public MLPVar(Config conf, int nInput, int nOuput, OptBase opt) : base(conf, nInput, nOuput, opt)
{
Layers = new DenseUnit[Conf.Layers];
for (var i = 0; i < Layers.Length; i++)
{
if (i == 0 && i == Conf.Layers - 1)
{
Layers[i] = new DenseUnit(this, InDim, OutDim);
}
else if (i == 0)
{
Layers[i] = new DenseMaskedUnit(this, InDim, HidDim);
}
else if (i == Conf.Layers - 1)
{
Layers[i] = new DenseUnit(this, HidDim, OutDim);
}
else
{
Layers[i] = new DenseMaskedUnit(this, HidDim, HidDim);
}
}
_top = conf.TrainTop;
}
public MLP(Config conf, DataSet train)
{
_train = train;
_conf = conf;
_rng = new RandomNumberGenerator(conf.Seed);
_rngChoice = new RandomNumberGenerator(conf.Seed);
_formemb = File.Exists(conf.EmbedFile)
? InitEmbed(conf.EmbedFile, train.Form, conf.InitRange)
: InitEmbed(train.Form, conf.InitRange);
_posemb = InitEmbed(train.PosTag, conf.InitRange);
_non = new Tensor(1, conf.HiddenSize * 2, true);
_surface = new BiLstmUnit(conf.EmbeddingSize * 2, conf.HiddenSize);
_act = new DenseUnit(conf.HiddenSize * 6, conf.HiddenSize);
_actOutput = new DenseUnit(conf.HiddenSize, 3);
_label = new DenseUnit(conf.HiddenSize * 6, conf.HiddenSize);
_labelOutput = new DenseUnit(conf.HiddenSize * 3, train.DepLabel.Count);
FixedParams.AddRange(_surface.SubmitParameters());
FixedParams.AddRange(_act.SubmitParameters());
FixedParams.AddRange(_label.SubmitParameters());
FixedParams.AddRange(_actOutput.SubmitParameters());
FixedParams.AddRange(_labelOutput.SubmitParameters());
FixedParams.Add(_non);
AllParams.AddRange(FixedParams);
AllParams.AddRange(_formemb);
AllParams.AddRange(_posemb);
_surface.Init((fin, fout) => new GlorotNormalInit(fin, fout, _rng).Next);
_act.Init((fin, fout) => new GlorotNormalInit(fin, fout, _rng).Next);
_label.Init((fin, fout) => new GlorotNormalInit(fin, fout, _rng).Next);
_actOutput.Init((fin, fout) => new GlorotNormalInit(fin, fout, _rng).Next);
_labelOutput.Init((fin, fout) => new GlorotNormalInit(fin, fout, _rng).Next);
switch (conf.OptType)
{
case OptType.sgd:
_opt = new Sgd(conf.LearningRate, conf.L2RegFactor, conf.ClipBound, conf.DecayRate,
train.Count);
break;
case OptType.adagrad:
_opt = new AdaGrad(conf.LearningRate, conf.L2RegFactor, conf.ClipBound, conf.Eps);
break;
case OptType.adam:
_opt = new Adam(conf.LearningRate, conf.L2RegFactor, conf.ClipBound, 0.999f, 0.9f, 1e-8f);
break;
default:
throw new ArgumentOutOfRangeException(nameof(conf.OptType), "unknown opt type");
}
_opt.Prepare(AllParams);
G.SetRng(conf.Seed);
}