public override object ForwardPropagation(object inputs)
{
float[,,] inputValues = (float[, , ])inputs;
float[,,] outputs = new float[
NeuralNetworkAPI.CalculateOutputLength(inputValues.GetLength(0), Height, StrideY, PaddingType),
NeuralNetworkAPI.CalculateOutputLength(inputValues.GetLength(1), Width, StrideX, PaddingType),
Kernels.Length
];
#if NET40_OR_GREATER
Parallel.For(0, Kernels.Length, i => {
ForwardPropagation(i);
});
#else
for (int i = 0; i < Kernels.Length; i++)
{
ForwardPropagation(i);
}
#endif
return(Outputs = outputs);
void ForwardPropagation(int i)
{
Kernels[i].ForwardPropagation(inputValues, ref outputs, i, ActivationFunctionType);
}
}
public void Load(INeuralNetworkLoader loader, JsonTextWriter textWriter = null)
{
if (loader is null)
{
throw new ArgumentNullException(nameof(loader));
}
JsonObject modelConfig = loader.GetModelConfig();
if (textWriter != null)
{
textWriter.BeginObject();
textWriter.Write(NeuralNetworkKeyword.ModelConfig);
new JsonTextBuilder(new JsonValueReader(modelConfig), textWriter).BuildObject();
}
IsSequential = modelConfig["class_name"] == "Sequential";
NeuralNetworkLayer previousLayer = null;
Layers.Clear();
foreach (JsonObject jsonLayer in modelConfig["config"]["layers"].Array)
{
string className = jsonLayer["class_name"];
JsonObject config = jsonLayer["config"];
string name = config["name"];
JsonValue activation = config["activation"];
JsonValue dataFormat = config["data_format"];
JsonValue padding = config["padding"];
NeuralNetworkLayer layer;
JsonArray inboundNodes = jsonLayer["inbound_nodes"][0];
foreach (JsonValue inboundNode in inboundNodes)
{
string targetName = inboundNode[0];
foreach (NeuralNetworkLayer currentLayer in Layers)
{
if (currentLayer.Name == targetName)
{
previousLayer = currentLayer;
break;
}
}
}
switch (className)
{
case "InputLayer": {
JsonArray batchInputShape = config["batch_input_shape"];
batchInputShape.RemoveAt(0);
List <int> inputShapes = JsonConvert.Deserialize <List <int> > (batchInputShape);
AddLayer(new NeuralNetworkInputLayer());
layer.InputShape = inputShapes.ToArray();
layer.OutputShape = layer.InputShape;
break;
}
case nameof(NeuralNetworkLayerType.Dense): {
int units = config["units"];
float[,] weights = loader.GetDenseWeights(name, units, previousLayer.OutputShape[0]);
float[] biases = loader.GetBiases(name, units);
if (textWriter != null)
{
WriteLayer(textWriter, name, weights, biases);
}
NeuralNetworkNeuron[] neurons = new NeuralNetworkNeuron[units];
for (int i = 0; i < neurons.Length; i++)
{
neurons[i] = new NeuralNetworkNeuron(previousLayer.OutputShape[0], biases[i]);
for (int n = 0; n < previousLayer.OutputShape[0]; n++)
{
neurons[i].Weights[n] = weights[n, i];
}
}
int[] outputShape = new int[] { units };
AddLayer(new NeuralNetworkDenseLayer(neurons, NeuralNetworkAPI.GetActivationFunctionType(activation)));
layer.OutputShape = outputShape;
break;
}
case nameof(NeuralNetworkLayerType.Conv2D): {
JsonArray kernelSize = config["kernel_size"];
JsonArray strides = config["strides"];
int units = config["filters"];
int width = kernelSize[1];
int height = kernelSize[0];
int channel = previousLayer.OutputShape[2];
int strideX = strides[1];
int strideY = strides[0];
float[,,,] weights = loader.GetConv2DWeights(name, width, height, channel, units);
float[] biases = loader.GetBiases(name, units);
if (textWriter != null)
{
WriteLayer(textWriter, name, weights, biases);
}
NeuralNetworkPaddingType paddingType = NeuralNetworkAPI.GetPaddingType(padding);
NeuralNetworkKernel[] kernels = new NeuralNetworkKernel[units];
for (int i = 0; i < kernels.Length; i++)
{
