protected override FloatTensor Forward(FloatTensor prediction, FloatTensor target)
{
// Note: prediction should be logits, basically pre-softmax. This method applies softmax first.
// TODO check shapes
FloatTensor softmax = Functional.Softmax(prediction);
FloatTensor output = ((target.Mul(softmax.Log1p())).Sum()).Mul(-1);
return(output);
}