Upsample layer?

[tirinox]

Hello,
I need upsample layer to implement U-net for segmentation.
Is there one? Will it be implemented?

[bryant1410]

Hey. We're not planning to implement it in the short term.

But you can give it a try! Check out the code and maybe you run into a way to do it.

[AndreJFBico]

Im still testing it but im using the following with MPSCNNUpsamplingNearest.

import Foundation
import MetalPerformanceShaders

open class Upscale: NetworkLayer {
    
    /// Used to determine the filename for this layers weights. (Ignored if there is no ParameterLoader)
    static var weightModifier: String = ""
    
    let size: ConvSize
    private var prevSize: LayerSize!
    
    var upscale: MPSCNNUpsamplingNearest!
    
    public init(size: ConvSize, id: String? = nil) {
        self.size = size
        super.init(id: id)
    }
    
    open override func initialize(network: Network, device: MTLDevice) {
        super.initialize(network: network, device: device)
        let incoming = getIncoming()
        prevSize = incoming[0].outputSize
        outputSize = LayerSize(f: prevSize.f,
                               w: size.strideX * prevSize.w,
                               h: size.strideY * prevSize.h)

        upscale = MPSCNNUpsamplingNearest(device: device, integerScaleFactorX: size.strideX, integerScaleFactorY: size.strideY)
        outputImage = MPSImage(device: device, imageDescriptor: MPSImageDescriptor(layerSize: outputSize))
    }
    
    open override func updatedCheckpoint(device: MTLDevice) {

    }

    open override func execute(commandBuffer: MTLCommandBuffer) {
        let incoming = getIncoming()
        upscale?.encode(commandBuffer: commandBuffer, sourceImage: incoming[0].outputImage, destinationImage: outputImage)
    }
}

[bryant1410]

Hey @AndreJFBico, you can send a PR with that if you want to have it in Bender!