private static LSTMGateWeight LoadLSTMGateWeights(BinaryReader br)
{
var w = br.ReadInt32();
var h = br.ReadInt32();
var vqSize = br.ReadInt32();
LSTMGateWeight gateWeight = new LSTMGateWeight();
Logger.WriteLine("Loading LSTM-Weight: width:{0}, height:{1}, vqSize:{2}...", w, h, vqSize);
var m = new float[w][];
gateWeight.weights = m;
gateWeight.layerSize = w;
gateWeight.denseFeatureSize = h;
for (var i = 0; i < w; i++)
{
m[i] = new float[h];
for (var j = 0; j < h; j++)
{
m[i][j] = br.ReadSingle();
}
}
return(gateWeight);
}
private void UpdateOutputGateWeights(LSTMGateWeight gateWeight, int i, float err)
{
var j = 0;
float[] learningrate_i = gateWeight.learningRate[i];
float[] weights_i = gateWeight.weightsDelta[i];
foreach (var denseFeature in DenseFeatureGroups)
{
int k = 0;
var denseFeatureSize = denseFeature.Length;
var moreItems = (denseFeatureSize % Vector <float> .Count);
while (k < denseFeatureSize - moreItems)
{
Vector <float> vecDelta = new Vector <float>(denseFeature, k);
vecDelta = vecDelta * err;
var w_i = new Vector <float>(weights_i, j);
w_i += vecDelta;
w_i.CopyTo(weights_i, j);
j += Vector <float> .Count;
k += Vector <float> .Count;
}
while (k < denseFeatureSize)
{
float delta = denseFeature[k] * err;
weights_i[j] += delta;
j++;
k++;
}
}
}
private void UpdateOutputGateWeights(LSTMGateWeight gateWeight, int i, float err)
{
var j = 0;
float[] learningrate_i = gateWeight.learningRate[i];
float[] weights_i = gateWeight.weightsDelta[i];
var moreItems = (DenseFeatureSize % Vector <float> .Count);
while (j < DenseFeatureSize - moreItems)
{
Vector <float> vecDelta = new Vector <float>(DenseFeature, j);
vecDelta = vecDelta * err;
var w_i = new Vector <float>(weights_i, j);
w_i += vecDelta;
w_i.CopyTo(weights_i, j);
j += Vector <float> .Count;
}
while (j < DenseFeatureSize)
{
float delta = DenseFeature[j] * err;
weights_i[j] += delta;
j++;
}
}
public override void Load(BinaryReader br)
{
LayerSize = br.ReadInt32();
SparseFeatureSize = br.ReadInt32();
DenseFeatureSize = br.ReadInt32();
AllocateMemoryForCells();
AllocateMemoryForLSTMCells();
//Create cells of each layer
CreateCell(br);
//Load weight matrix between each two layer pairs
//weight input->hidden
if (SparseFeatureSize > 0)
{
Logger.WriteLine("Loading sparse feature weights...");
sparseFeatureWeights = LoadLSTMWeights(br);
}
if (DenseFeatureSize > 0)
{
//weight fea->hidden
Logger.WriteLine("Loading dense feature weights...");
wDenseInputGate = LoadLSTMGateWeights(br);
wDenseCellGate = LoadLSTMGateWeights(br);
wDenseForgetGate = LoadLSTMGateWeights(br);
wDenseOutputGate = LoadLSTMGateWeights(br);
}
}
private void UpdateGateWeights(LSTMGateWeight gateWeight, int i, float featureDerivate, float c_yForget, float err)
{
var j = 0;
float[] deri_i = gateWeight.deri[i];
float[] learningrate_i = gateWeight.learningRate[i];
float[] weights_i = gateWeight.weights[i];
while (j < DenseFeatureSize)
{
var feature = new Vector <float>(DenseFeature, j);
var wd = feature * featureDerivate;
var wd_i = new Vector <float>(deri_i, j);
wd += wd_i * c_yForget;
wd.CopyTo(deri_i, j);
Vector <float> vecDelta = wd * err;
// vecDelta = RNNHelper.NormalizeGradient(vecDelta);
//Computing learning rate and update it
var wlr_i = new Vector <float>(learningrate_i, j);
var vecLearningRate = ComputeLearningRate(vecDelta, ref wlr_i);
wlr_i.CopyTo(learningrate_i, j);
lock (gateWeight.locker[i])
{
var w_i = new Vector <float>(weights_i, j);
w_i += vecLearningRate * vecDelta;
w_i.CopyTo(weights_i, j);
}
j += Vector <float> .Count;
}
}
private void UpdateOutputGateWeights(LSTMGateWeight gateWeight, int i, float err)
{
var j = 0;
float[] learningrate_i = gateWeight.learningRate[i];
float[] weights_i = gateWeight.weights[i];
while (j < DenseFeatureSize)
{
Vector <float> vecDelta = new Vector <float>(DenseFeature, j);
