Deep learning with automatic differentiation and backpropogation in Go.
Check main.go for example usage.
This is not a practical thing to use, just a project to experiment with machine learning in go. As such, this isn't structured like a library and I really recommend you don't use this in anything that you wish to continue working. This almost certainly contains a good amount of bugs.
calc/ contains the implementation of n-dimensional matrixes and the math operations on them. It implements a few of the heavier operations using the blas implementation from gonum.
dataset/ contains the MNIST dataset and a utility for loading it into a calc.NDArray.
model/ contains some keras-like utilities for building a tensor graph out of a few simple layer types, handling feeding input and updating the trainable parameters.
tensor/ contains implementations of various types of tensors, and an implementation of backwards automatic differentiation, used in backpropogation (check the gradientVisitor.VisitX implementation for each tensor to see the tensors generated for the gradients).
main.go contains a simple CNN model for classifying MNIST digits that can get to ~90% accuracy.