public Matrix <double>[] Gradients(Matrix <double> x, int action, double qtarget)
{
//Forward
Matrix <double> x1 = NNOperations.matrixmul(W [0], x);
Matrix <double> x2 = NNOperations.add(W [1], x1);
Matrix <double> x3 = NNOperations.sigm(x2);
Matrix <double> x4 = NNOperations.matrixmul(W [2], x3);
Matrix <double> x5 = NNOperations.add(W [3], x4);
//Loss
Matrix <double> ind = Matrix <double> .Build.Dense(x5.RowCount, 1);
ind [action, 0] = 1.0;
Matrix <double> x6 = x5.PointwiseMultiply(ind);
Matrix <double> ygold = ind.Multiply(qtarget);
Matrix <double>[] gx6_gyg = NNOperations.mse_b(x6, ygold, 1.0);
Matrix <double>[] gx5_gind = NNOperations.ewisemul_b(x5, ind, gx6_gyg [0]);
Matrix <double>[] gw3_gx4 = NNOperations.add_b(W[3], x4, gx5_gind[0]);
Matrix <double>[] gw2_gx3 = NNOperations.matrixmul_b(W[2], x3, gw3_gx4[1]);
Matrix <double> gx2 = NNOperations.sigm_b(x2, x3, gw2_gx3[1]);
Matrix <double>[] gw1_gx1 = NNOperations.add_b(W[1], x1, gx2);
Matrix <double>[] gw0_gx = NNOperations.matrixmul_b(W[0], x, gw1_gx1[1]);
Matrix <double> [] gradients = { gw0_gx[0], gw1_gx1[0], gw2_gx3[0], gw3_gx4[0] };
return(gradients);
}
public Matrix <double> lossbackward(Matrix <double>[] x)
{
//Layer1
Matrix <double>[] gx1_gyg = NNOperations.mse_b(x[2], x[3], 1.0);
Matrix <double>[] gx_gind = NNOperations.ewisemul_b(x[0], x[1], gx1_gyg [0]);
return(gx_gind [0]);
}
public Matrix <double>[] Gradients(Matrix <double> x, Vector <double> ygold)
{
//Forward
Matrix <double> x1 = NNOperations.matrixmul(W [0], x);
Matrix <double> x2 = NNOperations.add(W [1], x1);
Matrix <double> x3 = NNOperations.sigm(x2);
Matrix <double> x4 = NNOperations.matrixmul(W [2], x3);
Matrix <double> x5 = NNOperations.add(W [3], x4);
//Loss
//double loss = NNOperations.mse (x5.Column(0),ygold);
//Backward
Matrix <double> ym = ygold.ToColumnMatrix();
Matrix <double>[] gx5_gym = NNOperations.mse_b(x5, ym, 1.0);
Matrix <double>[] gw3_gx4 = NNOperations.add_b(W[3], x4, gx5_gym[0]);
Matrix <double>[] gw2_gx3 = NNOperations.matrixmul_b(W[2], x3, gw3_gx4[1]);
Matrix <double> gx2 = NNOperations.sigm_b(x2, x3, gw2_gx3[1]);
Matrix <double>[] gw1_gx1 = NNOperations.add_b(W[1], x1, gx2);
Matrix <double>[] gw0_gx = NNOperations.matrixmul_b(W[0], x, gw1_gx1[1]);
Matrix <double> [] gradients = { gw0_gx[0], gw1_gx1[0], gw2_gx3[0], gw3_gx4[0] };
return(gradients);
}
void Test()
{
Stopwatch t = new Stopwatch();
t.Start();
for (int i = 0; i < TestingSet.Size; i++)
{
Image image = TrainingSet.Data[i];
uint label = TrainingSet.Labels[i];
Vector labelVector = Vector.Create(10, 0f);
labelVector[(int)label] = 1;
Network.Input.Nodes = new Vector(image.NormalizedPixels);
Network.FeedForward();
Vector loss = NNOperations.OutputLoss(Network.Output, labelVector, Network.Properties.LossFunction);
float avgLoss = 0;
loss.ForEach(v => avgLoss += v);
avgLoss /= loss.Length;
Console.WriteLine(Network.Output.Nodes.ToString("Output (Index: " + i + ") (avg: " + avgLoss + ")"));
}
t.Stop();
Console.WriteLine("Testing time: " + t.ElapsedMilliseconds + "ms");
}
public Matrix <double>[] lastbackward(Matrix <double>[] x, int i, Matrix <double> gy)
{
//Layer1
Matrix <double>[] gw1_gx1 = NNOperations.add_b(W[i + 1], x[1], gy);
Matrix <double>[] gw0_gx = NNOperations.matrixmul_b(W[i], x[0], gw1_gx1[1]);
Matrix <double>[] gradients = { gw0_gx [0], gw1_gx1 [0], gw0_gx [1] };
return(gradients);
}
public Matrix <double>[] lastlayer(Matrix <double> x, int i)
{
//Layer1
Matrix <double> x1 = NNOperations.matrixmul(W [i], x);
Matrix <double> x2 = NNOperations.add(W [i + 1], x1);
Matrix <double>[] results = { x, x1, x2 };
return(results);
}
public Matrix <double>[] fcbackward(System.Func <Matrix <double>, Matrix <double>, Matrix <double>, Matrix <double> > activation_b, Matrix <double>[] x, int i, Matrix <double> gy)
{
//Layer1
Matrix <double> gx2 = activation_b(x[2], x[3], gy);
Matrix <double>[] gw1_gx1 = NNOperations.add_b(W[i + 1], x[1], gx2);
