qoncept/tensorswift

A lightweight library to calculate tensors in Swift, which has similar APIs to TensorFlow's

TensorSwift

TensorSwift is a lightweight library to calculate tensors, which has similar APIs to TensorFlow's. TensorSwift is useful to simulate calculating tensors in Swift using models trained by TensorFlow.

let a = Tensor(shape: [2, 3], elements: [1, 2, 3, 4, 5, 6])
let b = Tensor(shape: [2, 3], elements: [7, 8, 9, 10, 11, 12])
let sum = a + b // Tensor(shape: [2, 3], elements: [8, 10, 12, 14, 16, 18])
let mul = a * b // Tensor(shape: [2, 3], elements: [7, 16, 27, 40, 55, 72])

let c = Tensor(shape: [3, 1], elements: [7, 8, 9])
let matmul = a.matmul(c) // Tensor(shape: [2, 1], elements: [50, 122])

let zeros = Tensor(shape: [2, 3, 4])
let ones = Tensor(shape: [2, 3, 4], element: 1)

Deep MNIST for Experts

The following code shows how to simulate Deep MNIST for Experts, a tutorial of TensorFlow, by TensorSwift.

public struct Classifier {
    public let W_conv1: Tensor
    public let b_conv1: Tensor
    public let W_conv2: Tensor
    public let b_conv2: Tensor
    public let W_fc1: Tensor
    public let b_fc1: Tensor
    public let W_fc2: Tensor
    public let b_fc2: Tensor
    
    public func classify(_ x_image: Tensor) -> Int {
        let h_conv1 = (x_image.conv2d(filter: W_conv1, strides: [1, 1, 1]) + b_conv1).relu()
        let h_pool1 = h_conv1.maxPool(kernelSize: [2, 2, 1], strides: [2, 2, 1])
        
        let h_conv2 = (h_pool1.conv2d(filter: W_conv2, strides: [1, 1, 1]) + b_conv2).relu()
        let h_pool2 = h_conv2.maxPool(kernelSize: [2, 2, 1], strides: [2, 2, 1])
        
        let h_pool2_flat = h_pool2.reshaped([1, Dimension(7 * 7 * 64)])
        let h_fc1 = (h_pool2_flat.matmul(W_fc1) + b_fc1).relu()
        
        let y_conv = (h_fc1.matmul(W_fc2) + b_fc2).softmax()

        return y_conv.elements.enumerated().max { $0.1 < $1.1 }!.0
    }
}

Installation

Swift Package Manager

.Package(url: "[email protected]:qoncept/TensorSwift.git", from: "0.2.0"),

CocoaPods

pod 'TensorSwift', '~> 0.2'

Carthage

github "qoncept/TensorSwift" ~> 0.2

License

The MIT License

tensorswift's Issues

How to export weights from Python?

This is a pretty awesome project. I've trained some models on Tensorflow and want to export to the format you use loadFloatArray with. Any chance you can post some example Python code to export those weights?

Thanks for building this btw!

fatal error: `elements.count` must be greater than or equal to `shape.volume`: elements.count = 1024, shape.volume = 3136:

I have used my own images to trained by Tensorflow's DeepMnist example code.
And I used the way you give in the other issue to output the Models.
But when I replace your Models with mine and run, I meet the error:

fatal error: elements.count must be greater than or equal to shape.volume: elements.count = 1024, shape.volume = 3136:

I think maybe there is something wrong with my models.
But the DeepMnist training is successful:

And the output of models is successful, too.

What can I do to solve this problem? In order to replace the models successfully, should I do something with the Classifier Struct?

swift 3 - migration problems / tensorflow fyi

fyi - I'm getting close to having tensorflow via c_api in swift working.
could use some help though. I'm a bit of a newb when it comes to tensors.
checkout my branch -
https://github.com/johndpope/tensorflow/tree/swift/tensorflow/swift