public void matrixXvectorADD(LSTMCell[] dest, neuron[] srcvec, LSTMWeight[][] srcmatrix, int from, int to, int from2, int to2)
{
//ac mod
Parallel.For(0, (to - from), parallelOption, i =>
{
for (int j = 0; j < to2 - from2; j++)
{
dest[i + from].netIn += srcvec[j + from2].ac * srcmatrix[i][j].wInputInputGate;
}
});
}
private void saveLSTMWeight(LSTMWeight[][] weight, BinaryWriter fo)
{
if (weight == null || weight.Length == 0)
{
fo.Write(0);
fo.Write(0);
}
fo.Write(weight.Length);
fo.Write(weight[0].Length);
int w = weight.Length;
int h = weight[0].Length;
for (int i = 0; i < w; i++)
{
for (int j = 0; j < h; j++)
{
fo.Write((float)weight[i][j].wInputCell);
fo.Write((float)weight[i][j].wInputForgetGate);
fo.Write((float)weight[i][j].wInputInputGate);
fo.Write((float)weight[i][j].wInputOutputGate);
}
}
}
public LSTMWeight[][] loadLSTMWeight(BinaryReader br)
{
int w = br.ReadInt32();
int h = br.ReadInt32();
LSTMWeight[][] m = new LSTMWeight[w][];
for (int i = 0; i < w; i++)
{
m[i] = new LSTMWeight[h];
for (int j = 0; j < h; j++)
{
m[i][j].wInputCell = br.ReadSingle();
m[i][j].wInputForgetGate = br.ReadSingle();
m[i][j].wInputInputGate = br.ReadSingle();
m[i][j].wInputOutputGate = br.ReadSingle();
}
}
return m;
}
public override void initWeights()
{
int INPUT = m_TrainingSet.GetSparseDimension();
//create and initialise the weights from input to hidden layer
mat_input2hidden = new LSTMWeight[L1][];
for (int i = 0; i < L1; i++)
{
mat_input2hidden[i] = new LSTMWeight[INPUT + 1];
for (int j = 0; j <= INPUT; j++)
{
mat_input2hidden[i][j] = LSTMWeightInit(INPUT);
}
}
if (fea_size > 0)
{
mat_feature2hidden = new LSTMWeight[L1][];
for (int i = 0; i < L1; i++)
{
mat_feature2hidden[i] = new LSTMWeight[fea_size];
for (int j = 0; j < fea_size; j++)
{
mat_feature2hidden[i][j] = LSTMWeightInit(INPUT);
}
}
}
//Create and intialise the weights from hidden to output layer, these are just normal weights
double hiddenOutputRand = 1 / Math.Sqrt((double)L1);
mat_hidden2output = new Matrix(L1 + 1, L2);
// mat_hidden2output = new double[L1 + 1][];
for (int i = 0; i <= L1; i++)
{
// mat_hidden2output[i] = new double[L2];
for (int j = 0; j < L2; j++)
{
mat_hidden2output[i][j] = (((double)((rand() % 100) + 1) / 100) * 2 * hiddenOutputRand) - hiddenOutputRand;
}
}
}
public LSTMWeight[][] loadLSTMWeight(BinaryReader br)
{
int w = br.ReadInt32();
int h = br.ReadInt32();
int vqSize = br.ReadInt32();
Logger.WriteLine("Loading LSTM-Weight: width:{0}, height:{1}, vqSize:{2}...", w, h, vqSize);
List<double> codeBook = new List<double>();
for (int i = 0; i < vqSize; i++)
{
codeBook.Add(br.ReadDouble());
}
LSTMWeight[][] m = new LSTMWeight[w][];
for (int i = 0; i < w; i++)
{
m[i] = new LSTMWeight[h];
for (int j = 0; j < h; j++)
{
int vqIdx = br.ReadByte();
m[i][j].wInputCell = (float)codeBook[vqIdx];
vqIdx = br.ReadByte();
m[i][j].wInputForgetGate = (float)codeBook[vqIdx];
vqIdx = br.ReadByte();
m[i][j].wInputInputGate = (float)codeBook[vqIdx];
vqIdx = br.ReadByte();
m[i][j].wInputOutputGate = (float)codeBook[vqIdx];
}
}
return m;
}
public override void initWeights()
{
//create and initialise the weights from input to hidden layer
input2hidden = new LSTMWeight[L1][];
for (int i = 0; i < L1; i++)
{
input2hidden[i] = new LSTMWeight[L0];
for (int j = 0; j < L0; j++)
{
input2hidden[i][j] = LSTMWeightInit();
}
}
if (DenseFeatureSize > 0)
{
feature2hidden = new LSTMWeight[L1][];
for (int i = 0; i < L1; i++)
{
feature2hidden[i] = new LSTMWeight[DenseFeatureSize];
for (int j = 0; j < DenseFeatureSize; j++)
{
feature2hidden[i][j] = LSTMWeightInit();
}
}
}
//Create and intialise the weights from hidden to output layer, these are just normal weights
Hidden2OutputWeight = new Matrix<double>(L2, L1);
for (int i = 0; i < Hidden2OutputWeight.GetHeight(); i++)
{
for (int j = 0; j < Hidden2OutputWeight.GetWidth(); j++)
{
Hidden2OutputWeight[i][j] = RandInitWeight();
}
}
}
private void saveLSTMWeight(LSTMWeight[][] weight, BinaryWriter fo)
{
int w = weight.Length;
int h = weight[0].Length;
int vqSize = 256;
Logger.WriteLine("Saving LSTM weight matrix. width:{0}, height:{1}, vqSize:{2}", w, h, vqSize);
fo.Write(weight.Length);
fo.Write(weight[0].Length);
//Build vector quantization model
VectorQuantization vq = new VectorQuantization();
for (int i = 0; i < w; i++)
{
for (int j = 0; j < h; j++)
{
vq.Add(weight[i][j].wInputCell);
vq.Add(weight[i][j].wInputForgetGate);
vq.Add(weight[i][j].wInputInputGate);
vq.Add(weight[i][j].wInputOutputGate);
}
}
double distortion = vq.BuildCodebook(vqSize);
Logger.WriteLine("Distortion: {0}", distortion);
//Save VQ codebook into file
fo.Write(vqSize);
for (int j = 0; j < vqSize; j++)
{
fo.Write(vq.CodeBook[j]);
}
for (int i = 0; i < w; i++)
{
for (int j = 0; j < h; j++)
{
fo.Write((byte)vq.ComputeVQ(weight[i][j].wInputCell));
fo.Write((byte)vq.ComputeVQ(weight[i][j].wInputForgetGate));
fo.Write((byte)vq.ComputeVQ(weight[i][j].wInputInputGate));
fo.Write((byte)vq.ComputeVQ(weight[i][j].wInputOutputGate));
}
}
}
