public static ANNDef ReadANNFromKeras(string fnH5Weights, string fnJSONStructure)
{
ANNDef ann = new ANNDef();
Console.WriteLine("** LOAD file " + fnH5Weights);
if (!File.Exists(fnH5Weights))
{
throw new Exception("*** File not found");
}
const bool READ_ONLY = true;
long file = Hdf5.OpenFile(fnH5Weights, READ_ONLY);
if (file < 0)
{
throw new Exception("unable to find/open file " + fnH5Weights);
}
bool inputLayerIsSparse = false;
JsonDocument o = JsonDocument.Parse(File.ReadAllText(fnJSONStructure));
List <(string, ANNDef)> subnets = new();
inputLayerIsSparse = ExtractNetwork(ann, subnets, file, inputLayerIsSparse, o.RootElement, null);
if (subnets.Count > 0)
{
ann.InputSubnetworks = subnets;
}
H5G.close(file);
return(ann);
}
public ANNLayerDef(ANNDef parent, string name, int widthIn, int widthOut, Func <float[], object, float[]> customFunc)
{
Parent = parent;
Name = name;
WidthIn = widthIn;
WidthOut = widthOut;
Subtype = LayerSubtype.CustomFunc;
CustomFunc = customFunc;
}
/// <summary>
/// Constructor
/// </summary>
/// <param name="def"></param>
/// <param name="multiCount"></param>
public ANNCalcResult(ANNDef def, int multiCount)
{
if (multiCount <= 0)
{
throw new ArgumentException("multiCount must be positive");
}
Parent = def;
MultiCount = multiCount;
LayerOutputBuffersSingle = new float[def.Layers.Count][];
for (int i = 0; i < def.Layers.Count; i++)
{
LayerOutputBuffersSingle[i] = new AlignedFloatArray(def.Layers[i].WidthOut, 128).GetManagedArray();
}
}
private static bool ExtractNetwork(ANNDef ann, List <(string, ANNDef)> subnets, long file,
public ANNLayerDef(ANNDef parent, string name, int widthIn, int widthOut,
float[,] weights, float[] biases,
ActivationType activation, LayerSubtype subtype = LayerSubtype.Normal,
float[] movingMeans = null, float[] movingVariances = null,
float[] betas = null, float[] gammas = null,
float clipValue = float.NaN,
int sparseBinaryInputMaxNonzerPerRow = 0)
{
Parent = parent;
Name = name;
Activation = activation;
WidthIn = widthIn;
WidthOut = widthOut;
Weights = weights;
WeightsTr = new float[weights.GetLength(1), weights.GetLength(0)];
WeightsFP16Tr = new FP16[weights.GetLength(1), weights.GetLength(0)];
InputLayerMaxCountNonzeroPerRow = sparseBinaryInputMaxNonzerPerRow;
if (weights.GetLength(0) != widthIn || weights.GetLength(1) != widthOut)
{
throw new Exception("incorrect weight shape");
}
//Weights1D = new float[WidthIn * WidthOut];
Weights1D = new AlignedFloatArray(WidthIn * WidthOut, 128).GetManagedArray();
Weights1DTr = new AlignedFloatArray(WidthIn * WidthOut, 128).GetManagedArray();
int offset = 0;
for (int i = 0; i < WidthIn; i++)
{
for (int j = 0; j < WidthOut; j++)
{
WeightsTr[j, i] = weights[i, j];
WeightsFP16Tr[j, i] = (FP16)weights[i, j];
Weights1D[offset++] = weights[i, j];
Weights1DTr[j * WidthIn + i] = weights[i, j];
}
}
// Prepare copy of weights for AVX2 (8 floats at a time) where stride is 8
if (WidthOut >= 8)
{
Weights1DTrBlocked8 = MathUtils.BlockedMatrix(Weights1DTr, WidthOut, WidthIn, 8, 8);
}
Biases = new AlignedFloatArray(biases.Length, 128, biases).GetManagedArray();
Subtype = subtype;
MovingMeans = movingMeans;
float[] MovingVariances = movingVariances;
Betas = betas;
Gammas = gammas;
ClipValueMax = clipValue;
if (MovingVariances != null)
{
// Precompute standard deviations for speed
MovingStdDevs = new AlignedFloatArray(MovingVariances.Length, 128, biases).GetManagedArray();
for (int i = 0; i < MovingStdDevs.Length; i++)
{
MovingStdDevs[i] = (float)System.Math.Sqrt(MovingVariances[i] + EPSILON);
}
}
}