public void TestMethod5()
{
const int T = 3; // Length of utterance (time)
const int L = 5; // Alphabet size
int[] axes = new[] { T, L };
Tensor x = new Tensor(null, axes);
x.Set(1.0f);
Tensor y = new Tensor(null, axes);
Vectors.SoftMax(x.Length, x.Weights, 0, L, y.Weights, 0);
y.Weights[0] = 0;
CTCLoss loss = new CTCLoss();
int[] labels = { 1, 2, 3 };
float expected = -(float)Math.Log(y.Weights[(0 * L) + labels[0]] *
y.Weights[(1 * L) + labels[1]] *
y.Weights[(2 * L) + labels[2]]);
float score = loss.Loss(y, labels, false);
Assert.AreEqual(expected, score, 1e-6);
score = loss.Loss(y, labels, true);
Assert.AreEqual(expected, score, 1e-6);
}
private ILoss <int[]> CreateLoss()
{
ILoss <int[]> loss = null;
switch (this.TaskParameters.Loss.Name)
{
case "LogLikelihood":
default:
loss = new LogLikelihoodLoss();
break;
case "CTC":
loss = new CTCLoss();
break;
}
JsonSerializer jsonSerializer = new JsonSerializer();
using (JTokenReader jtokenReader = new JTokenReader(this.TaskParameters.Loss.Parameters))
{
jsonSerializer.Populate(jtokenReader, loss);
}
return(loss);
}
public void TestMethod2()
{
const int T = 5;
const int L = 6;
int[] axes = new[] { T, L };
Tensor y = new Tensor(null, axes);
y.Set(new float[]
{
0.633766f, 0.221185f, 0.0917319f, 0.0129757f, 0.0142857f, 0.0260553f,
0.111121f, 0.588392f, 0.278779f, 0.0055756f, 0.00569609f, 0.010436f,
0.0357786f, 0.633813f, 0.321418f, 0.00249248f, 0.00272882f, 0.0037688f,
0.0663296f, 0.643849f, 0.280111f, 0.00283995f, 0.0035545f, 0.00331533f,
0.458235f, 0.396634f, 0.123377f, 0.00648837f, 0.00903441f, 0.00623107f
});
CTCLoss loss = new CTCLoss()
{
BlankLabelIndex = 5
};
int[] labels = { 0, 1, 2, 1, 0 /*, 0, 1, 1, 0*/ };
float expected = -(float)Math.Log(y.Weights[(0 * L) + labels[0]] *
y.Weights[(1 * L) + labels[1]] *
y.Weights[(2 * L) + labels[2]] *
y.Weights[(3 * L) + labels[3]] *
y.Weights[(4 * L) + labels[4]]);
float score = loss.Loss(y, labels, true);
Helpers.AreArraysEqual(
new float[]
{
1, 0, 0, 0, 0, 0,
0, 1, 0, 0, 0, 0,
0, 0, 1, 0, 0, 0,
0, 1, 0, 0, 0, 0,
1, 0, 0, 0, 0, 0
},
y.Gradient);
}
public void TestMethod3()
{
const int T = 5;
const int L = 6;
int[] axes = new[] { T, L };
Tensor y = new Tensor(null, axes);
y.Set(new float[]
{
0.30176f, 0.28562f, 0.0831517f, 0.0862751f, 0.0816851f, 0.161508f,
0.24082f, 0.397533f, 0.0557226f, 0.0546814f, 0.0557528f, 0.19549f,
0.230246f, 0.450868f, 0.0389607f, 0.038309f, 0.0391602f, 0.202456f,
0.280884f, 0.429522f, 0.0326593f, 0.0339046f, 0.0326856f, 0.190345f,
0.423286f, 0.315517f, 0.0338439f, 0.0393744f, 0.0339315f, 0.154046f
});
CTCLoss loss = new CTCLoss()
{
BlankLabelIndex = 5
};
int[] labels = { 0, 1, 1, 0 };
float expected = 5.42262f; // from tensorflow
float score = loss.Loss(y, labels, true);
Assert.AreEqual(expected, score, 1e-4);
Helpers.AreArraysEqual(
new float[]
{
1, 0, 0, 0, 0, 0,
0, 1, 0, 0, 0, 0,
0, 0, 0, 0, 0, 1,
0, 1, 0, 0, 0, 0,
1, 0, 0, 0, 0, 0
},
y.Gradient);
}