public void run(object param) { param pa = (param)param; List <DataStep> x1 = pa.datastep; List <Matrix> xpro = new List <Matrix>(); ForwdBackwdProp g = new ForwdBackwdProp(_train); for (int i = 0; i < x1.Count; i++) { DataStep x = x1[i]; List <Matrix> add = new List <Matrix>(); for (int k = 0; k < 5; k++) { add.Add(Global.wordEmbedding[x.inputs[k]]); } List <Matrix> returnObj2 = Global.GRNNLayer1.activate(add, g); List <Matrix> returnObj3 = Global.GRNNLayer2.activate(returnObj2, g); List <Matrix> returnObj4 = Global.GRNNLayer3.activate(returnObj3, g); List <Matrix> returnObj5 = Global.GRNNLayer4.activate(returnObj4, g); xpro.Add(returnObj5[0]); } List <Matrix> returnObj6 = Global.upLSTMLayer.activate(xpro, g); List <Matrix> returnObj7 = Global.upLSTMLayerr.activate(reverse(xpro), g); List <Matrix> sum = new List <Matrix>(); for (int inde = 0; inde < returnObj6.Count(); inde++) { sum.Add(g.Add(returnObj6[inde], returnObj7[returnObj7.Count - inde - 1])); } for (int i = 0; i < returnObj6.Count; i++) { Matrix returnObj9 = Global.feedForwardLayer.Activate(sum[i], g); double loss = LossSoftmax.getLoss(returnObj9, x1[i].goldOutput); if (double.IsNaN(loss) || double.IsInfinity(loss)) { Console.WriteLine("WARNING!!!"); Global.swLog.WriteLine("WARNING!!!"); pa.mre.Set(); return; } LossSoftmax.getGrad(returnObj9, x1[i].goldOutput); } g.backwardProp(); pa.mre.Set(); }
public List <Matrix> activate(DataStep x, ForwdBackwdProp g) { List <int> input = x.inputs; List <int> bigram = x.bigram; Matrix final = new Matrix(Global.hiddenDim, 1); List <Matrix> outputs = new List <Matrix>(); Matrix _h_tm1 = Matrix.newMatrix_0(_hiddenDim, 1); Matrix _s_tm1 = Matrix.newMatrix_0(_hiddenDim, 1); for (int i = 0; i < input.Count; i++) { //input gate //Matrix con = g.ConcatVectors(Global.wordEmbedding[input[i]], Global.BigramwordEmbedding[bigram[i]]); //Matrix inputt = g.tanhNonlin(g.Mul(_hw, con)); Matrix inputt = Global.wordEmbedding[input[i]]; Matrix sum0 = g.Mul(_wix, inputt); Matrix sum1 = g.Mul(_wih, _h_tm1); Matrix inputGate = g.sigNonlin(g.Add(g.Add(sum0, sum1), _iBias)); //forget gate Matrix sum2 = g.Mul(_wfx, inputt); Matrix sum3 = g.Mul(_wfh, _h_tm1); Matrix forgetGate = g.sigNonlin(g.Add(g.Add(sum2, sum3), _fBias)); //output gate Matrix sum4 = g.Mul(_wox, inputt); Matrix sum5 = g.Mul(_woh, _h_tm1); Matrix outputGate = g.sigNonlin(g.Add(g.Add(sum4, sum5), _Bias)); //write operation on cells Matrix sum6 = g.Mul(_wcx, inputt); Matrix sum7 = g.Mul(_wch, _h_tm1); Matrix cellInput = g.tanhNonlin(g.Add(g.Add(sum6, sum7), _cBias)); //compute new cell activation Matrix retainCell = g.Elmul(forgetGate, _s_tm1); Matrix writeCell = g.Elmul(inputGate, cellInput); Matrix cellAct = g.Add(retainCell, writeCell); //compute hidden state as gated, saturated cell activations Matrix output = g.Elmul(outputGate, g.tanhNonlin(cellAct)); //if (i == 0) //{ // final = output; //} //else //{ // final = g.ConcatVectors(final, output); //} //final = g.Add(final, output); outputs.Add(output); //rollover activations for next iteration _h_tm1 = output; _s_tm1 = cellAct; //_h = g.Add(output, _h); } return(outputs); }