam optimizer
• Early stopping based on the loss on the validation set
• Parallelization using OpenMP
• Final accuracy over 90%

Usage

Build the project

mkdir build
cd build
cmake ..
make

Train the model

You can run the training using the network binary. The network learns the weights from the train data and predicts the labels of the test data.

./network ../data  # provide path to the directory containing the data sets

Evaluate the accuracy

You can check the accuracy by using the included evaluator.

python3 ../evaluator/evaluate.py test_predictions.csv ../data/fashion_mnist_test_labels.csv