public MnistFullLayerNeuralNetwork(IMTLCommandQueue commandQueueIn)
{
// CommandQueue to be kept around
commandQueue = commandQueueIn;
device = commandQueueIn.Device;
// Initialize MPSImage from descriptors
SrcImage = new MPSImage(device, SID);
dstImage = new MPSImage(device, DID);
// setup convolution layer (which is a fully-connected layer)
// cliprect, offset is automatically set
layer = SlimMPSCnnFullyConnected.Create(kernelWidth: 28, kernelHeight: 28,
inputFeatureChannels: 1, outputFeatureChannels: 10,
neuronFilter: null, device: device,
kernelParamsBinaryName: "NN");
// prepare softmax layer to be applied at the end to get a clear label
softmax = new MPSCnnSoftMax(device);
}
public MnistFullLayerNeuralNetwork (IMTLCommandQueue commandQueueIn)
{
// CommandQueue to be kept around
commandQueue = commandQueueIn;
device = commandQueueIn.Device;
// Initialize MPSImage from descriptors
SrcImage = new MPSImage (device, SID);
dstImage = new MPSImage (device, DID);
// setup convolution layer (which is a fully-connected layer)
// cliprect, offset is automatically set
layer = SlimMPSCnnFullyConnected.Create (kernelWidth: 28, kernelHeight: 28,
inputFeatureChannels: 1, outputFeatureChannels: 10,
neuronFilter: null, device: device,
kernelParamsBinaryName: "NN");
// prepare softmax layer to be applied at the end to get a clear label
softmax = new MPSCnnSoftMax (device);
}
public MnistDeepConvNeuralNetwork(IMTLCommandQueue commandQueueIn)
: base(commandQueueIn)
{
// use device for a little while to initialize
var device = commandQueueIn.Device;
pool = new MPSCnnPoolingMax(device, 2, 2, 2, 2)
{
Offset = new MPSOffset {
X = 1, Y = 1, Z = 0
},
EdgeMode = MPSImageEdgeMode.Clamp
};
relu = new MPSCnnNeuronReLU(device, 0);
// Initialize MPSImage from descriptors
c1Image = new MPSImage(device, c1id);
p1Image = new MPSImage(device, p1id);
c2Image = new MPSImage(device, c2id);
p2Image = new MPSImage(device, p2id);
fc1Image = new MPSImage(device, fc1id);
// setup convolution layers
conv1 = SlimMPSCnnConvolution.Create(kernelWidth: 5,
kernelHeight: 5,
inputFeatureChannels: 1,
outputFeatureChannels: 32,
neuronFilter: relu,
device: device,
kernelParamsBinaryName: "conv1",
padding: true,
strideX: 1,
strideY: 1,
destinationFeatureChannelOffset: 0,
groupNum: 1);
conv2 = SlimMPSCnnConvolution.Create(kernelWidth: 5,
kernelHeight: 5,
inputFeatureChannels: 32,
outputFeatureChannels: 64,
neuronFilter: relu,
device: device,
kernelParamsBinaryName: "conv2",
padding: true,
strideX: 1,
strideY: 1,
destinationFeatureChannelOffset: 0,
groupNum: 1);
// same as a 1x1 convolution filter to produce 1x1x10 from 1x1x1024
fc1 = SlimMPSCnnFullyConnected.Create(kernelWidth: 7,
kernelHeight: 7,
inputFeatureChannels: 64,
outputFeatureChannels: 1024,
neuronFilter: null,
device: device,
kernelParamsBinaryName: "fc1",
destinationFeatureChannelOffset: 0);
fc2 = SlimMPSCnnFullyConnected.Create(kernelWidth: 1,
kernelHeight: 1,
inputFeatureChannels: 1024,
outputFeatureChannels: 10,
neuronFilter: null,
device: device,
kernelParamsBinaryName: "fc2");
}
