public WeightMatrix Perform(WeightMatrix input, WeightMatrix state, ComputeGraph g)
{
WeightMatrix context;
List <WeightMatrix> atten = new List <WeightMatrix>();
var stateRepeat = g.RepeatRows(state, input.Rows);
var baiseInput = new WeightMatrix(input.Rows, 1, 1);
var inputb = g.concatColumns(input, baiseInput);
var uh = g.mul(inputb, Ua);
baiseInput = new WeightMatrix(stateRepeat.Rows, 1, 1);
stateRepeat = g.concatColumns(stateRepeat, baiseInput);
var wc = g.mul(stateRepeat, Wa);
var gg = g.tanh(g.add(uh, wc));
var aa = g.mul(gg, V);
var res = g.Softmax(aa);
var weighted = g.weightRows(input, res);;
context = g.sumColumns(weighted);
return(context);
}
private void Encode(Random r, out List <string> OutputSentence, out ComputeGraph g, out double cost, List <WeightMatrix> encoded)
{
var sentIndex = r.Next(0, InputSequences.Count);
var inputSentence = InputSequences[sentIndex];
var reversSentence = InputSequences[sentIndex].ToList();
reversSentence.Reverse();
OutputSentence = OutputSequences[sentIndex];
g = new ComputeGraph();
cost = 0.0;
for (int i = 0; i < inputSentence.Count; i++)
{
int ix_source = wordToIndex[inputSentence[i]];
int ix_source2 = wordToIndex[reversSentence[i]];
var x = g.PeekRow(Embedding, ix_source);
var eOutput = encoder.Encode(x, g);
var x2 = g.PeekRow(Embedding, ix_source2);
var eOutput2 = ReversEncoder.Encode(x2, g);
encoded.Add(g.concatColumns(eOutput, eOutput2));
}
//if (UseDropout)
//{
// encoded = g.Dropout(encoded, 0.2);
//}
}
public List <string> Predict(List <string> inputSeq)
{
ReversEncoder.Reset();
encoder.Reset();
decoder.Reset();
List <string> result = new List <string>();
var G2 = new ComputeGraph(false);
List <string> revseq = inputSeq.ToList();
revseq.Reverse();
List <WeightMatrix> encoded = new List <WeightMatrix>();
for (int i = 0; i < inputSeq.Count; i++)
{
int ix = wordToIndex[inputSeq[i]];
int ix2 = wordToIndex[revseq[i]];
var x2 = G2.PeekRow(Embedding, ix);
var o = encoder.Encode(x2, G2);
var x3 = G2.PeekRow(Embedding, ix2);
var eOutput2 = ReversEncoder.Encode(x3, G2);
var d = G2.concatColumns(o, eOutput2);
encoded.Add(d);
}
var ix_input = 1;
while (true)
{
var x = G2.PeekRow(Embedding, ix_input);
var eOutput = decoder.Decode(x, encoded.LastOrDefault(), G2);
if (UseDropout)
{
for (int i = 0; i < eOutput.Weight.Length; i++)
{
eOutput.Weight[i] *= 0.2;
}
}
var o = G2.add(
G2.mul(eOutput, this.Whd), this.bd);
if (UseDropout)
{
for (int i = 0; i < o.Weight.Length; i++)
{
o.Weight[i] *= 0.2;
}
}
var probs = G2.SoftmaxWithCrossEntropy(o);
var maxv = probs.Weight[0];
var maxi = 0;
for (int i = 1; i < probs.Weight.Length; i++)
{
if (probs.Weight[i] > maxv)
{
maxv = probs.Weight[i];
maxi = i;
}
}
var pred = maxi;
if (pred == 0)
{
break; // END token predicted, break out
}
if (result.Count > max_word)
{
break;
} // something is wrong
var letter2 = indexToWord[pred];
result.Add(letter2);
ix_input = pred;
}
return(result);
}