kernels[i] = new NeuralNetworkKernel(width, height, channel, strideX, strideY, biases[i], paddingType);
for (int c = 0; c < channel; c++)
{
for (int y = 0; y < height; y++)
{
for (int x = 0; x < width; x++)
{
kernels[i].Weights[y, x, c] = weights[y, x, c, i];
}
}
}
}
int[] outputShape = new int[] {
NeuralNetworkAPI.CalculateOutputLength(previousLayer.OutputShape[0], height, strideY, paddingType),
NeuralNetworkAPI.CalculateOutputLength(previousLayer.OutputShape[1], width, strideX, paddingType),
units
};
AddLayer(new NeuralNetworkConv2DLayer(
kernels,
width,
height,
strideX,
strideY,
NeuralNetworkAPI.GetActivationFunctionType(activation), paddingType)
);
layer.OutputShape = outputShape;
break;
}
case nameof(NeuralNetworkLayerType.MaxPooling2D): {
JsonArray poolSize = config["pool_size"];
JsonArray strides = config["strides"];
int width = poolSize[1];
int height = poolSize[0];
int strideX = strides[1];
int strideY = strides[0];
NeuralNetworkPaddingType paddingType = NeuralNetworkAPI.GetPaddingType(padding);
int[] outputShape = new int[] {
NeuralNetworkAPI.CalculateOutputLength(previousLayer.OutputShape[0], height, strideY, paddingType),
NeuralNetworkAPI.CalculateOutputLength(previousLayer.OutputShape[1], width, strideX, paddingType),
previousLayer.OutputShape[2]
};
AddLayer(new NeuralNetworkMaxPooling2DLayer(width, height, strideX, strideY, paddingType));
layer.OutputShape = outputShape;
break;
}
case nameof(NeuralNetworkLayerType.Dropout): {
layer = new NeuralNetworkDropoutLayer();
layer.OutputShape = previousLayer.OutputShape;
break;
}
case nameof(NeuralNetworkLayerType.Flatten): {
NeuralNetworkDataFormatType dataFormatType = NeuralNetworkAPI.GetDataFormatType(dataFormat);
int sum = 1;
NeuralNetworkLayer lastLayer = Layers[Layers.Count - 1];
foreach (int length in lastLayer.OutputShape)
{
sum *= length;
}
int[] outputShape = new int[] { sum };
AddLayer(new NeuralNetworkFlattenLayer(dataFormatType));
layer.OutputShape = outputShape;
break;
}
case nameof(NeuralNetworkLayerType.Add): {
AddLayer(new NeuralNetworkAddLayer());
layer.OutputShape = previousLayer.OutputShape;
break;
}
case nameof(NeuralNetworkLayerType.BatchNormalization): {
AddLayer(new NeuralNetworkBatchNormalizationLayer(
loader.GetValues(name, "beta", previousLayer.OutputShape[2]),
loader.GetValues(name, "gamma", previousLayer.OutputShape[2]),
loader.GetValues(name, "moving_mean", previousLayer.OutputShape[2]),
loader.GetValues(name, "moving_variance", previousLayer.OutputShape[2]),
config["epsilon"]
));
layer.OutputShape = previousLayer.OutputShape;
break;
}
case nameof(NeuralNetworkLayerType.LeakyReLU): {
AddLayer(new NeuralNetworkLeakyReLULayer(config["alpha"]));
layer.OutputShape = previousLayer.OutputShape;
break;
}
case nameof(NeuralNetworkLayerType.ZeroPadding2D): {
int paddingX = config["padding"][1][0];
int paddingY = config["padding"][0][0];
AddLayer(new NeuralNetworkZeroPadding2DLayer(paddingX, paddingY));
int[] outputShape = new int[] {
previousLayer.OutputShape[0] + paddingY * 2,
previousLayer.OutputShape[1] + paddingX * 2,
previousLayer.OutputShape[2]
};
layer.OutputShape = outputShape;
break;
}
case nameof(NeuralNetworkLayerType.UpSampling2D): {
int sizeX = config["size"][1];
int sizeY = config["size"][0];
AddLayer(new NeuralNetworkUpSampling2DLayer(sizeX, sizeY));
int[] outputShape = new int[] { previousLayer.OutputShape[0] * sizeY, previousLayer.OutputShape[1] * sizeX, previousLayer.OutputShape[2] };
layer.OutputShape = outputShape;
break;
}
case nameof(NeuralNetworkLayerType.Concatenate): {
AddLayer(new NeuralNetworkConcatenateLayer());
int channels = 0;
foreach (NeuralNetworkLayer currentLayer in layer.Previous)
{
