Here we go.
- Usage of gonum for matrix calculations;
- Fully-connected layer without bias;
- ReLU layer;
- Pooling layer. Only 'max' supported as pooling function currently;
- Convolution layer without bias;
- Gradient descent with inertia as backpropagation function for convolution and fully-connected layers;
- Conv2D, Pool2D, ZeroPadding, Im2Col, Rotate, Reshape, ControursPadding, Flatten functions;
- Training process;
- Some simple examples for MLP, CNN;
- Activation functions and its derivatives: tanh, sigmoid, atan, softplus, gaussian;
- Graphivz for fully-connected layer;
- Pretty prints on layers;
- Import/Export JSON files.
Future and current works are here: https://github.com/LdDl/cnns/blob/master/ROADMAP.md