public static LSTMNetwork Create(int hiddenNeuronCount, int inputDim)
{
var graph = new NetworkGraph();
var inputs = graph.Block("inputs", inputDim, "", true);
var previousH = graph.Block("previousH", hiddenNeuronCount, "ht", true);
var previousC = graph.Block("previousC", hiddenNeuronCount, "ct_tanh", true);
var f_gate_W = graph.NeuronLayer("f_gate_W", hiddenNeuronCount);
var f_gate_U = graph.NeuronLayer("f_gate_U", hiddenNeuronCount);
var i_gate_W = graph.NeuronLayer("i_gate_W", hiddenNeuronCount);
var i_gate_U = graph.NeuronLayer("i_gate_U", hiddenNeuronCount);
var o_gate_W = graph.NeuronLayer("o_gate_W", hiddenNeuronCount);
var o_gate_U = graph.NeuronLayer("o_gate_U", hiddenNeuronCount);
var c_gate_W = graph.NeuronLayer("c_gate_W", hiddenNeuronCount);
var c_gate_U = graph.NeuronLayer("c_gate_U", hiddenNeuronCount);
var plus = graph.Operation("plus", new AddOperation());
var multiply = graph.Operation("multiply", new MultiplyOperation());
var sigmoid = graph.Operation("sigmoid", new SigmoidOperation());
var tanh = graph.Operation("tanh", new TanhOperation());
var fc = graph.Operation("fc", new FullConnectionOperation());
var f_gate_W_fc_out = graph.Connect("f_gate_W_fc_out", new GraphNode[] { inputs, f_gate_W }, new GraphNode[] { fc });
var i_gate_W_fc_out = graph.Connect("i_gate_W_fc_out", new GraphNode[] { inputs, i_gate_W }, new GraphNode[] { fc });
var o_gate_W_fc_out = graph.Connect("o_gate_W_fc_out", new GraphNode[] { inputs, o_gate_W }, new GraphNode[] { fc });
var c_gate_W_fc_out = graph.Connect("c_gate_W_fc_out", new GraphNode[] { inputs, c_gate_W }, new GraphNode[] { fc });
var f_gate_U_fc_out = graph.Connect("f_gate_U_fc_out", new GraphNode[] { previousH, f_gate_U }, new GraphNode[] { fc });
var i_gate_U_fc_out = graph.Connect("i_gate_U_fc_out", new GraphNode[] { previousH, i_gate_U }, new GraphNode[] { fc });
var o_gate_U_fc_out = graph.Connect("o_gate_U_fc_out", new GraphNode[] { previousH, o_gate_U }, new GraphNode[] { fc });
var c_gate_U_fc_out = graph.Connect("c_gate_U_fc_out", new GraphNode[] { previousH, c_gate_U }, new GraphNode[] { fc });
var f_value = graph.Connect("f_value", new GraphNode[] { f_gate_W_fc_out, f_gate_U_fc_out }, new GraphNode[] { plus, sigmoid });
var i_value = graph.Connect("i_value", new GraphNode[] { i_gate_W_fc_out, i_gate_U_fc_out }, new GraphNode[] { plus, sigmoid });
var o_value = graph.Connect("o_value", new GraphNode[] { o_gate_W_fc_out, o_gate_U_fc_out }, new GraphNode[] { plus, sigmoid });
var c1_value = graph.Connect("c1_value", new GraphNode[] { c_gate_W_fc_out, c_gate_U_fc_out }, new GraphNode[] { plus, tanh });
var f_c_value = graph.Connect("f_c_value", new GraphNode[] { f_value, previousC }, new GraphNode[] { multiply });
var i_c1_value = graph.Connect("i_c1_value", new GraphNode[] { i_value, c1_value }, new GraphNode[] { multiply });
var ct_tanh = graph.Connect("ct_tanh", new GraphNode[] { f_c_value, i_c1_value }, new GraphNode[] { plus, tanh }, true);
var ht = graph.Connect("ht", new GraphNode[] { ct_tanh, o_value }, new GraphNode[] { multiply }, true);
ct_tanh.IsMonitoringNode = true;
ht.IsMonitoringNode = true;
graph.Compile();
LSTMNetwork network = new LSTMNetwork(graph);
return(network);
}
public static LSTMNetwork Create2(int hiddenNeuronCount, int inputDim)
{
var graph = new ComputeGraph();
var inputs = graph.DeclareInput("inputs", inputDim);
var previousH = graph.DeclareCopyFrom("previousH", hiddenNeuronCount, "ht");
var previousC = graph.DeclareCopyFrom("previousC", hiddenNeuronCount, "ct_tanh");
var f_gate_W = graph.DeclareFC("f_gate_W", inputDim, hiddenNeuronCount, inputs);
var f_gate_U = graph.DeclareFC("f_gate_U", hiddenNeuronCount, hiddenNeuronCount, previousH);
var i_gate_W = graph.DeclareFC("i_gate_W", inputDim, hiddenNeuronCount, inputs);
var i_gate_U = graph.DeclareFC("i_gate_U", hiddenNeuronCount, hiddenNeuronCount, previousH);
var o_gate_W = graph.DeclareFC("o_gate_W", inputDim, hiddenNeuronCount, inputs);
var o_gate_U = graph.DeclareFC("o_gate_U", hiddenNeuronCount, hiddenNeuronCount, previousH);
var c_gate_W = graph.DeclareFC("c_gate_W", inputDim, hiddenNeuronCount, inputs);
var c_gate_U = graph.DeclareFC("c_gate_U", hiddenNeuronCount, hiddenNeuronCount, previousH);
var f_value = graph.DeclarePlus("f_value", f_gate_W, f_gate_U);
var i_value = graph.DeclarePlus("i_value", i_gate_W, i_gate_U);
var o_value = graph.DeclarePlus("o_value", o_gate_W, o_gate_U);
var c1_value = graph.DeclarePlus("c1_value", c_gate_W, c_gate_U);
var f_sigmoid_value = graph.DeclareSigmoid("f_sigmoid_value", f_value);
var i_sigmoid_value = graph.DeclareSigmoid("i_sigmoid_value", i_value);
var o_sigmoid_value = graph.DeclareSigmoid("o_sigmoid_value", o_value);
var c1_tanh_value = graph.DeclareTanh("c1_tanh_value", c1_value);
var f_c_value = graph.DeclareMultiply("f_c_value", f_sigmoid_value, previousC);
var i_c1_value = graph.DeclareMultiply("i_c1_value", i_sigmoid_value, c1_tanh_value);
var f_c_i_c1_plus = graph.DeclarePlus("f_c_i_c1_plus", f_c_value, i_c1_value);
var ct_tanh = graph.DeclareTanh("ct_tanh", f_c_i_c1_plus);
var ht = graph.DeclareMultiply("ht", ct_tanh, o_sigmoid_value);
ct_tanh.ResultsDim = hiddenNeuronCount;
ht.ResultsDim = hiddenNeuronCount;
ct_tanh.IsMonitoring = true;
ht.IsMonitoring = true;
graph.Compile();
LSTMNetwork network = new LSTMNetwork(graph);
return(network);
}
