The Swift machine learning library.

@KacheFlowe @joobei Hey guys, sorry for my apparent lack of communication over the last little while. New startup, crazy hours, no sleep... you know the routine! Definitely haven't left Swift AI behind though, and in fact it's finally made its way back up my priority list.

It's been some time since we've discussed Recurrent Neural Networks - any progress or thoughts from you guys? I've decided that it's the next component I want to tackle, partly because it would be particularly useful for some side projects that I have in mind. I definitely don't want to duplicate any effort that people have already made, if any, so I'd love to hear what you guys think about the idea.

I get an error saying: Module file was created by an older version of the compiler; rebuild 'APKit' and try again

when I try to run the iOS example project.

Hi. Since Github eliminated private messaging I find it difficult to comunicate with developers. I'm just wondering if there's a feature I can tackle or code I can take off yoru plate.

I am looking to implement an RNN in swift and I'm wondering if there's already some effort underway to add an RNN to Swift-AI. So that there is no duplication of effort. Is there some RNN effort by somebody?

Thanks in advance.

Can someone explain me these lines: let x = (-500 + (Float(index) * 1000) / Float(self.numPoints)) / 100 try! self.network.update(inputs: [x]) let answer = self.sineFunc(x) print("\(x):\(answer)") try! self.network.backpropagate(answer: [answer]) I don't get where the AI calculates the points here.

I believe there was a discussion about it should be possible to use the framework just by dragging in a single file (currently FFNN.swift). Right now, FFNN conforming to Storage is preventing this from happening. What's the issue with moving this snippet into Storage.swift?

Added Storage functionalities so we can load a NN from a file. I've not used the NSCoding protocol to avoid the @objc directive. ActivationFunction now is of type String, we could also use Int

Someone pointed out this code base to me recently, and it's very cool stuff. While poking around I found things like this:

https://github.com/collinhundley/Swift-AI/blob/master/Source/Vector.swift#L14

While I also use a straight-up swift array to back my own publicly-available vectors at https://github.com/liscio/SMUGMath-Swift, I have recently found that it's not a great idea in practice—Arrays aren't necessarily guaranteed to be contiguous in memory. Instead I would recommend that you use ContiguousArray<Float> for the flat definition, and also surround your Accelerate accesses using withUnsafe{,Mutable}Pointer as I do in the code here:

https://github.com/liscio/SMUGMath-Swift/blob/master/SMUGMath/RealVectorOperations.swift

I don't yet have concrete plans to put my new code into the repo above because it is still in progress and tightly coupled with something else I'm working on.

Anyway—really cool repo, and subject matter!

Hi! I've developed a similar library, except focused on customization instead of performance. The only feature I miss from my implementation is the ability to specify an arbitrary number of hidden layers, and it would be nice to see it in Swift-AI.

The way I implemented it, the designated initializer would take an [Int] for the hidden parameter, each Int being the size of that hidden layer. The rest of the interface stays pretty much the same.

P.s. why are all member accesses qualified with self.? This is Swift man :)

"Swift AI is a high-performance deep learning library " --> but where is the code? Is it this only library https://github.com/Swift-AI/NeuralNet ?

I accept conversion (no choice) but then 2 error messages: 1.

public func toString(decimalPlaces: Int) -> String {
    let power = pow(10.0, Double(decimalPlaces))
    return "\(round(power * self) / power)"
}

--> cannot use mutating member on immutable value "self" is immutable ...

and

let providerRef = CGDataProvider(data: Data(bytes: UnsafePointer(&data), count: data.count * sizeof(PixelData)))

--> ambiguous use of 'init'

Hi Team,

Is it possible to classify the images using this library? like categories the images or detecting the object in image?

Need assistance on this.

Hello, I am creating the iOS Mobile application to detect snoring, I need to apply Mfcc and Spectrogram to the audio file and need to pass these two in the Tensorflow lite I need to apply Mfcc [Mel-frequency cepstral coefficients] and Spectrogram to the audio file. Can you suggest me any solution for this?

Thanks in Advance!

Tried to use a simple Neural Network without an activation layer at the output. The output needs to be a continuous number (example price). The API requires an activation function (linear), but the back-propagation is not implemented.

This probably isn't an issue for this SDK, I just added this as a "discussion"...

I've been messing around with getting Neural Networks to play games, for example, but I want to take it a step further: Would it be possible to use something like "Mendel" (an iOS SDK for Genetic Algorithms) to evolve a Neural Network using this SDK?

• Tanmay Bakshi (TajyMany)