Exemple #1
0
        public static void AddLSTM(
            this NeuralModelBase model,
            out Neuron input,
            out Neuron output,
            BiasNeuron biasNeuron = null,
            WeightInitializer weightInitializer = null,
            string groupName = "LSTM")
        {
            var concatNeur = model.AddNeuron(
                sampleNeuron: new Neuron(-1, null)
                );

            // Multiply Gate
            var sigmoid1 = model.AddNeuron(
                sampleNeuron: new Neuron(-1, ActivationFunctions.Sigmoid)
                );

            model.AddConnection(concatNeur, sigmoid1, weightInitializer)
            .isTransferConnection = true;

            var multiplyGate = model.AddNeuron(
                sampleNeuron: new Neuron(-1, null)
            {
                ValueCollector = new MultValueCollector()
            }
                );

            model.AddConnection(sigmoid1, multiplyGate, weightInitializer);

            // Addition gate
            var sigmoid2 = model.AddNeuron(
                sampleNeuron: new Neuron(-1, ActivationFunctions.Sigmoid)
                );

            model.AddConnection(concatNeur, sigmoid2, weightInitializer)
            .isTransferConnection = true;

            var tanh = model.AddNeuron(
                sampleNeuron: new Neuron(-1, ActivationFunctions.TanH)
                );

            model.AddConnection(concatNeur, tanh, weightInitializer)
            .isTransferConnection = true;

            var sigmoidAndTanhMultGate = model.AddNeuron(
                sampleNeuron: new Neuron(-1, null)
            {
                ValueCollector = new MultValueCollector()
            }
                );

            model.AddConnection(sigmoid2, sigmoidAndTanhMultGate, weightInitializer);
            model.AddConnection(tanh, sigmoidAndTanhMultGate, weightInitializer);

            var additionGate = model.AddNeuron(new Neuron(-1, null));

            model.AddConnection(multiplyGate, additionGate, weightInitializer)
            .isTransferConnection = true;
            model.AddConnection(sigmoidAndTanhMultGate, additionGate, weightInitializer)
            .isTransferConnection = true;

            // Tanh gate
            var sigmoid3 = model.AddNeuron(
                sampleNeuron: new Neuron(-1, ActivationFunctions.Sigmoid)
                );

            model.AddConnection(concatNeur, sigmoid3, weightInitializer)
            .isTransferConnection = true;

            var finalMult = model.AddNeuron(
                sampleNeuron: new Neuron(-1, null)
            {
                ValueCollector = new MultValueCollector()
            }
                );

            var tanhGate = model.AddNeuron(
                sampleNeuron: new Neuron(-1, ActivationFunctions.TanH)
                );

            model.AddConnection(additionGate, tanhGate, weightInitializer)
            .isTransferConnection = true;

            model.AddConnection(sigmoid3, finalMult, weightInitializer);
            model.AddConnection(tanhGate, finalMult, weightInitializer)
            .isTransferConnection = true;

            // Adding memory neurons
            var finalMultMem = model.AddNeuron(
                sampleNeuron: new MemoryNeuron(-1, finalMult.InnovationNb)
                );

            model.AddConnection(finalMultMem, concatNeur, weightInitializer)
            .isTransferConnection = true;

            var cellStateMem = model.AddNeuron(
                sampleNeuron: new MemoryNeuron(-1, additionGate.InnovationNb)
                );

            model.AddConnection(cellStateMem, multiplyGate, weightInitializer)
            .isTransferConnection = true;

            // Connecting bias
            if (biasNeuron != null)
            {
                model.AddConnection(biasNeuron, sigmoid1);
                model.AddConnection(biasNeuron, sigmoid2);
                model.AddConnection(biasNeuron, tanh);
                model.AddConnection(biasNeuron, sigmoid3);
            }

            // Assign neuron group
            concatNeur.group = groupName;
            sigmoid1.group   = groupName;
            sigmoid2.group   = groupName;
            sigmoid3.group   = groupName;
            tanh.group       = groupName;

            additionGate.group = groupName;
            multiplyGate.group = groupName;
            tanhGate.group     = groupName;

            sigmoidAndTanhMultGate.group = groupName;
            finalMult.group = groupName;

            cellStateMem.group = groupName;
            finalMultMem.group = groupName;

            // Assigning out's
            input  = concatNeur;
            output = finalMult;
        }