channels += currentLayer.OutputShape[2];
}
int[] outputShape = new int[] { previousLayer.OutputShape[0], previousLayer.OutputShape[1], channels };
layer.OutputShape = outputShape;
break;
}
default:
throw new NotImplementedException(className);
}
void AddLayer(NeuralNetworkLayer newLayer)
{
newLayer.Name = name;
if (IsSequential)
{
if (Layers.Count > 0)
{
newLayer.InputShape = previousLayer.OutputShape;
previousLayer.Nexts.Add(newLayer);
newLayer.Previous = new NeuralNetworkLayer[] { previousLayer };
}
previousLayer = newLayer;
}
else
{
List <NeuralNetworkLayer> previous = new List <NeuralNetworkLayer> ();
foreach (JsonValue inboundNode in inboundNodes)
{
string targetName = inboundNode[0];
foreach (NeuralNetworkLayer currentLayer in Layers)
{
if (currentLayer.Name == targetName)
{
currentLayer.Nexts.Add(newLayer);
previous.Add(currentLayer);
previousLayer = currentLayer;
}
}
}
if (previous.Count > 0)
{
newLayer.Previous = previous.ToArray();
newLayer.InputShape = newLayer.Previous[0].OutputShape;
}
}
Layers.Add(newLayer);
layer = newLayer;
}
}
if (textWriter != null)
{
textWriter.EndObject();
}
if (IsSequential)
{
OutputLayers = new NeuralNetworkLayer[] { previousLayer };
}
else
{
JsonArray outputLayers = modelConfig["config"]["output_layers"];
OutputLayers = new NeuralNetworkLayer[outputLayers.Count];
for (int i = 0; i < outputLayers.Count; i++)
{
string targetName = outputLayers[i][0];
OutputLayers[i] = Layers.Find(layer => layer.Name == targetName);
}
}
}
public override object ForwardPropagation(object inputs)
{
float[,,] inputValues = (float[, , ])inputs;
int inputWidth = inputValues.GetLength(1);
int inputHeight = inputValues.GetLength(0);
int inputChannel = inputValues.GetLength(2);
int outputWidth = NeuralNetworkAPI.CalculateOutputLength(inputValues.GetLength(1), Width, StrideX, PaddingType);
int outputHeight = NeuralNetworkAPI.CalculateOutputLength(inputValues.GetLength(0), Height, StrideY, PaddingType);
int paddingX;
int paddingY;
switch (PaddingType)
{
case NeuralNetworkPaddingType.Valid:
paddingX = 0;
paddingY = 0;
break;
case NeuralNetworkPaddingType.Same:
paddingX = (Width - 1) / 2;
paddingY = (Height - 1) / 2;
break;
default:
throw new NotImplementedException(PaddingType.ToString());
}
int startX = 0 - paddingX;
int startY = 0 - paddingY;
int endX = inputWidth + paddingX - Width;
int endY = inputHeight + paddingY - Height;
float[,,] outputs = new float[outputHeight, outputWidth, inputChannel];
#if NET40_OR_GREATER
Parallel.For(0, inputChannel, c => {
int outputX = 0;
int outputY = 0;
for (int y = startY; y <= endY; y += StrideY, outputY++)
{
for (int x = startX; x <= endX; x += StrideX, outputX++)
{
bool first = true;
float max = 0;
int ex = x + Width;
int eY = y + Height;
for (int cy = y; cy < eY; cy++)
{
for (int cx = x; cx < ex; cx++)
{
float value;
if (cy < 0 || cy >= endY || cx < 0 || cx >= endX)
{
value = 0;
}
else
{
value = inputValues[cy, cx, c];
}
if (first || max < value)
{
first = false;
max = value;
}
}
}
outputs[outputY, outputX, c] = max;
}
outputX = 0;
}
});
#else
for (int c = 0; c < inputChannel; c++)
{
int outputX = 0;
int outputY = 0;
for (int y = startY; y <= endY; y += StrideY, outputY++)
{
for (int x = startX; x <= endX; x += StrideX, outputX++)
{
bool first = true;
float max = 0;
int ex = x + Width;
int eY = y + Height;
for (int cy = y; cy < eY; cy++)
{
for (int cx = x; cx < ex; cx++)
{
float value;
if (cy < 0 || cy >= endY || cx < 0 || cx >= endX)
{
value = 0;
}
else
{
value = inputValues[cy, cx, c];
}
if (first || max < value)
{
first = false;
max = value;
}
}
}
outputs[outputY, outputX, c] = max;
}
outputX = 0;
}
}
#endif
return(Outputs = outputs);
}