Matrix <double>[] gw0_gx = NNOperations.matrixmul_b(W[i], x[0], gw1_gx1[1]);
Matrix <double>[] gradients = { gw0_gx [0], gw1_gx1 [0], gw0_gx [1] };
return(gradients);
}
public Matrix <double>[] fclayer(System.Func <Matrix <double>, Matrix <double> > activation, Matrix <double> x, int i)
{
//Layer1
Matrix <double> x1 = NNOperations.matrixmul(W [i], x);
Matrix <double> x2 = NNOperations.add(W [i + 1], x1);
Matrix <double> x3 = activation(x2);
Matrix <double>[] results = { x, x1, x2, x3 };
return(results);
}
public double Loss(Matrix <double> x, int action, double qtarget)
{
Matrix <double> x5 = Forward(x);
Matrix <double> ind = Matrix <double> .Build.Dense(x5.RowCount, 1);
ind [action, 0] = 1.0;
Matrix <double> x6 = x5.PointwiseMultiply(ind);
ind [action, 0] = qtarget;
return(NNOperations.mse(x6, ind));
}
public Matrix <double> Forward(Matrix <double> x)
{
//Layer1
Matrix <double> x1 = NNOperations.matrixmul(W [0], x);
Matrix <double> x2 = NNOperations.add(W [1], x1);
Matrix <double> x3 = NNOperations.sigm(x2);
//Layer2
Matrix <double> x4 = NNOperations.matrixmul(W [2], x3);
Matrix <double> x5 = NNOperations.add(W [3], x4);
return(x5);
}
public Matrix <double>[] losslayer(Matrix <double> x, int action, double qtarget)
{
//Layer1
Matrix <double> ind = Matrix <double> .Build.Dense(x.RowCount, 1);
ind [action, 0] = 1.0;
Matrix <double> x1 = x.PointwiseMultiply(ind);
Matrix <double> ygold = ind.Multiply(qtarget);
Matrix <double> loss = M.Dense(1, 1, NNOperations.mse(x1, ygold));
Matrix <double>[] results = { x, ind, x1, ygold, loss };
return(results);
}
public NNBayes(int[] numNeurons, params object[] parameters)
{
float l2Penalty = 0.1F;
_preds = new TFOutput[numSamples];
_session = new TFSession();
_graph = _session.Graph;
_graph.Seed = Global.Random.Next();
_layers = new BayesLayer[numNeurons.Length - 1];
_input = _graph.Placeholder(TFDataType.Float, new TFShape(-1, numNeurons[0]));
_output = _graph.Placeholder(TFDataType.Float, new TFShape(-1));
_decay = _graph.Placeholder(TFDataType.Float, new TFShape(1));
for (int i = 0; i < numNeurons.Length - 1; i++)
{
_layers[i] = new BayesLayer(_graph, numNeurons[i], numNeurons[i + 1]);
}
TFOutput likelihood = _graph.Const(0F);
for (int i = 0; i < numSamples; i++)
{
TFOutput act = _input;
foreach (BayesLayer layer in _layers)
{
// TFOutput W = layer.SampleW(_graph);
// TFOutput b = layer.Sampleb(_graph);
// TFOutput z = _graph.Add(_graph.MatMul(act, W), b);
TFOutput z = layer.Samplez(_graph, act);
if (layer == _layers.Last())
{
TFOutput pred = _graph.Reshape(z, _graph.Const(new TFShape(-1)));
_preds[i] = pred;
TFOutput sample_likelihood = NNOperations.LogOfNormal(_graph, pred, _output, _graph.Const(1F));
likelihood = _graph.Add(likelihood, sample_likelihood);
}
else
{
act = _graph.Relu(z);
}
}
}
TFOutput kl_W = _graph.Const(0F);
TFOutput kl_b = _graph.Const(0F);
foreach (BayesLayer layer in _layers)
{
kl_W = _graph.Add(kl_W, _graph.ReduceSum(NNOperations.KLUnivariateNormal(_graph, layer.Mu_W, NNOperations.LogTrans(_graph, layer.Phi_W), _graph.Const(0F), _graph.Const((float)(1 / (Math.Sqrt(l2Penalty)))))));
kl_b = _graph.Add(kl_b, _graph.ReduceSum(NNOperations.KLUnivariateNormal(_graph, layer.Mu_b, NNOperations.LogTrans(_graph, layer.Phi_b), _graph.Const(0F), _graph.Const((float)(1 / Math.Sqrt(l2Penalty))))));
}
TFOutput kl = _graph.Add(kl_W, kl_b);
likelihood = _graph.Div(likelihood, _graph.Const((float)numSamples));
_cost = _graph.ReduceMean(_graph.Sub(_graph.Mul(_decay, kl), likelihood));
optimizer = new AdamOptimizer();
optimizer.AddBayesLayer(_graph, _layers, _cost);
optimizer.Apply(_graph);
var runner = _session.GetRunner();
foreach (BayesLayer layer in _layers)
{
layer.Init(runner);
}
optimizer.Init(runner);
runner.Run();
}
public double Loss(Matrix <double> x, Vector <double> ygold)
{
Matrix <double> y = Forward(x);
return(NNOperations.mse(y, ygold.ToColumnMatrix()));
}