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);
}
public WeightMatrix Perform(List <WeightMatrix> input, WeightMatrix state, ComputeGraph g)
{
WeightMatrix context;
List <WeightMatrix> atten = new List <WeightMatrix>();
foreach (var h_j in input)
{
var uh = g.add(g.mul(h_j, Ua), bUa);
var wc = g.add(g.mul(state, Wa), bWa);
var gg = g.tanh(g.add(uh, wc));
var aa = g.mul(gg, V);
atten.Add(aa);
}
var res = g.Softmax(atten);
var cmax = res[0].Weight[0];
int maxAtt = 0;
for (int i = 1; i < res.Count; i++)
{
if (res[i].Weight[0] > cmax)
{
cmax = res[i].Weight[0];
maxAtt = i;
}
}
this.MaxIndex = maxAtt;
context = g.scalemul(input[0], res[0]);
for (int hj = 1; hj < input.Count; hj++)
{
context = g.add(context, g.scalemul(input[hj], res[hj]));
}
return(context);
}
private double DecodeOutput(List <string> OutputSentence, ComputeGraph g, double cost, List <WeightMatrix> encoded)
{
int ix_input = 1;
for (int i = 0; i < OutputSentence.Count + 1; i++)
{
int ix_target = 0;
if (i == OutputSentence.Count)
{
ix_target = 0;
}
else
{
ix_target = wordToIndex[OutputSentence[i]];
}
var x = g.PeekRow(Embedding, ix_input);
var eOutput = decoder.Decode(x, encoded.LastOrDefault(), g);
if (UseDropout)
{
eOutput = g.Dropout(eOutput, 0.2);
}
var o = g.add(
g.mul(eOutput, this.Whd), this.bd);
if (UseDropout)
{
o = g.Dropout(o, 0.2);
}
var probs = g.SoftmaxWithCrossEntropy(o);
cost += -Math.Log(probs.Weight[ix_target]);
o.Gradient = probs.Weight;
o.Gradient[ix_target] -= 1;
ix_input = ix_target;
}
return(cost);
}
public WeightMatrix Step(WeightMatrix input, ComputeGraph innerGraph)
{
var hidden_prev = ht;
var cell_prev = ct;
var cell = this;
var h0 = innerGraph.mul(input, cell.Wix);
var h1 = innerGraph.mul(hidden_prev, cell.Wih);
var input_gate = innerGraph.sigmoid(
innerGraph.add(
innerGraph.add(h0, h1),
cell.bi
)
);
var h2 = innerGraph.mul(input, cell.Wfx);
var h3 = innerGraph.mul(hidden_prev, cell.Wfh);
var forget_gate = innerGraph.sigmoid(
innerGraph.add(
innerGraph.add(h2, h3),
cell.bf
)
);
var h4 = innerGraph.mul(input, cell.Wox);
var h5 = innerGraph.mul(hidden_prev, cell.Woh);
var output_gate = innerGraph.sigmoid(
innerGraph.add(
innerGraph.add(h4, h5),
cell.bo
)
);
var h6 = innerGraph.mul(input, cell.Wcx);
var h7 = innerGraph.mul(hidden_prev, cell.Wch);
var cell_write = innerGraph.tanh(
innerGraph.add(
innerGraph.add(h6, h7),
cell.bc
)
);
// compute new cell activation
var retain_cell = innerGraph.eltmul(forget_gate, cell_prev); // what do we keep from cell
var write_cell = innerGraph.eltmul(input_gate, cell_write); // what do we write to cell
var cell_d = innerGraph.add(retain_cell, write_cell); // new cell contents
// compute hidden state as gated, saturated cell activations
var hidden_d = innerGraph.eltmul(output_gate, innerGraph.tanh(cell_d));
this.ht = hidden_d;
this.ct = cell_d;
return(ht);
}
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);
}