public void PredictOne(Sent sent)
{
G.Need = false;
var c = new ArcStandardConfig(sent, false, null);
GetTokenRepr(sent, false);
while (!c.IsTerminal())
{
GetTrans(c, out Tensor op, out Tensor label);
// in labelid 0 is root
// in mapping 1 is root, 0 is always unk
var labelid = label.W.MaxIndex();
var plabel = _train.DepLabel[labelid + 1];
var scores = op.W.Storage;
var optScore = float.NegativeInfinity;
var optTrans = ArcStandardConfig.Op.shift;
for (var j = 0; j < 3; ++j)
{
if (scores[j] > optScore && c.CanApply((ArcStandardConfig.Op)j, plabel))
{
optScore = scores[j];
optTrans = (ArcStandardConfig.Op)j;
}
}
c.Apply(optTrans, plabel);
}
}
public float TrainOne(Sent sent)
{
G.Need = true;
var c = new ArcStandardConfig(sent, false, _rngChoice);
GetTokenRepr(sent, true);
var losses = 0f;
while (!c.IsTerminal())
{
var oracle = c.GetOracle(_conf.OraType, out float[] target, out string slabel);
GetTrans(c, out Tensor op, out Tensor label);
G.SoftmaxWithCrossEntropy(op, target, out float loss);
losses += loss;
if (slabel != null)
{
// in mapping 0 is unk, 1 is root
var labelid = _train.DepLabel[slabel] - 1;
G.SoftmaxWithCrossEntropy(label, labelid, out loss);
losses += loss;
}
c.Apply(oracle, slabel);
}
G.Backward();
_opt.Update(FixedParams);
_opt.Update(VariedParams);
VariedParams.Clear();
G.Clear();
return(losses / sent.Count);
}
public void TestOne(Sent sent)
{
G.Need = false;
var c = new ArcStandardConfig(sent, false, null);
GetTokenRepr(sent, false);
while (!c.IsTerminal())
{
GetTrans(c, out Tensor op, out Tensor label);
// in labelid 0 is root
// in mapping 1 is root, 0 is always unk
var labelid = label.W.MaxIndex();
var plabel = _train.DepLabel[labelid + 1];
var opid = op.W.MaxIndex();
var optTrans = (ArcStandardConfig.Op)opid;
c.Apply(optTrans, plabel);
}
}