Swiftpack.co - wvabrinskas/Neuron as Swift Package

Swiftpack.co is a collection of thousands of indexed Swift packages. Search packages.
See all packages published by wvabrinskas.
wvabrinskas/Neuron 2.0.1
A neural network library for Swift
⭐️ 40
🕓 1 week ago
iOS macOS watchOS tvOS
.package(url: "https://github.com/wvabrinskas/Neuron.git", from: "2.0.1")



Feel free to send me suggestions on how to improve this. I would be delighted to learn more!! You can also feel free to assign issues here. Run the unit tests as well to learn how the project works!

Full Documentation

Grand Re-Opening!

Version 2.0 of Neuron is here! This new version of Neuron is a complete rewrite from the ground up of the architecture. It is much more streamlined, with faster execution. Its usage also aligns more with commonly used ML frameworks like Keras and PyTorch.


Neuron has been a pet project of mine for years now. I set off to learn the basics of ML and I figured the best way to learn it was to implement it myself. I decided on Swift because it was the language I knew the most and I knew it would be challenging to optimize for ML as it has a lot of overhead. What you're seeing here in this repository is an accumulation of my work over the past 2 years or so. It is my baby. I decided to make this open source as I wanted to share what I've learned with the ML and Swift community. I wanted to give users of this framework the opportunity to learn and implement ML in their projects or apps. Have fun!

There is still a lot missing in this framework but with this rewrite I brought a lot more flexibity to the framework to allow for playing around with different architectures and models. There are some example models provided with the framework, like Classifier, GAN, WGAN, and WGANGP. I am always working on this project and will continue to provide updates.



Generated 7's from a WGAN. Trained on MNIST 7's for 10 epochs. 16 - 32 kernels on the generator.

Contribution Policies

Filing Issues

Feel free to file issues about the framework here or contact me through the Discord. I am open to all suggestions on how to improve the framework.

Pull Requests

There are automated tests that run when a PR is created to the develop or master branches. These tests must pass before a PR can be merged. All PRs must merge into the develop branch.


All features must be branched off the develop branch.

Important Note: GPU Support (WIP)

Currently there is no GPU execution, at least not as how I would like it. Everything runs on the CPU, with some C optimizations for certain mathematical functions. Neuron will run multithreaded on the CPU with somewhat decent speed depending on the model. However a very large model with serveral kernels and convolutions will take a while. This is something I want to get working ASAP however Metal is very difficult to work with, especially with my limited knowledge and my desire to write everything from scratch.

Quick Start Guide

To get started with Neuron it all begins with setting up a Sequential object. This object is responsible for organizing the forward and backward passes through the network.

Build a Network

Let's build an MNIST classifier network. We need to build a Sequential object to handle our layers.

let network = Sequential {
    Conv2d(filterCount: 16,
            inputSize: TensorSize(array: [28,28,1]),
            padding: .same,
            initializer: initializer),
    LeakyReLu(limit: 0.2),
    Conv2d(filterCount: 32,
            padding: .same,
            initializer: initializer),
    LeakyReLu(limit: 0.2),
    Dense(64, initializer: initializer),
    LeakyReLu(limit: 0.2),
    Dense(10, initializer: initializer),

Sequential takes in one property which is block that returns an array of Layer types. init(_ layers: () -> [Layer]). The order here matters. The first layer is a Conv2d layer with 16 filters, padding .same, and an initializer. The default initializer is .heNormal.

First layer note:

You can see here the first layer is the only layer where the inputSize is specified. This is because all the other layer's inputSize are automatically calculated when added to an Optimizer.

Picking an Optimizer

Neuron uses a protocol that defines what's needed for an Opitmizer. There are currently three provided optimizers bundled with Neuron.

  • Adam
  • SGD
  • RMSProp

All optimizers are interchangeable. Optimizers are the "brain" of the network. All function calls to train the network should be called through your specific Optimizer. Let's build an Adam optimizer for this classifier.

let optim = Adam(network,
                 learningRate: 0.0001,
                 l2Normalize: false)

The first parameter here is the network we defined above. learningRate is the the size of the steps the optimizer will take when .step() is called. l2Normalize defines if the optimizer will normalize the gradients before they are applied to the weights. Default value for this is false. Adam also takes in properties for its beta1, beta2, and epsilon properties.

Set up training model

By this point you are ready to train the Optimizer and the network. You could create your own training cycle to create the classifier or you can use the built in provided Classifier model.

The Classifier model is a completely optional class that does a lot of the heavy lifting for you when training the network. Let's use that for now.

let classifier = Classifier(optimizer: optim,
                            epochs: 10,
                            batchSize: 32,
                            threadWorkers: 8,
                            log: false)

Here we create a Classifier object. We pass in the Adam Optimizer we defined earlier, the number of epochs to train for, the batch size, and the number of multi-threaded workers to use. 8 or 16 is usually a good enough number for this. This will split the batch up over multiple threads allowing for faster exectution.

Building the MNIST dataset

Next step to get the MNIST dataset. Neuron provides this locally to you through the MNIST() object.

let data = await MNIST().build()

We can use the Async/Await build function for simplicity. Datasets also support Combine publishers.

Time to train!

To train the network using the Classifier object just call classifier.fit(data.training, data.val). That's it! The Classifier will now train for the specified number of epochs and report to the MetricProvider if set.