vecDelta = vecDelta * err;
// vecDelta = RNNHelper.NormalizeGradient(vecDelta);
var wlr_i = new Vector <float>(learningrate_i, j);
var vecLearningRate = ComputeLearningRate(vecDelta, ref wlr_i);
wlr_i.CopyTo(learningrate_i, j);
lock (gateWeight.locker[i])
{
var w_i = new Vector <float>(weights_i, j);
w_i += vecLearningRate * vecDelta;
w_i.CopyTo(weights_i, j);
}
j += Vector <float> .Count;
}
}
public LSTMGateWeight CloneSharedWeights()
{
LSTMGateWeight gateWeight = new LSTMGateWeight();
gateWeight.InitWeights(layerSize, denseFeatureSize);
gateWeight.weights = weights;
gateWeight.weightsDelta = weightsDelta;
gateWeight.learningRate = learningRate;
return(gateWeight);
}
private void UpdateGateWeights(int curState, LSTMGateWeight gateWeight, int i, float featureDerivate, float c_yForget, float err)
{
var j = 0;
float[] deri_i = gateWeight.deri[i];
float[] learningrate_i = gateWeight.learningRate[i];
float[] weights_i = gateWeight.weights[i];
while (j < DenseFeatureSize - Vector <float> .Count)
{
var feature = new Vector <float>(DenseFeature, j);
var wd = feature * featureDerivate;
if (curState > 0)
{
var wd_i = new Vector <float>(deri_i, j);
wd += wd_i * c_yForget;
}
wd.CopyTo(deri_i, j);
Vector <float> vecDelta = wd * err;
vecDelta = RNNHelper.NormalizeGradient(vecDelta);
var wlr_i = new Vector <float>(learningrate_i, j);
var vecLearningRate = ComputeLearningRate(vecDelta, ref wlr_i);
var w_i = new Vector <float>(weights_i, j);
w_i += vecLearningRate * vecDelta;
w_i.CopyTo(weights_i, j);
wlr_i.CopyTo(learningrate_i, j);
j += Vector <float> .Count;
}
while (j < DenseFeatureSize)
{
var wd = DenseFeature[j] * featureDerivate;
if (curState > 0)
{
wd += deri_i[j] * c_yForget;
}
deri_i[j] = wd;
float delta = wd * err;
delta = RNNHelper.NormalizeGradient(delta);
var wlr_i = learningrate_i[j];
var learningRate = ComputeLearningRate(delta, ref wlr_i);
weights_i[j] += learningRate * delta;
learningrate_i[j] = wlr_i;
j++;
}
}
private void UpdateGateWeights(LSTMGateWeight gateWeight, int i, float featureDerivate, float c_yForget, float err)
{
var j = 0;
float[] deri_i = gateWeight.deri[i];
float[] learningrate_i = gateWeight.learningRate[i];
float[] weights_i = gateWeight.weightsDelta[i];
foreach (var denseFeature in DenseFeatureGroups)
{
int k = 0;
var denseFeatureSize = denseFeature.Length;
var moreItems = (denseFeatureSize % Vector <float> .Count);
while (k < denseFeatureSize - moreItems)
{
var feature = new Vector <float>(denseFeature, k);
var wd = feature * featureDerivate;
var wd_i = new Vector <float>(deri_i, j);
wd += wd_i * c_yForget;
wd.CopyTo(deri_i, j);
Vector <float> vecDelta = wd * err;
var w_i = new Vector <float>(weights_i, j);
w_i += vecDelta;
w_i.CopyTo(weights_i, j);
j += Vector <float> .Count;
k += Vector <float> .Count;
}
while (k < denseFeatureSize)
{
var wd = denseFeature[k] * featureDerivate;
wd += deri_i[j] * c_yForget;
deri_i[j] = wd;
weights_i[j] += wd * err;
j++;
k++;
}
}
}
public override void InitializeWeights(int sparseFeatureSize, int denseFeatureSize)
{
SparseFeatureSize = sparseFeatureSize;
DenseFeatureSize = denseFeatureSize;
if (DenseFeatureSize % Vector <float> .Count != 0)
{
DenseFeatureSize += (Vector <float> .Count - (DenseFeatureSize % Vector <float> .Count));
}
InitializeCellWeights(null);
if (SparseFeatureSize > 0)
{
sparseFeatureWeights = new Vector4[LayerSize][];
for (var i = 0; i < LayerSize; i++)
{
sparseFeatureWeights[i] = new Vector4[SparseFeatureSize];
for (var j = 0; j < SparseFeatureSize; j++)
{
sparseFeatureWeights[i][j] = InitializeLSTMWeight();
}
}
}
if (DenseFeatureSize > 0)
{
wDenseInputGate = new LSTMGateWeight();
wDenseForgetGate = new LSTMGateWeight();
wDenseCellGate = new LSTMGateWeight();
wDenseOutputGate = new LSTMGateWeight();
wDenseInputGate.InitWeights(LayerSize, DenseFeatureSize);
