public void ForwardVolumes()
{
Random random = new Random(0);
ClassificationNetworkTrainer trainer = new ClassificationNetworkTrainer();
SGD sgd = new SGD()
{
LearningRate = 0.0001f,
Momentum = 0.0f
};
ClassificationNetwork network = ClassificationNetwork.FromArchitecture("1x1x2~5N~5N~3N", this.classes);
// lets test 100 random point and label settings
// note that this should work since l2 and l1 regularization are off
// an issue is that if step size is too high, this could technically fail...
for (int k = 0; k < 100; k++)
{
int gti = (int)Math.Floor(random.NextDouble() * 3);
Tensor x = new Tensor(null, new Shape(Shape.BWHC, 1, 1, 1, 2));
x.Set(new float[] { ((float)random.NextDouble() * 2) - 1, ((float)random.NextDouble() * 2) - 1 });
Tensor pv = network.Forward(null, x).Clone() as Tensor;
trainer.RunEpoch(
k,
network,
Enumerable.Repeat((x, new string[] { this.classes[gti] }), 1),
sgd,
new LogLikelihoodLoss(),
CancellationToken.None);
Tensor pv2 = network.Forward(null, x).Clone() as Tensor;
Assert.IsTrue(pv2.Weights[gti] > pv.Weights[gti], "k: {0}, gti: {1}, pv2[gti]: {2}, pv[gti]: {3}", k, gti, pv2.Weights[gti], pv.Weights[gti]);
}
}
public void ForwardToProbability()
{
ClassificationNetwork network = ClassificationNetwork.FromArchitecture("1x1x2~5N~5N~3N", this.classes);
Tensor x = new Tensor(null, new Shape(Shape.BWHC, 1, 1, 1, 2));
x.Set(new float[] { 0.2f, -0.3f });
Tensor probability = network.Forward(null, x);
Assert.AreEqual(3, probability.Length); // 3 classes output
for (int i = 0; i < probability.Length; i++)
{
Assert.IsTrue(probability.Weights[i] > 0.0);
Assert.IsTrue(probability.Weights[i] < 1.0);
}
Assert.AreEqual(1.0, probability.Weights[0] + probability.Weights[1] + probability.Weights[2], 1e-6);
}
public void XorTest1()
{
const int AlphabetSize = 16;
const int VectorSize = 4;
const int BatchSize = 3000;
const int Epochs = 200;
const int TestBatchSize = 3000;
Random random = new Random(0);
string[] classes = Enumerable.Range(0, AlphabetSize).Select(v => v.ToString(CultureInfo.InvariantCulture)).ToArray();
ClassificationNetwork network = ClassificationNetwork.FromArchitecture("1x1x4~80-80-80-16LSTM", classes);
float[] vectors = new RandomGeneratorF().Generate(AlphabetSize * VectorSize);
(Tensor, int[]) createSample(int size)
{
Tensor input = new Tensor(null, new[] { size, 1, 1, VectorSize });
int[] truth = new int[size];
int v = 0;
for (int i = 0; i < size; i++)
{
v ^= random.Next(0, AlphabetSize);
Vectors.Copy(VectorSize, vectors, v * VectorSize, input.Weights, i * VectorSize);
if (i > 0)
{
truth[i - 1] = v;
}
}
return(input, truth);
}
// train the network
Trainer <int[]> trainer = new Trainer <int[]>()
{
ClipValue = 2.0f
};
SGD sgd = new SGD();
ILoss <int[]> loss = new LogLikelihoodLoss();
for (int epoch = 0; epoch < Epochs; epoch++)
{
(Tensor, int[])sample = createSample(BatchSize);
TrainingResult result = trainer.RunEpoch(
network,
Enumerable.Repeat(sample, 1),
epoch,
sgd,
loss,
CancellationToken.None);
Console.WriteLine(result.CostLoss);
}
// test the network
(Tensor x, int[] expected) = createSample(TestBatchSize);
Tensor y = network.Forward(null, x);
////y.Reshape(testBatchSize - 1);
////expected.Reshape(testBatchSize - 1);
float error = loss.Loss(y, expected, false);
Console.WriteLine(y);
Console.WriteLine(expected);
////Console.WriteLine(y.Axes[1]);
Console.WriteLine(error);
////Assert.IsTrue(errorL1 < 0.01, errorL1.ToString(CultureInfo.InvariantCulture));
}