/// <inheritdoc />
public float Loss(Tensor y, int[] expected, bool calculateGradient)
{
if (y == null)
{
throw new ArgumentNullException(nameof(y));
}
if (expected == null)
{
throw new ArgumentNullException(nameof(expected));
}
#pragma warning disable SA1312 // Variable names must begin with lower-case letter
int L = expected.Length; // Number of labels
int T = y.Axes[0]; // Number of mini-batches (time)
int A = y.Strides[0]; // Number of classes (alphabet size)
#pragma warning restore SA1312 // Variable names must begin with lower-case letter
int[] labels = CTCLoss.InsertBlanks(expected, A, this.BlankLabelIndex, out int repeats);
if (L + repeats > T)
{
if (calculateGradient)
{
Vectors.Copy(y.Length, y.Weights, 0, y.Gradient, 0);
}
// not enough elements to compute
return(float.PositiveInfinity);
}
#pragma warning disable SA1312 // Variable names must begin with lower-case letter
int S = labels.Length; // Number of labels with blanks
#pragma warning restore SA1312 // Variable names must begin with lower-case letter
// convert predicted probabilities into log space
float[] ylog = new float[y.Length];
Vectors.Log(y.Length, y.Weights, 0, ylog, 0);
// compute alphas
float[] alphas = new float[T * S];
////CTCLoss.CTCComputeAlphas(T, A, S, ylog, labels, alphas);
NativeMethods.CTCComputeAlphas(T, A, S, ylog, labels, alphas);
float logLossA = Mathematics.LogSumExp(alphas[alphas.Length - 1], alphas[alphas.Length - 2]);
if (float.IsNegativeInfinity(logLossA))
{
if (calculateGradient)
{
Vectors.Copy(y.Length, y.Weights, 0, y.Gradient, 0);
}
return(float.PositiveInfinity);
}
if (calculateGradient)
{
// compute betas
float[] betas = new float[T * S];
NativeMethods.CTCComputeBetas(T, A, S, ylog, labels, betas);
////float logLossB = MKL.LogSumExp(betas.Weights[0], betas.Weights[1]);
// compute unnormalized gradient
Mathematics.Add(alphas.Length, betas, 0, alphas, 0);
NativeMethods.CTCReduceAlphasBetas(T, A, S, alphas, labels, y.Gradient);
Mathematics.Sub(y.Length, ylog, 0, y.Gradient, 0);
Mathematics.SubC(y.Length, logLossA, y.Gradient, 0);
Vectors.Exp(y.Length, y.Gradient, 0);
// remove NaN
// NaN may come from various sources (for instance log(y) where y = 0)
Arrays.Replace(y.Length, y.Gradient, 0, float.NaN, 0.0f, y.Gradient, 0);
Debug.Assert(!float.IsNaN(y.Gradient[0]), "Tensor contains invalid weight.");
}
Debug.Assert(!float.IsNaN(logLossA), "Calculated loss is invalid.");
return(-logLossA);
}