wDenseForgetGate.InitWeights(LayerSize, DenseFeatureSize);
wDenseCellGate.InitWeights(LayerSize, DenseFeatureSize);
wDenseOutputGate.InitWeights(LayerSize, DenseFeatureSize);
}
InitializeInternalTrainingParameters();
Logger.WriteLine(
"Initializing weights, sparse feature size: {0}, dense feature size: {1}, random value is {2}",
SparseFeatureSize, DenseFeatureSize, RNNHelper.rand.NextDouble());
}
private void SaveLSTMweights(LSTMGateWeight gateWeight, BinaryWriter fo, bool bVQ = false)
{
float[][] weights = gateWeight.weights;
var w = weights.Length;
var h = weights[0].Length;
Logger.WriteLine($"Saving LSTM gate weight matrix. width: {w}, height: {h}");
fo.Write(w);
fo.Write(h);
fo.Write(0);
for (var i = 0; i < w; i++)
{
for (var j = 0; j < h; j++)
{
fo.Write(weights[i][j]);
}
}
}
private void UpdateGateWeights(LSTMGateWeight gateWeight, int i, float featureDerivate, float c_yForget, float err)
{
var j = 0;
float[] deri_i = gateWeight.deri[i];
float[] learningrate_i = gateWeight.learningRate[i];
float[] weights_i = gateWeight.weightsDelta[i];
var moreItems = (DenseFeatureSize % Vector <float> .Count);
while (j < DenseFeatureSize - moreItems)
{
var feature = new Vector <float>(DenseFeature, j);
var wd = feature * featureDerivate;
var wd_i = new Vector <float>(deri_i, j);
wd += wd_i * c_yForget;
wd.CopyTo(deri_i, j);
Vector <float> vecDelta = wd * err;
var w_i = new Vector <float>(weights_i, j);
w_i += vecDelta;
w_i.CopyTo(weights_i, j);
j += Vector <float> .Count;
}
while (j < DenseFeatureSize)
{
var wd = DenseFeature[j] * featureDerivate;
wd += deri_i[j] * c_yForget;
deri_i[j] = wd;
float delta = wd * err;
weights_i[j] += delta;
j++;
}
}
public override void InitializeWeights(int sparseFeatureSize, int denseFeatureSize)
{
SparseFeatureSize = sparseFeatureSize;
DenseFeatureSize = denseFeatureSize;
CreateCell(null);
if (SparseFeatureSize > 0)
{
sparseFeatureWeights = new Vector4[LayerSize][];
sparseFeatureToHiddenDeri = new Vector3[LayerSize][];
for (var i = 0; i < LayerSize; i++)
{
sparseFeatureWeights[i] = new Vector4[SparseFeatureSize];
sparseFeatureToHiddenDeri[i] = new Vector3[SparseFeatureSize];
for (var j = 0; j < SparseFeatureSize; j++)
{
sparseFeatureWeights[i][j] = InitializeLSTMWeight();
}
}
}
if (DenseFeatureSize > 0)
{
wDenseInputGate = new LSTMGateWeight();
wDenseForgetGate = new LSTMGateWeight();
wDenseCellGate = new LSTMGateWeight();
wDenseOutputGate = new LSTMGateWeight();
wDenseInputGate.Init(LayerSize, DenseFeatureSize);
wDenseForgetGate.Init(LayerSize, DenseFeatureSize);
wDenseCellGate.Init(LayerSize, DenseFeatureSize);
wDenseOutputGate.Init(LayerSize, DenseFeatureSize, false);
}
Logger.WriteLine(
"Initializing weights, sparse feature size: {0}, dense feature size: {1}, random value is {2}",
SparseFeatureSize, DenseFeatureSize, RNNHelper.rand.NextDouble());
}
private void UpdateOutputGateWeights(LSTMGateWeight gateWeight, int i, float err)
{
var j = 0;
float[] learningrate_i = gateWeight.learningRate[i];
float[] weights_i = gateWeight.weights[i];
while (j < DenseFeatureSize - Vector <float> .Count)
{
Vector <float> vecDelta = new Vector <float>(DenseFeature, j);
vecDelta = vecDelta * err;
vecDelta = RNNHelper.NormalizeGradient(vecDelta);
var wlr_i = new Vector <float>(learningrate_i, j);
var vecLearningRate = ComputeLearningRate(vecDelta, ref wlr_i);
var w_i = new Vector <float>(weights_i, j);
w_i += vecLearningRate * vecDelta;
w_i.CopyTo(weights_i, j);
wlr_i.CopyTo(learningrate_i, j);
j += Vector <float> .Count;
}
while (j < DenseFeatureSize)
{
float delta = DenseFeature[j] * err;
delta = RNNHelper.NormalizeGradient(delta);
var wlr_i = learningrate_i[j];
var learningRate = ComputeLearningRate(delta, ref wlr_i);
weights_i[j] += learningRate * delta;
learningrate_i[j] = wlr_i;
j++;
}
}
