NNGo: Neural Network framework in Golang built from scratch

Neuralnetwork is a hands-on approach to machine learning in Go. For those of you, who have tried Tensorflow for Go or keras, it will fit you perfectly. The framework is still in development, there is still a lot to be implemented. Appreciate every feedback possible.

Inspiration Behind NNGo

The primary goal for NNGo is to be a highly performant machine learning/graph computation-based framework. It should bring the appeal of Go (simple compilation and deployment process to the ML world). There is a long way ahead of us regarding deployment, efficiency and managebility, but baby steps, right? :)

The secondary goal for NNGo is to provide a platform for exploration for non-standard deep-learning and neural network related things. Using our framework, you'll be able to expand the horizon of deep learning by exploring the highly abstract tool for extracting the most of the data as well as the algorithms.

Installation

go get -u github.com/timothy102/neuralnetwork
import nn "github.com/timothy102/neuralnetwork 

How the Tensor Package Works

Mimicking the Keras architecture, TensorGo works by implementing the unbounded interface method able to reproduce any form or value ensuring tensor scalability. This was accomplished using the reflect module in Golang. In order to initialize a tensor, you can either define a placeholder, the tensor constructor or avoid it all together by implementing the higher abstract level of the NNGo library for ML.

#1
cube := [][][]float64{}
tensor := nn.NewTensor(cube)

#2
shape := []int{2, 3, 4}
t := nn.Placeholder(shape)

Try your first NNGo Program

result := nn.Add(tensor, t)
res := tensor.Add(t)

Both solutions yield the same result 😃

Getting Started with the framework

The NNGo environement is structured similarly to Keras' layers API.

model := nn.Sequential([]Layer{
  Conv2D(64,3, 1,Valid),
  MaxPooling2D(2),
  Conv2D(32,3, 1, DefaultPadding()),
  MaxPooling2D(2),
  Flatten(),
  Dense(128, Relu),
  Dense(32,Tanh),
  Softmax(10)
}, "sequential")

model.Compile(RMSprop, CrossEntropy, Mae)

history := model.Train(dataX, dataY, numEpochs)

Contact

Please, feel free to reach out on LinkedIn, gmail. For more, check my medium article.

https://towardsdatascience.com/golang-as-the-new-machine-learning-powerforce-e1b74b10b83b

https://www.linkedin.com/in/tim-cvetko-32842a1a6/

License

Licensed under the MIT LICENSE