Feed Forward Neural Network from scratch

This project has been implemented using two ways.

• withEigen: Using Eigen library Eigen
• withoutEigen: This part is based on vector library in std namespace std::vector

Available activation function choices.

1. SoftMax
2. ReLU
3. Tanh
4. Sigmoid
5. Linear

Available loss function choices.

1. CrossEntropy
2. MeanSquare
3. FocalLoss
4. FocalLoss_b

Available optimizer choices.

1. Standard gradient descent.
2. Standard gradient descent with momentum.

Please look at the documentation for further details.

Installation.

    mkdir build
    cd build
    cmake -DINSTALL_PREFIX=/path/to/install/FFNeuralNetwork/ /path/to/FFNeuralNetwork/source
    make
    make install (optional)

In case, build is required for the version with the eigen integeration. Add the following flag.

    mkdir build
    cd build
    cmake -DWITHOUTEIGEN=OFF . ..
    make

If you also want to compile and run the tests, also add this -DTESTS=ON. This option needs internet access, because the process will download the googletest library.

Getting started.

Please refer to some of examples provided along this package. These examples illustrate a way on how to create an architect for the neural network, and train it on a dataset. You might need to download the iris dataset from University of California Irvine dataset repository to see these examples play. Please refer to the documentation for further information.

Build the documentation.

• Install doxygen
• To build the documentation.

    mkdir build
    cd build
    cmake ..
    make docs