Simple Neural Network Library (v2)

This project is my attempt to create a simple neural network framework in C++ (cpp17) that is powerful enough to give decent results in MNIST dataset in decent running time.

The project is inspired by the simplicity of Keras. e.g. creating a simple network for MNIST:

    models::sequential m;
    m.add("dense(units=300, input=784)");
    m.add("dense(units=10)");
    m.compile("loss=cross_entropy(), optimizer=sgd(learning_rate=0.5)");
    m.summary();

Console Output:

See demo_mnist.cpp for full code.

---

Index

• Results
• Building Project
• Documentation
• Features

---

Results

Experiment:

• Dataset: MNIST
• Hidden Layers: 1 Dense Layer (300 neurons)
• Learning Rate: 0.5 (Decay = 0)
• Batch Size: 32

See demo_mnist.cpp for full code.

Here is a comparision of different Loss-Functions (test-accuracy)

Autoencoder Experiment:

• Dataset: MNIST
• Autoencoder (784->256->32, 32->256->784)
• Cross-entropy with Adam(learning_rate=0.001)
• Batch Size: 32

See autoencoder_mnist.cpp for full code.

Here is the reconstruction of some random examples.

---

Building Project

Building examples

git clone https://github.com/rahulsrma26/simpleNN.git
cd simpleNN
mkdir build
cd build

• Linux and MacOS (make)

  cmake ..
  make

• Windows

  Visual Studio 2017 and later can directly open a CMake project. However, if you want to build using Developer Command prompt and have cmake installed then run:

  cmake ..
  msbuild simpleNN.sln /t:Build /p:Configuration=Release

Running examples

# After build first download MNIST dataset
python ../scripts/download_mnist.py ../data/mnist
# or you can download fashion mnist
# python ../scripts/download_fashion.py ../data/fashion
# mnist example
bin/demo_mnist ../data/mnist
# spiral example
bin/spiral_test
# mnist model with save/restore
bin/mnist_model -h

For windows users who are using MSVC check '.exe' files in bin\Release folder.

Even More ways to build

You can also be build using ninja across any platform.

• Using ninja

  cmake .. -GNinja
  ninja -v

• If you want to use mingw (GCC) instead of MSVC under windows then chocolatey can be used to install dependencies for ninja build.

  choco intall -y mingw ninja cmake

  After that append cmake install path ("C:\Program Files\CMake\bin") to Environment Variable.