Retrieving Metrics

All Optimizers support the addition of a MetricsReporter object. This object will track all metrics you ask it to during the initialization. If the metric isn't supported by your netowrk setup it will report a 0.

let reporter = MetricsReporter(frequency: 1,
                                metricsToGather: [.loss,

optim.metricsReporter = reporter

optim.metricsReporter?.receive = { metrics in
  let accuracy = metrics[.accuracy] ?? 0
  let loss = metrics[.loss] ?? 0
  //let valLoss = metrics[.valLoss] ?? 0
  print("training -> ", "loss: ", loss, "accuracy: ", accuracy)

The metricsToGather array is a Set of Metric definitions.

public enum Metric: String {
  case loss = "Training Loss"
  case accuracy = "Accuracy"
  case valLoss = "Validation Loss"
  case generatorLoss = "Generator Loss"
  case criticLoss = "Critic Loss"
  case gradientPenalty = "Gradient Penalty"
  case realImageLoss = "Real Image Loss"
  case fakeImageLoss = "Fake Image Loss"
  case valAccuracy = "Validation Accuracy"

MetricReporter will call receive when it is updated.

Exporting your model

Once the model has trained to your liking you can export the model to a .smodel file. This model cvan be then imported later using the Sequential intializer. The export will not export your Optimizer settings, only the Trainable specified in the Optimizer.

Neuron provides a helper object for exporting called ExportHelper. The usage is simple:

// defined: 
public static func getModel<T: Codable>(filename: String = "model", model: T) -> URL?

// usage:
ExportHelper.export(filename: "my_model", model: network)

This will return a URL for you to access your .smodel file.

Pretty-print your network

You can also print your network to the console by calling print on the Sequential object. It will pretty print your network as below:

Finishing up

Keep playing around with your new model and enjoy the network! Share your model on the Discord or ask for some other models that others have made!

Basics Background

How does Neuron work?


The main backbone of Neuron is the Tensor object. This object is basically a glorified 3D array of numbers. All Tensor objects are 3D arrays however they can contain any type of array in-between. Its size is defined by a TensorSize object defining columns, rows, depth.

public class Tensor: Equatable, Codable {
  public init() {
    self.value = []
    self.context = TensorContext()
  public init(_ data: Scalar? = nil, context: TensorContext = TensorContext()) {
    if let data = data {
      self.value = [[[data]]]
    } else {
      self.value = []
    self.context = context
  public init(_ data: [Scalar], context: TensorContext = TensorContext()) {
    self.value = [[data]]
    self.context = context
  public init(_ data: [[Scalar]], context: TensorContext = TensorContext()) {
    self.value = [data]
    self.context = context
  public init(_ data: Data, context: TensorContext = TensorContext()) {
    self.value = data
    self.context = context

Above are the initializers that Tensor supports. More in-depth documentation on Tensor can be found here.


You can perform basic arithmetic opterations directly to a Tensor object as well.

static func * (Tensor, Tensor.Scalar) -> Tensor
static func * (Tensor, Tensor) -> Tensor
static func + (Tensor, Tensor) -> Tensor
static func + (Tensor, Tensor.Scalar) -> Tensor
static func - (Tensor, Tensor.Scalar) -> Tensor
static func - (Tensor, Tensor) -> Tensor
static func / (Tensor, Tensor.Scalar) -> Tensor
static func / (Tensor, Tensor) -> Tensor
static func == (Tensor, Tensor) -> Bool

Building a backpropagation graph

You can attach a Tensor to another Tensor's graph by calling setGraph(_ tensor: Tensor) on the Tensor whose graph you'd like to set.

let inputTensor = Tensor([1,2,3,4])
var outputTensor = Tensor([2])


Doing so will set the inputTensor as the graph to the outputTensor. This means that when calling .gradients on the outputTensor the operation will look as such:

delta -> outputTensor.context(inputTensor) -> gradients w.r.t to inputTensor

Unless you're building a graph yourself or doing something custom, you'll never have to set a graph yourself. This will be handled by the Sequential object.


More in-depth TensorContext documentation can be found here.

Neuron performs gradient descent operations using Tensor objects and their accompanying TensorContext. Tensor objects contain an internal property called context which is of type TensorContext. TensorContext is an object that contains the backpropagtion information for that given Tensor. As of right now Neuron doesn't have a full auto-grad setup yet however Tensor objects with their TensorContext provides some type of auto-grad.

public struct TensorContext: Codable {
  public typealias TensorBackpropResult = (input: Tensor, weight: Tensor)
  public typealias TensorContextFunction = (_ inputs: Tensor, _ gradient: Tensor) -> TensorBackpropResult
  var backpropagate: TensorContextFunction
  public init(backpropagate: TensorContextFunction? = nil) {
    let defaultFunction = { (input: Tensor, gradient: Tensor) in
      return (Tensor(gradient.value), Tensor())
    self.backpropagate = backpropagate ?? defaultFunction
  public func encode(to encoder: Encoder) throws {}
  public init(from decoder: Decoder) throws {
    self = TensorContext()

When calling .gradients(delta: SomeTensor) on a Tensor that has an attached graph it wil automatically backpropagate all the way through the graph and return a Tensor.Gradient object.

public struct Gradient {
  let input: [Tensor]
  let weights: [Tensor]
  let biases: [Tensor]
  public init(input: [Tensor] = [],
              weights: [Tensor] = [],
              biases: [Tensor] = []) {
    self.input = input
    self.weights = weights
    self.biases = biases

A Tensor.Gradient object will contain all the gradients you'll need to perform a backpropagation step in the Optimizer. This object contains gradients w.r.t the input, w.r.t the weights, and w.r.t the biases of the graph.


Stars: 40
Last commit: 1 week ago
jonrohan Something's broken? Yell at me @ptrpavlik. Praise and feedback (and money) is also welcome.

Release Notes

1 week ago

Updated NumSwift dependency to tag 2.0.0

Swiftpack is being maintained by Petr Pavlik | @ptrpavlik | @swiftpackco | API | Analytics