Swift DataBunch - how to derive _tensorHandle indices

SwiftAI
1

Can anyone explain the DataBunch used to automatically batch examples, from nb03 of SwiftAI repo? Here’s the relevant code. In particular, can anyone explain this line, and how the indices were derived:

self.yb = Labels.init(_handles: _handles[xEnd..<_handles.endIndex])

The full code snippet can be seen here, for reference:

public struct DataBatch<Inputs: Differentiable & TensorGroup, Labels: TensorGroup>: TensorGroup {
public var xb: Inputs
public var yb: Labels

public init(xb: Inputs, yb: Labels){ (self.xb,self.yb) = (xb,yb) }

public var _tensorHandles: [_AnyTensorHandle] {
    xb._tensorHandles + yb._tensorHandles
}

public init<C: RandomAccessCollection>(_handles: C) where C.Element: _AnyTensorHandle {
    let xStart = _handles.startIndex
    let xEnd = _handles.index(
        xStart, offsetBy: Int(Inputs._tensorHandleCount))
    self.xb = Inputs.init(_handles: _handles[xStart..<xEnd])
    self.yb = Labels.init(_handles: _handles[xEnd..<_handles.endIndex])
}

}

2

Those are lines of code needed for TensorGroup, I don’t understand them myself. They will disappear in the next release so I wouldn’t worry too much about them :wink:

3

Merci beaucoup