Base class for neural network layers.
SDKs
- iOS 10.0+
- macOS 10.13+
- Mac Catalyst 13.0+
- tvOS 10.0+
Framework
- Metal Performance Shaders
Declaration
class MPSCNNKernel : MPSKernel
Overview
An MPSCNNKernel object consumes one MPSImage object and produces one MPSImage object.
The region overwritten in the destination image is described by the clip property. The top left corner of the region consumed (ignoring adjustments for filter size—for example, convolution filter size) is given by the offset property. The size of the region consumed is a function of the size of the clip property and any subsampling caused by pixel strides at work (for example, strideInPixelsX/strideInPixelsY in the MPSCNNPooling class). Wherever the offset and clip properties would cause an {x,y} pixel address not in the image to be read, the edge property is used to determine what value to read there.
The z or depth component of the offset, origin and size properties indexes which images to use.
If the
MPSImageobject contains only a single image, then these values should beoffset,.z = 0 clip, andRect .origin .z = 0 clip.Rect .size .depth = 1 If the
MPSImageobject contains multiple images, then the value ofclipdetermines the number of images to process. Both the source and destinationRect .size .depth MPSImageobjects must have at least this many images. The value ofoffsetrefers to the starting image index of the source. Thus, the value of.z offsetmust be.z + clip Rect .size .depth <=source. Similarly, the value of.number Of Images clipdetermines the starting image index of the destination. Thus, the value ofRect .origin .z clipmust beRect .origin .z + clip Rect .size .depth <=destination..number Of Images
The destination property can be used to control where the kernel will start writing in terms of feature channel dimension. For example, if the destination has 64 channels and thdestinatione kernel outputs 32 channels, channels 0-31 of the destination will be populated by the kernel (by default). But if you want the kernel to populate channels 32-63 of the destination, you can set the value of destination to 32. Suppose you have a source of dimensions w x h x Ni, where N is the number of channels, which goes through a convolution filter C0 which produces the output O0 = w x h x N0 and C1 which produces the output O1 = w x h x N1 followed by concatenation which produces O = w x h x (N0 + N1). You can achieve this by creating an MPSImage object with dimensions w x h x (N0 + N1) and using this as the destination of both convolutions as follows:
C0: destination, this will outputFeature Channel Offset = 0 N0channels starting at channel0of destination thus populating[0,N0-1]channels.C1: destination, this will outputFeature Channel Offset = N0 N1channels starting at channelN0of destination thus populating[N0,N0+N1-1]channels.